xml
stringlengths 7
23
| proceedings
stringlengths 58
222
| year
stringdate 2005-01-01 00:00:00
2024-01-01 00:00:00
| url
stringlengths 1
64
| language documentation
stringclasses 1
value | has non-English?
stringclasses 4
values | topics
stringclasses 7
values | language coverage
stringclasses 32
values | title
stringlengths 32
161
| abstract
stringlengths 176
2.45k
|
|---|---|---|---|---|---|---|---|---|---|
2024.acl.xml | Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) | 2024 | https://aclanthology.org/2024.acl-long.812 | x | 0 | general safety, LLM alignment | null | PsySafe: A Comprehensive Framework for Psychological-based Attack, Defense, and Evaluation of Multi-agent System Safety | Multi-agent systems, when enhanced with Large Language Models (LLMs), exhibit profound capabilities in collective intelligence. However, the potential misuse of this intelligence for malicious purposes presents significant risks. To date, comprehensive research on the safety issues associated with multi-agent systems remains limited. In this paper, we explore these concerns through the innovative lens of agent psychology, revealing that the dark psychological states of agents constitute a significant threat to safety.To tackle these concerns, we propose a comprehensive framework (PsySafe) grounded in agent psychology, focusing on three key areas: firstly, identifying how dark personality traits in agents can lead to risky behaviors; secondly, evaluating the safety of multi-agent systems from the psychological and behavioral perspectives, and thirdly, devising effective strategies to mitigate these risks.Our experiments reveal several intriguing phenomena, such as the collective dangerous behaviors among agents, agents’ self-reflection when engaging in dangerous behavior, and the correlation between agents’ psychological assessments and dangerous behaviors. We anticipate that our framework and observations will provide valuable insights for further research into the safety of multi-agent systems. We make our data and code publicly accessible at https://github.com/AI4Good24/PsySafe. |
2024.acl.xml | Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) | 2024 | https://aclanthology.org/2024.acl-long.824 | x | 0 | others | null | Calibrating Large Language Models Using Their Generations Only | As large language models (LLMs) are increasingly deployed in user-facing applications, building trust and maintaining safety by accurately quantifying a model’s confidence in its prediction becomes even more important. However, finding effective ways to calibrate LLMs—especially when the only interface to the models is their generated text—remains a challenge. We propose APRICOT (Auxiliary prediction of confidence targets): A method to set confidence targets and train an additional model that predicts an LLM’s confidence based on its textual input and output alone. This approach has several advantages: It is conceptually simple, does not require access to the target model beyond its output, does not interfere with the language generation, and has a multitude of potential usages, for instance by verbalizing the predicted confidence or using it to re-prompting the LLM to accurately reflecting its uncertainty. We show how our approach performs competitively in terms of calibration error for white-box and black-box LLMs on closed-book question-answering to detect incorrect LLM answers. |
2024.acl.xml | Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) | 2024 | https://aclanthology.org/2024.acl-long.830 | null | 1 | general safety, LLM alignment | Chinese, English | SafetyBench: Evaluating the Safety of Large Language Models | With the rapid development of Large Language Models (LLMs), increasing attention has been paid to their safety concerns. Consequently, evaluating the safety of LLMs has become an essential task for facilitating the broad applications of LLMs. Nevertheless, the absence of comprehensive safety evaluation benchmarks poses a significant impediment to effectively assess and enhance the safety of LLMs. In this work, we present SafetyBench, a comprehensive benchmark for evaluating the safety of LLMs, which comprises 11,435 diverse multiple choice questions spanning across 7 distinct categories of safety concerns. Notably, SafetyBench also incorporates both Chinese and English data, facilitating the evaluation in both languages. Our extensive tests over 25 popular Chinese and English LLMs in both zero-shot and few-shot settings reveal a substantial performance advantage for GPT-4 over its counterparts, and there is still significant room for improving the safety of current LLMs. We also demonstrate that the measured safety understanding abilities in SafetyBench are correlated with safety generation abilities. Data and evaluation guidelines are available at https://github.com/thu-coai/SafetyBench. Submission entrance and leaderboard are available at https://llmbench.ai/safety. |
2024.acl.xml | Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) | 2024 | https://aclanthology.org/2024.acl-long.842 | x | 0 | jailbreaking attacks | null | Emulated Disalignment: Safety Alignment for Large Language Models May Backfire! | Large language models (LLMs) undergo safety alignment to ensure safe conversations with humans. However, this paper introduces a training-free attack method capable of reversing safety alignment, converting the outcomes of stronger alignment into greater potential for harm by accessing only LLM output token distributions. Specifically, our method achieves this reversal by contrasting the output token distribution of a safety-aligned language model (e.g., Llama-2-chat) against its pre-trained version (e.g., Llama-2), so that the token predictions are shifted towards the opposite direction of safety alignment.We name this method emulated disalignment (ED) because sampling from this contrastive distribution provably emulates the result of fine-tuning to minimize a safety reward.Our experiments with ED across three evaluation datasets and four model families (Llama-1, Llama-2, Mistral, and Alpaca) show that ED doubles the harmfulness of pre-trained models and outperforms strong baselines, achieving the highest harmful rates in 43 out of 48 evaluation subsets by a large margin.Eventually, given ED’s reliance on language model output token distributions, which particularly compromises open-source models, our findings highlight the need to reassess the open accessibility of language models, even if they have been safety-aligned.Code is available at https://github.com/ZHZisZZ/emulated-disalignment. |
2024.acl.xml | Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) | 2024 | https://aclanthology.org/2024.acl-long.845 | null | 1 | general safety, LLM alignment | Scottish Gaelic, Ukrainian, Hindi, Thai, Simplified Chinese, Hebrew, English, Bengali, Standard Arabic, Italian, Zulu, English | Aya Model: An Instruction Finetuned Open-Access Multilingual Language Model | Recent breakthroughs in large language models (LLMs) have centered around a handful of data-rich languages. What does it take to broaden access to breakthroughs beyond first-class citizen languages? Our work introduces Aya, a massively multilingual generative language model that follows instructions in 101 languages of which over 50% are considered as lower-resourced. Aya outperforms mT0 and BLOOMZ on the majority of tasks while covering double the number of languages. We introduce extensive new evaluation suites that broaden the state-of-art for multilingual eval across 99 languages —— including discriminative and generative tasks, human evaluation, and simulated win rates that cover both held-out tasks and in-distribution performance. Furthermore, we conduct detailed investigations on the optimal finetuning mixture composition, data pruning, as well as the toxicity, bias, and safety of our models. |
2024.acl.xml | Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) | 2024 | https://aclanthology.org/2024.acl-short.75 | null | 1 | toxicity, bias | Arabic, Bengali, German, Spanish, French, Hindi, Indonesian, Italian, Japanese, Korean, Marathi, Malay, Dutch, Portuguese, Swahili, Tamil, Telugu, Thai, Turkish, Vietnamese | SeeGULL Multilingual: a Dataset of Geo-Culturally Situated Stereotypes | While generative multilingual models are rapidly being deployed, their safety and fairness evaluations are largely limited to resources collected in English. This is especially problematic for evaluations targeting inherently socio-cultural phenomena such as stereotyping, where it is important to build multilingual resources that reflect the stereotypes prevalent in respective language communities. However, gathering these resources, at scale, in varied languages and regions pose a significant challenge as it requires broad socio-cultural knowledge and can also be prohibitively expensive. To overcome this critical gap, we employ a recently introduced approach that couples LLM generations for scale with culturally situated validations for reliability, and build SeeGULL Multilingual, a global-scale multilingual dataset of social stereotypes, containing over 25K stereotypes, spanning 23 pairs of languages and regions they are common in, with human annotations, and demonstrate its utility in identifying gaps in model evaluations. |
2024.acl.xml | Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) | 2024 | https://aclanthology.org/2024.acl-demos.18 | null | null | others | null | DrugWatch: A Comprehensive Multi-Source Data Visualisation Platform for Drug Safety Information | Drug safety research is crucial for maintaining public health, often requiring comprehensive data support. However, the resources currently available to the public are limited and fail to provide a comprehensive understanding of the relationship between drugs and their side effects. This paper introduces “DrugWatch”, an easy-to-use and interactive multi-source information visualisation platform for drug safety study. It allows users to understand common side effects of drugs and their statistical information, flexibly retrieve relevant medical reports, or annotate their own medical texts with our automated annotation tool. Supported by NLP technology and enriched with interactive visual components, we are committed to providing researchers and practitioners with a one-stop information analysis, retrieval, and annotation service. The demonstration video is available at https://www.youtube.com/watch?v=RTqDgxzETjw. We also deployed an online demonstration system at https://drugwatch.net/. |
2024.acl.xml | Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) | 2024 | https://aclanthology.org/2024.acl-demos.19 | null | 1 | general safety, LLM alignment | Chinese | OpenEval: Benchmarking Chinese LLMs across Capability, Alignment and Safety | The rapid development of Chinese large language models (LLMs) poses big challenges for efficient LLM evaluation. While current initiatives have introduced new benchmarks or evaluation platforms for assessing Chinese LLMs, many of these focus primarily on capabilities, usually overlooking potential alignment and safety issues. To address this gap, we introduce OpenEval, an evaluation testbed that benchmarks Chinese LLMs across capability, alignment and safety. For capability assessment, we include 12 benchmark datasets to evaluate Chinese LLMs from 4 sub-dimensions: NLP tasks, disciplinary knowledge, commonsense reasoning and mathematical reasoning. For alignment assessment, OpenEval contains 7 datasets that examines the bias, offensiveness and illegalness in the outputs yielded by Chinese LLMs. To evaluate safety, especially anticipated risks (e.g., power-seeking, self-awareness) of advanced LLMs, we include 6 datasets. In addition to these benchmarks, we have implemented a phased public evaluation and benchmark update strategy to ensure that OpenEval is in line with the development of Chinese LLMs or even able to provide cutting-edge benchmark datasets to guide the development of Chinese LLMs. In our first public evaluation, we have tested a range of Chinese LLMs, spanning from 7B to 72B parameters, including both open-source and proprietary models. Evaluation results indicate that while Chinese LLMs have shown impressive performance in certain tasks, more attention should be directed towards broader aspects such as commonsense reasoning, alignment, and safety. |
2024.acl.xml | Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) | 2024 | https://aclanthology.org/2024.acl-srw.29 | null | 1 | jailbreaking attacks | Italian | Compromesso! Italian Many-Shot Jailbreaks undermine the safety of Large Language Models | As diverse linguistic communities and users adopt Large Language Models (LLMs), assessing their safety across languages becomes critical. Despite ongoing efforts to align these models with safe and ethical guidelines, they can still be induced into unsafe behavior with jailbreaking, a technique in which models are prompted to act outside their operational guidelines. What research has been conducted on these vulnerabilities was predominantly on English, limiting the understanding of LLM behavior in other languages. We address this gap by investigating Many-Shot Jailbreaking (MSJ) in Italian, underscoring the importance of understanding LLM behavior in different languages. We base our analysis on a newly created Italian dataset to identify unique safety vulnerabilities in 4 families of open-source LLMs.We find that the models exhibit unsafe behaviors even with minimal exposure to harmful prompts, and–more alarmingly–this tendency rapidly escalates with more demonstrations. |
2024.acl.xml | Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) | 2024 | https://aclanthology.org/2024.acl-tutorials.5 | x | 0 | others | null | Vulnerabilities of Large Language Models to Adversarial Attacks | This tutorial serves as a comprehensive guide on the vulnerabilities of Large Language Models (LLMs) to adversarial attacks, an interdisciplinary field that blends perspectives from Natural Language Processing (NLP) and Cybersecurity. As LLMs become more complex and integrated into various systems, understanding their security attributes is crucial. However, current research indicates that even safety-aligned models are not impervious to adversarial attacks that can result in incorrect or harmful outputs. The tutorial first lays the foundation by explaining safety-aligned LLMs and concepts in cybersecurity. It then categorizes existing research based on different types of learning architectures and attack methods. We highlight the existing vulnerabilities of unimodal LLMs, multi-modal LLMs, and systems that integrate LLMs, focusing on adversarial attacks designed to exploit weaknesses and mislead AI systems. Finally, the tutorial delves into the potential causes of these vulnerabilities and discusses potential defense mechanisms. |
2024.semeval.xml | Proceedings of the 18th International Workshop on Semantic Evaluation (SemEval-2024) | 2024 | https://aclanthology.org/2024.semeval-1.91 | null | null | others | null | D-NLP at SemEval-2024 Task 2: Evaluating Clinical Inference Capabilities of Large Language Models | Large language models (LLMs) have garnered significant attention and widespread usage due to their impressive performance in various tasks. However, they are not without their own set of challenges, including issues such as hallucinations, factual inconsistencies, and limitations in numerical-quantitative reasoning. Evaluating LLMs in miscellaneous reasoning tasks remains an active area of research. Prior to the breakthrough of LLMs, Transformers had already proven successful in the medical domain, effectively employed for various natural language understanding (NLU) tasks. Following this trend, LLMs have also been trained and utilized in the medical domain, raising concerns regarding factual accuracy, adherence tosafety protocols, and inherent limitations. In this paper, we focus on evaluating the natural language inference capabilities of popular open-source and closed-source LLMs using clinical trial reports as the dataset. We present the performance results of each LLM and further analyze their performance on a development set, particularly focusing on challenging instances that involve medical abbreviations and require numerical-quantitative reasoning. Gemini, our leading LLM, achieved a test set F1-score of 0.748, securing the ninth position on the task scoreboard. Our work is the first of its kind, offering a thorough examination of the inference capabilities of LLMs within the medical domain. |
2024.eacl.xml | Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers) | 2024 | https://aclanthology.org/2024.eacl-long.77 | x | null | general safety, LLM alignment | null | Plan-Grounded Large Language Models for Dual Goal Conversational Settings | Training Large Language Models (LLMs) to follow user instructions has shown to supply the LLM with ample capacity to converse fluently while being aligned with humans. Yet, it is not completely clear how an LLM can lead a plan-grounded conversation in mixed-initiative settings where instructions flow in both directions of the conversation, i.e. both the LLM and the user provide instructions to one another. In this paper, we tackle a dual goal mixed-initiative conversational setting where the LLM not only grounds the conversation on an arbitrary plan but also seeks to satisfy both a procedural plan and user instructions. The LLM is then responsible for guiding the user through the plan and, at the same time, adapting to new circumstances, answering questions, and activating safety guardrails when needed. We propose a novel LLM that grounds the dialogue on a procedural plan, can take the dialogue initiative, and enforces guardrails on the system’s behavior, while also improving the LLM’s responses to unexpected user behavior. Experiments in controlled settings and with real users show that the best-performing model, which we call PlanLLM, achieves a 2.1x improvement over a strong baseline. Moreover, experiments also show good generalization to unseen domains. |
2024.eacl.xml | Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers) | 2024 | https://aclanthology.org/2024.eacl-long.175 | null | 1 | general safety, LLM alignment | German, English | Gradient-Based Language Model Red Teaming | Red teaming is a common strategy for identifying weaknesses in generative language models (LMs) by producing adversarial prompts that trigger models to generate unsafe responses. Red teaming is instrumental for both model alignment and evaluation, but is labor-intensive and difficult to scale when done by humans. In this paper, we present Gradient-Based Red Teaming (GBRT), a novel red teaming method for automatically generating diverse prompts that are likely to cause an LM to output unsafe responses. GBRT is a form of prompt learning, trained by scoring an LM response with a safety classifier and then backpropagating through the frozen safety classifier and LM to update the prompt. To improve the coherence of input prompts, we introduce two variants that add a realism loss and fine-tune a pretrained model to generate the prompts instead of learning the prompts directly. Our experiments show that GBRT is more effective at finding prompts that trigger an LM to generate unsafe responses than a strong reinforcement learning-based red teaming approach and works even when the LM has been fine-tuned to produce safer outputs. |
2023.gem.xml | Proceedings of the Third Workshop on Natural Language Generation, Evaluation, and Metrics (GEM) | 2023 | https://aclanthology.org/2023.gem-1.10 | null | 0 | general safety, LLM alignment | null | Unveiling Safety Vulnerabilities of Large Language Models | As large language models become more prevalent, their possible harmful or inappropriate responses are a cause for concern. This paper introduces a unique dataset containing adversarial examples in the form of questions, we call AttaQ, designed to provoke such harmful or inappropriate responses. We assess the efficacy of our dataset by analyzing the vulnerabilities of various models when subjected to it. Additionally, we introduce a novel automatic approach for identifying and naming vulnerable semantic regions — input semantic areas for which the model is likely to produce harmful outputs. This is achieved through the application of specialized clustering techniques that consider both the semantic similarity of the input attacks and the harmfulness of the model’s responses.Automatically identifying vulnerable semantic regions enhances the evaluation of model weaknesses, facilitating targeted improvements to its safety mechanisms and overall reliability. |
2023.semeval.xml | Proceedings of the 17th International Workshop on Semantic Evaluation (SemEval-2023) | 2023 | https://aclanthology.org/2023.semeval-1.113 | null | 0 | toxicity, bias | null | CNLP-NITS at SemEval-2023 Task 10: Online sexism prediction, PREDHATE! | Online sexism is a rising issue that threatens women’s safety, fosters hostile situations, and upholds social inequities. We describe a task SemEval-2023 Task 10 for creating English-language models that can precisely identify and categorize sexist content on internet forums and social platforms like Gab and Reddit as well to provide an explainability in order to address this problem. The problem is divided into three hierarchically organized subtasks: binary sexism detection, sexism by category, and sexism by fine-grained vector. The dataset consists of 20,000 labelled entries. For Task A, pertained models like Convolutional Neural Network (CNN) and Bidirectional Long Short-Term Memory (BiLSTM), which is called CNN-BiLSTM and Generative Pretrained Transformer 2 (GPT-2) models were used, as well as the GPT-2 model for Task B and C, and have provided experimental configurations. According to our findings, the GPT-2 model performs better than the CNN-BiLSTM model for Task A, while GPT-2 is highly accurate for Tasks B and C on the training, validation and testing splits of the training data provided in the task. Our proposed models allow researchers to create more precise and understandable models for identifying and categorizing sexist content in online forums, thereby empowering users and moderators. |
2023.clinicalnlp.xml | Proceedings of the 5th Clinical Natural Language Processing Workshop | 2023 | https://aclanthology.org/2023.clinicalnlp-1.17 | null | null | others | null | Can Large Language Models Safely Address Patient Questions Following Cataract Surgery? | Recent advances in large language models (LLMs) have generated significant interest in their application across various domains including healthcare. However, there is limited data on their safety and performance in real-world scenarios. This study uses data collected using an autonomous telemedicine clinical assistant. The assistant asks symptom-based questions to elicit patient concerns and allows patients to ask questions about their post-operative recovery. We utilise real-world postoperative questions posed to the assistant by a cohort of 120 patients to examine the safety and appropriateness of responses generated by a recent popular LLM by OpenAI, ChatGPT. We demonstrate that LLMs have the potential to helpfully address routine patient queries following routine surgery. However, important limitations around the safety of today’s models exist which must be considered. |
2023.banglalp.xml | Proceedings of the First Workshop on Bangla Language Processing (BLP-2023) | 2023 | https://aclanthology.org/2023.banglalp-1.28 | null | 1 | toxicity, bias | Bangla | Semantics Squad at BLP-2023 Task 1: Violence Inciting Bangla Text Detection with Fine-Tuned Transformer-Based Models | This study investigates the application of Transformer-based models for violence threat identification. We participated in the BLP-2023 Shared Task 1 and in our initial submission, BanglaBERT large achieved 5th position on the leader-board with a macro F1 score of 0.7441, approaching the highest baseline of 0.7879 established for this task. In contrast, the top-performing system on the leaderboard achieved an F1 score of 0.7604. Subsequent experiments involving m-BERT, XLM-RoBERTa base, XLM-RoBERTa large, BanglishBERT, BanglaBERT, and BanglaBERT large models revealed that BanglaBERT achieved an F1 score of 0.7441, which closely approximated the baseline. Remarkably, m-BERT and XLM-RoBERTa base also approximated the baseline with macro F1 scores of 0.6584 and 0.6968, respectively. A notable finding from our study is the under-performance by larger models for the shared task dataset, which requires further investigation. Our findings underscore the potential of transformer-based models in identifying violence threats, offering valuable insights to enhance safety measures on online platforms. |
2023.banglalp.xml | Proceedings of the First Workshop on Bangla Language Processing (BLP-2023) | 2023 | https://aclanthology.org/2023.banglalp-1.30 | null | 1 | toxicity, bias | Bengali | Team Error Point at BLP-2023 Task 1: A Comprehensive Approach for Violence Inciting Text Detection using Deep Learning and Traditional Machine Learning Algorithm | In the modern digital landscape, social media platforms have the dual role of fostering unprecedented connectivity and harboring a dark underbelly in the form of widespread violence-inciting content. Pioneering research in Bengali social media aims to provide a groundbreaking solution to this issue. This study thoroughly investigates violence-inciting text classification using a diverse range of machine learning and deep learning models, offering insights into content moderation and strategies for enhancing online safety. Situated at the intersection of technology and social responsibility, the aim is to empower platforms and communities to combat online violence. By providing insights into model selection and methodology, this work makes a significant contribution to the ongoing dialogue about the challenges posed by the darker aspects of the digital era. Our system scored 31.913 and ranked 26 among the participants. |
2023.findings.xml | Findings of the Association for Computational Linguistics: EACL 2023 | 2023 | https://aclanthology.org/2023.findings-acl.157 | x | 0 | jailbreaking attacks | null | Diffusion Theory as a Scalpel: Detecting and Purifying Poisonous Dimensions in Pre-trained Language Models Caused by Backdoor or Bias | Pre-trained Language Models (PLMs) may be poisonous with backdoors or bias injected by the suspicious attacker during the fine-tuning process. A core challenge of purifying potentially poisonous PLMs is precisely finding poisonous dimensions. To settle this issue, we propose the Fine-purifying approach, which utilizes the diffusion theory to study the dynamic process of fine-tuning for finding potentially poisonous dimensions. According to the relationship between parameter drifts and Hessians of different dimensions, we can detect poisonous dimensions with abnormal dynamics, purify them by resetting them to clean pre-trained weights, and then fine-tune the purified weights on a small clean dataset. To the best of our knowledge, we are the first to study the dynamics guided by the diffusion theory for safety or defense purposes. Experimental results validate the effectiveness of Fine-purifying even with a small clean dataset. |
2023.findings.xml | Findings of the Association for Computational Linguistics: EACL 2023 | 2023 | https://aclanthology.org/2023.findings-acl.212 | null | 0 | general safety, LLM alignment | null | Adversarial Textual Robustness on Visual Dialog | Adversarial robustness evaluates the worst-case performance scenario of a machine learning model to ensure its safety and reliability. For example, cases where the user input contains a minimal change, e.g. a synonym, which causes the previously correct model to return a wrong answer. Using this scenario, this study is the first to investigate the robustness of visually grounded dialog models towards textual attacks. We first aim to understand how multimodal input components contribute to model robustness. Our results show that models which encode dialog history are more robust by providing redundant information. This is in contrast to prior work which finds that dialog history is negligible for model performance on this task. We also evaluate how to generate adversarial test examples which successfully fool the model but remain undetected by the user/software designer. Our analysis shows that the textual, as well as the visual context are important to generate plausible attacks. |
2023.findings.xml | Findings of the Association for Computational Linguistics: EACL 2023 | 2023 | https://aclanthology.org/2023.findings-acl.267 | x | 0 | others | null | TextVerifier: Robustness Verification for Textual Classifiers with Certifiable Guarantees | When textual classifiers are deployed in safety-critical workflows, they must withstand the onslaught of AI-enabled model confusion caused by adversarial examples with minor alterations. In this paper, the main objective is to provide a formal verification framework, called TextVerifier, with certifiable guarantees on deep neural networks in natural language processing against word-level alteration attacks. We aim to provide an approximation of the maximal safe radius by deriving provable bounds both mathematically and automatically, where a minimum word-level L_0 distance is quantified as a guarantee for the classification invariance of victim models. Here, we illustrate three strengths of our strategy: i) certifiable guarantee: effective verification with convergence to ensure approximation of maximal safe radius with tight bounds ultimately; ii) high-efficiency: it yields an efficient speed edge by a novel parallelization strategy that can process a set of candidate texts simultaneously on GPUs; and iii) reliable anytime estimation: the verification can return intermediate bounds, and robustness estimates that are gradually, but strictly, improved as the computation proceeds. Furthermore, experiments are conducted on text classification on four datasets over three victim models to demonstrate the validity of tightening bounds. Our tool TextVerifier is available at https://github.com/TrustAI/TextVerifer. |
2023.findings.xml | Findings of the Association for Computational Linguistics: EACL 2023 | 2023 | https://aclanthology.org/2023.findings-acl.701 | null | 0 | general safety, LLM alignment | null | Foveate, Attribute, and Rationalize: Towards Physically Safe and Trustworthy AI | Users’ physical safety is an increasing concern as the market for intelligent systems continues to grow, where unconstrained systems may recommend users dangerous actions that can lead to serious injury. Covertly unsafe text is an area of particular interest, as such text may arise from everyday scenarios and are challenging to detect as harmful. We propose FARM, a novel framework leveraging external knowledge for trustworthy rationale generation in the context of safety. In particular, FARM foveates on missing knowledge to qualify the information required to reason in specific scenarios and retrieves this information with attribution to trustworthy sources. This knowledge is used to both classify the safety of the original text and generate human-interpretable rationales, shedding light on the risk of systems to specific user groups and helping both stakeholders manage the risks of their systems and policymakers to provide concrete safeguards for consumer safety. Our experiments show that FARM obtains state-of-the-art results on the SafeText dataset, showing absolute improvement in safety classification accuracy by 5.9%. |
2023.findings.xml | Findings of the Association for Computational Linguistics: EACL 2023 | 2023 | https://aclanthology.org/2023.findings-acl.719 | null | null | privacy | null | Membership Inference Attacks against Language Models via Neighbourhood Comparison | Membership Inference attacks (MIAs) aim to predict whether a data sample was present in the training data of a machine learning model or not, and are widely used for assessing the privacy risks of language models. Most existing attacks rely on the observation that models tend toassign higher probabilities to their training samples than non-training points. However, simple thresholding of the model score in isolation tends to lead to high false-positive rates as it does not account for the intrinsic complexity of a sample. Recent work has demonstrated that reference-based attacks which compare model scores to those obtained from a reference model trained on similar data can substantially improve the performance of MIAs.However, in order to train reference models, attacks of this kind make the strong and arguably unrealistic assumption that an adversary has access to samples closely resembling the original training data. Therefore, we investigate their performance in more realistic scenarios and find that they are highly fragile in relation to the data distribution used to train reference models. To investigate whether this fragility provides a layer of safety, we propose and evaluate neighbourhood attacks, which compare model scores for a given sample to scores of synthetically generated neighbour texts and therefore eliminate the need for access to the training data distribution. We show that, in addition to being competitive with reference-based attacks that have perfect knowledge about the training data distribution, our attack clearly outperforms existing reference-free attacks as well as reference-based attacks with imperfect knowledge, which demonstrates the need for a reevaluation of the threat model of adversarial attacks. |
2023.findings.xml | Findings of the Association for Computational Linguistics: EACL 2023 | 2023 | https://aclanthology.org/2023.findings-emnlp.30 | x | 0 | general safety, LLM alignment | null | Towards Agile Text Classifiers for Everyone | Text-based safety classifiers are widely used for content moderation and increasingly to tune generative language model behavior - a topic of growing concern for the safety of digital assistants and chatbots. However, different policies require different classifiers, and safety policies themselves improve from iteration and adaptation. This paper introduces and evaluates methods for agile text classification, whereby classifiers are trained using small, targeted datasets that can be quickly developed for a particular policy. Experimenting with 7 datasets from three safety-related domains, comprising 15 annotation schemes, led to our key finding: prompt-tuning large language models, like PaLM 62B, with a labeled dataset of as few as 80 examples can achieve state-of-the-art performance. We argue that this enables a paradigm shift for text classification, especially for models supporting safer online discourse. Instead of collecting millions of examples to attempt to create universal safety classifiers over months or years, classifiers could be tuned using small datasets, created by individuals or small organizations, tailored for specific use cases, and iterated on and adapted in the time-span of a day. |
2023.findings.xml | Findings of the Association for Computational Linguistics: EACL 2023 | 2023 | https://aclanthology.org/2023.findings-emnlp.88 | x | 0 | toxicity, bias | null | Toxicity in chatgpt: Analyzing persona-assigned language models | Large language models (LLMs) have shown incredible capabilities and transcended the natural language processing (NLP) community, with adoption throughout many services like healthcare, therapy, education, and customer service. Since users include people with critical information needs like students or patients engaging with chatbots, the safety of these systems is of prime importance. Legislation has recognized its significance and recently drafted a “Blueprint For An AI Bill Of Rights” which calls for domain experts to identify risks and potential impact of AI systems. To this end, we systematically evaluate toxicity in over half a million generations of ChatGPT, a popular dialogue-based LLM. We find that setting the system parameter of ChatGPT by assigning it a persona, say that of the boxer Muhammad Ali, significantly increases the toxicity of generations. Depending on the persona assigned to ChatGPT, its toxicity can increase up to 6\times, with outputs engaging in incorrect stereotypes, harmful dialogue, and hurtful opinions. Furthermore, we find concerning patterns where specific entities (e.g., certain races) are targeted more than others (3\times more) irrespective of the assigned persona, reflecting discriminatory biases in the model. Our findings show that multiple provisions in the legislative blueprint are being violated, and we hope that the broader AI community rethinks the efficacy of current safety guardrails and develops better techniques that lead to robust, safe, and trustworthy AI. |
2023.findings.xml | Findings of the Association for Computational Linguistics: EACL 2023 | 2023 | https://aclanthology.org/2023.findings-emnlp.129 | null | null | others | null | ExplainCPE: A Free-text Explanation Benchmark of Chinese Pharmacist Examination | In the field of Large Language Models (LLMs), researchers are increasingly exploring their effectiveness across a wide range of tasks. However, a critical area that requires further investigation is the interpretability of these models, particularly the ability to generate rational explanations for their decisions. Most existing explanation datasets are limited to the English language and the general domain, which leads to a scarcity of linguistic diversity and a lack of resources in specialized domains, such as medical. To mitigate this, we propose ExplainCPE, a challenging medical dataset consisting of over 7K problems from Chinese Pharmacist Examination, specifically tailored to assess the model-generated explanations. From the overall results, only GPT-4 passes the pharmacist examination with a 75.7% accuracy, while other models like ChatGPT fail. Further detailed analysis of LLM-generated explanations reveals the limitations of LLMs in understanding medical text and executing computational reasoning. With the increasing importance of AI safety and trustworthiness, ExplainCPE takes a step towards improving and evaluating the interpretability of LLMs in the medical domain. |
2023.findings.xml | Findings of the Association for Computational Linguistics: EACL 2023 | 2023 | https://aclanthology.org/2023.findings-emnlp.143 | x | 0 | jailbreaking attacks | null | Attack Prompt Generation for Red Teaming and Defending Large Language Models | Large language models (LLMs) are susceptible to red teaming attacks, which can induce LLMs to generate harmful content. Previous research constructs attack prompts via manual or automatic methods, which have their own limitations on construction cost and quality. To address these issues, we propose an integrated approach that combines manual and automatic methods to economically generate high-quality attack prompts. Specifically, considering the impressive capabilities of newly emerged LLMs, we propose an attack framework to instruct LLMs to mimic human-generated prompts through in-context learning. Furthermore, we propose a defense framework that fine-tunes victim LLMs through iterative interactions with the attack framework to enhance their safety against red teaming attacks. Extensive experiments on different LLMs validate the effectiveness of our proposed attack and defense frameworks. Additionally, we release a series of attack prompts datasets named SAP with varying sizes, facilitating the safety evaluation and enhancement of more LLMs. |
2023.findings.xml | Findings of the Association for Computational Linguistics: EACL 2023 | 2023 | https://aclanthology.org/2023.findings-emnlp.272 | x | 0 | jailbreaking attacks | null | Multi-step Jailbreaking Privacy Attacks on ChatGPT | With the rapid progress of large language models (LLMs), many downstream NLP tasks can be well solved given appropriate prompts. Though model developers and researchers work hard on dialog safety to avoid generating harmful content from LLMs, it is still challenging to steer AI-generated content (AIGC) for the human good. As powerful LLMs are devouring existing text data from various domains (e.g., GPT-3 is trained on 45TB texts), it is natural to doubt whether the private information is included in the training data and what privacy threats can these LLMs and their downstream applications bring. In this paper, we study the privacy threats from OpenAI’s ChatGPT and the New Bing enhanced by ChatGPT and show that application-integrated LLMs may cause new privacy threats. To this end, we conduct extensive experiments to support our claims and discuss LLMs’ privacy implications. |
2023.findings.xml | Findings of the Association for Computational Linguistics: EACL 2023 | 2023 | https://aclanthology.org/2023.findings-emnlp.365 | x | 0 | toxicity, bias | null | HARE: Explainable Hate Speech Detection with Step-by-Step Reasoning | With the proliferation of social media, accurate detection of hate speech has become critical to ensure safety online. To combat nuanced forms of hate speech, it is important to identify and thoroughly explain hate speech to help users understand its harmful effects. Recent benchmarks have attempted to tackle this issue by training generative models on free-text annotations of implications in hateful text. However, we find significant reasoning gaps in the existing annotations schemes, which may hinder the supervision of detection models. In this paper, we introduce a hate speech detection framework, **HARE**, which harnesses the reasoning capabilities of large language models (LLMs) to fill these gaps in explanations of hate speech, thus enabling effective supervision of detection models. Experiments on SBIC and Implicit Hate benchmarks show that our method, using model-generated data, consistently outperforms baselines, using existing free-text human annotations. Analysis demonstrates that our method enhances the explanation quality of trained models and improves generalization to unseen datasets. Our code is available at https://github.com/joonkeekim/hare-hate-speech.git. |
2023.findings.xml | Findings of the Association for Computational Linguistics: EACL 2023 | 2023 | https://aclanthology.org/2023.findings-emnlp.388 | null | 0 | general safety, LLM alignment | null | ASSERT: Automated Safety Scenario Red Teaming for Evaluating the Robustness of Large Language Models | As large language models are integrated into society, robustness toward a suite of prompts is increasingly important to maintain reliability in a high-variance environment.Robustness evaluations must comprehensively encapsulate the various settings in which a user may invoke an intelligent system. This paper proposes ASSERT, Automated Safety Scenario Red Teaming, consisting of three methods – semantically aligned augmentation, target bootstrapping, and adversarial knowledge injection. For robust safety evaluation, we apply these methods in the critical domain of AI safety to algorithmically generate a test suite of prompts covering diverse robustness settings – semantic equivalence, related scenarios, and adversarial. We partition our prompts into four safety domains for a fine-grained analysis of how the domain affects model performance. Despite dedicated safeguards in existing state-of-the-art models, we find statistically significant performance differences of up to 11% in absolute classification accuracy among semantically related scenarios and error rates of up to 19% absolute error in zero-shot adversarial settings, raising concerns for users’ physical safety. |
2023.findings.xml | Findings of the Association for Computational Linguistics: EACL 2023 | 2023 | https://aclanthology.org/2023.findings-emnlp.700 | x | 0 | general safety, LLM alignment | null | InstructSafety: A Unified Framework for Building Multidimensional and Explainable Safety Detector through Instruction Tuning | Safety detection has been an increasingly important topic in recent years and it has become even more necessary to develop reliable safety detection systems with the rapid development of large language models. However, currently available safety detection systems have limitations in terms of their versatility and interpretability. In this paper, we first introduce InstructSafety, a safety detection framework that unifies 7 common sub-tasks for safety detection. These tasks are unified into a similar form through different instructions. We then conduct a comprehensive survey of existing safety detection datasets and process 39 human-annotated datasets for instruction tuning. We also construct adversarial samples to enhance the model’s robustness. After fine-tuning Flan-T5 on the collected data, we have developed Safety-Flan-T5, a multidimensional and explainable safety detector. We conduct comprehensive experiments on a variety of datasets and tasks, and demonstrate the strong performance of Safety-Flan-T5 in comparison to supervised baselines and served APIs (Perspective API, ChatGPT and InstructGPT). We will release the processed data, fine-tuned Safety-Flan-T5 and related code for public use. |
2023.findings.xml | Findings of the Association for Computational Linguistics: EACL 2023 | 2023 | https://aclanthology.org/2023.findings-emnlp.716 | x | 0 | jailbreaking attacks | null | Attention-Enhancing Backdoor Attacks Against BERT-based Models | Recent studies have revealed that Backdoor Attacks can threaten the safety of natural language processing (NLP) models. Investigating the strategies of backdoor attacks will help to understand the model’s vulnerability. Most existing textual backdoor attacks focus on generating stealthy triggers or modifying model weights. In this paper, we directly target the interior structure of neural networks and the backdoor mechanism. We propose a novel Trojan Attention Loss (TAL), which enhances the Trojan behavior by directly manipulating the attention patterns. Our loss can be applied to different attacking methods to boost their attack efficacy in terms of attack successful rates and poisoning rates. It applies to not only traditional dirty-label attacks, but also the more challenging clean-label attacks. We validate our method on different backbone models (BERT, RoBERTa, and DistilBERT) and various tasks (Sentiment Analysis, Toxic Detection, and Topic Classification). |
2023.findings.xml | Findings of the Association for Computational Linguistics: EACL 2023 | 2023 | https://aclanthology.org/2023.findings-emnlp.796 | x | 0 | general safety, LLM alignment | null | Using In-Context Learning to Improve Dialogue Safety | While large neural-based conversational models have become increasingly proficient dialogue agents, recent work has highlighted safety issues with these systems. For example, these systems can be goaded into generating toxic content, often perpetuating social biases or stereotypes. We investigate a retrieval-based approach for reducing bias and toxicity in responses from chatbots. It uses in-context learning to steer a model towards safer generations. Concretely, to generate a response to an unsafe dialogue context, we retrieve demonstrations of safe responses to similar dialogue contexts. We find our method performs competitively with existing approaches to dialogue safety without requiring training. We also show, using automatic and human evaluation, that reductions in toxicity obtained using our approach are not at the cost engagingness or coherency. Finally, we note our method can be used in compliment to existing dialogue safety approaches, such as RLHF. |
2023.findings.xml | Findings of the Association for Computational Linguistics: EACL 2023 | 2023 | https://aclanthology.org/2023.findings-emnlp.896 | null | null | others | null | BioDEX: Large-Scale Biomedical Adverse Drug Event Extraction for Real-World Pharmacovigilance | Timely and accurate extraction of Adverse Drug Events (ADE) from biomedical literature is paramount for public safety, but involves slow and costly manual labor. We set out to improve drug safety monitoring (pharmacovigilance, PV) through the use of Natural Language Processing (NLP). We introduce BioDEX, a large-scale resource for Biomedical adverse Drug Event eXtraction, rooted in the historical output of drug safety reporting in the U.S. BioDEX consists of 65k abstracts and 19k full-text biomedical papers with 256k associated document-level safety reports created by medical experts. The core features of these reports include the reported weight, age, and biological sex of a patient, a set of drugs taken by the patient, the drug dosages, the reactions experienced, and whether the reaction was life threatening. In this work, we consider the task of predicting the core information of the report given its originating paper. We estimate human performance to be 72.0% F1, whereas our best model achieves 59.1% F1 (62.3 validation), indicating significant headroom. We also begin to explore ways in which these models could help professional PV reviewers. Our code and data are available at https://github.com/KarelDO/BioDEX. |
2023.findings.xml | Findings of the Association for Computational Linguistics: EACL 2023 | 2023 | https://aclanthology.org/2023.findings-emnlp.935 | x | 0 | general safety, LLM alignment | null | DialGuide: Aligning Dialogue Model Behavior with Developer Guidelines | Dialogue models are able to generate coherent and fluent responses, but they can still be challenging to control and may produce non-engaging, unsafe results. This unpredictability diminishes user trust and can hinder the use of the models in the real world. To address this, we introduce DialGuide, a novel framework for controlling dialogue model behavior using natural language rules, or guidelines. These guidelines provide information about the context they are applicable to and what should be included in the response, allowing the models to generate responses that are more closely aligned with the developer’s expectations and intent. We evaluate DialGuide on three tasks in open-domain dialogue response generation: guideline selection, response generation, and response entailment verification. Our dataset contains 10,737 positive and 15,467 negative dialogue context-response-guideline triplets across two domains - chit-chat and safety. We provide baseline models for the tasks and benchmark their performance. We also demonstrate that DialGuide is effective in the dialogue safety domain, producing safe and engaging responses that follow developer guidelines. |
2023.genbench.xml | Proceedings of the 1st GenBench Workshop on (Benchmarking) Generalisation in NLP | 2023 | https://aclanthology.org/2023.genbench-1.8 | null | 0 | hallucination, factuality | null | Walking a Tightrope – Evaluating Large Language Models in High-Risk Domains | High-risk domains pose unique challenges that require language models to provide accurate and safe responses. Despite the great success of large language models (LLMs), such as ChatGPT and its variants, their performance in high-risk domains remains unclear. Our study delves into an in-depth analysis of the performance of instruction-tuned LLMs, focusing on factual accuracy and safety adherence. To comprehensively assess the capabilities of LLMs, we conduct experiments on six NLP datasets including question answering and summarization tasks within two high-risk domains: legal and medical. Further qualitative analysis highlights the existing limitations inherent in current LLMs when evaluating in high-risk domains. This underscores the essential nature of not only improving LLM capabilities but also prioritizing the refinement of domain-specific metrics, and embracing a more human-centric approach to enhance safety and factual reliability. Our findings advance the field toward the concerns of properly evaluating LLMs in high-risk domains, aiming to steer the adaptability of LLMs in fulfilling societal obligations and aligning with forthcoming regulations, such as the EU AI Act. |
2023.ccl.xml | Proceedings of the 22nd Chinese National Conference on Computational Linguistics | 2023 | https://aclanthology.org/2023.ccl-2.8 | null | null | others | null | Through the Lens of Core Competency: Survey on Evaluation of Large Language Models | “From pre-trained language model (PLM) to large language model (LLM), the field of naturallanguage processing (NLP) has witnessed steep performance gains and wide practical uses. Theevaluation of a research field guides its direction of improvement. However, LLMs are extremelyhard to thoroughly evaluate for two reasons. First of all, traditional NLP tasks become inade-quate due to the excellent performance of LLM. Secondly, existing evaluation tasks are difficultto keep up with the wide range of applications in real-world scenarios. To tackle these problems,existing works proposed various benchmarks to better evaluate LLMs. To clarify the numerousevaluation tasks in both academia and industry, we investigate multiple papers concerning LLMevaluations. We summarize 4 core competencies of LLM, including reasoning, knowledge, relia-bility, and safety. For every competency, we introduce its definition, corresponding benchmarks,and metrics. Under this competency architecture, similar tasks are combined to reflect corre-sponding ability, while new tasks can also be easily added into the system. Finally, we give oursuggestions on the future direction of LLM’s evaluation.” |
2023.ccl.xml | Proceedings of the 22nd Chinese National Conference on Computational Linguistics | 2023 | https://aclanthology.org/2023.ccl-4.2 | null | null | others | null | Safety and Ethical Concerns of Large Language Models | “Recent months have witnessed significant progress in the field of large language models (LLMs).Represented by ChatGPT and GPT-4, LLMs perform well in various natural language process-ing tasks and have been applied to many downstream applications to facilitate people’s lives. However, there still exist safety and ethical concerns. Specifically, LLMs suffer from social bias,robustness problems, and poisoning issues, all of which may induce LLMs to spew harmful con-tents. We propose this tutorial as a gentle introduction to the safety and ethical issues of LLMs.” |
2023.alta.xml | Proceedings of the 21st Annual Workshop of the Australasian Language Technology Association | 2023 | https://aclanthology.org/2023.alta-1.8 | null | null | others | null | Catching Misdiagnosed Limb Fractures in the Emergency Department Using Cross-institution Transfer Learning | We investigated the development of a Machine Learning (ML)-based classifier to identify abnormalities in radiology reports from Emergency Departments (EDs) that can help automate the radiology report reconciliation process. Often, radiology reports become available to the ED only after the patient has been treated and discharged, following ED clinician interpretation of the X-ray. However, occasionally ED clinicians misdiagnose or fail to detect subtle abnormalities on X-rays, so they conduct a manual radiology report reconciliation process as a safety net. Previous studies addressed this problem of automated reconciliation using ML-based classification solutions that require data samples from the target institution that is heavily based on feature engineering, implying lower transferability between hospitals. In this paper, we investigated the benefits of using pre-trained BERT models for abnormality classification in a cross-institutional setting where data for fine-tuning was unavailable from the target institution. We also examined how the inclusion of synthetically generated radiology reports from ChatGPT affected the performance of the BERT models. Our findings suggest that BERT-like models outperform previously proposed ML-based methods in cross-institutional scenarios, and that adding ChatGPT-generated labelled radiology reports can improve the classifier’s performance by reducing the number of misdiagnosed discharged patients. |
2023.tacl.xml | Transactions of the Association for Computational Linguistics, Volume 11 | 2023 | https://aclanthology.org/2023.tacl-1.69 | null | 0 | others | null | Calibrated Interpretation: Confidence Estimation in Semantic Parsing | Sequence generation models are increasingly being used to translate natural language into programs, i.e., to perform executable semantic parsing. The fact that semantic parsing aims to predict programs that can lead to executed actions in the real world motivates developing safe systems. This in turn makes measuring calibration—a central component to safety—particularly important. We investigate the calibration of popular generation models across four popular semantic parsing datasets, finding that it varies across models and datasets. We then analyze factors associated with calibration error and release new confidence-based challenge splits of two parsing datasets. To facilitate the inclusion of calibration in semantic parsing evaluations, we release a library for computing calibration metrics.1 |
2023.tacl.xml | Transactions of the Association for Computational Linguistics, Volume 11 | 2023 | https://aclanthology.org/2023.tacl-1.85 | null | 1 | hallucination, factuality | Afrikaans, Arabic, Asturian, Bengali, Czech, Welsh, German, Greek, Spanish, Persian, Finnish, French, Hebrew, Hindi, Croatian, Hungarian, Indonesian, Japanese, Korean, Lithuanian, Dutch, Occitan, Polish, Portuguese, Russian, Swedish, Swahili, Tagalog, Turkish, Vietnamese, Chinese | Hallucinations in Large Multilingual Translation Models | Hallucinated translations can severely undermine and raise safety issues when machine translation systems are deployed in the wild. Previous research on the topic focused on small bilingual models trained on high-resource languages, leaving a gap in our understanding of hallucinations in multilingual models across diverse translation scenarios. In this work, we fill this gap by conducting a comprehensive analysis—over 100 language pairs across various resource levels and going beyond English-centric directions—on both the M2M neural machine translation (NMT) models and GPT large language models (LLMs). Among several insights, we highlight that models struggle with hallucinations primarily in low-resource directions and when translating out of English, where, critically, they may reveal toxic patterns that can be traced back to the training data. We also find that LLMs produce qualitatively different hallucinations to those of NMT models. Finally, we show that hallucinations are hard to reverse by merely scaling models trained with the same data. However, employing more diverse models, trained on different data or with different procedures, as fallback systems can improve translation quality and virtually eliminate certain pathologies. |
2023.acl.xml | Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) | 2023 | https://aclanthology.org/2023.acl-long.2 | null | 1 | general safety, LLM alignment | Chinese | SafeConv: Explaining and Correcting Conversational Unsafe Behavior | One of the main challenges open-domain end-to-end dialogue systems, or chatbots, face is the prevalence of unsafe behavior, such as toxic languages and harmful suggestions. However, existing dialogue datasets do not provide enough annotation to explain and correct such unsafe behavior. In this work, we construct a new dataset called SafeConv for the research of conversational safety: (1) Besides the utterance-level safety labels, SafeConv also provides unsafe spans in an utterance, information able to indicate which words contribute to the detected unsafe behavior; (2) SafeConv provides safe alternative responses to continue the conversation when unsafe behavior detected, guiding the conversation to a gentle trajectory. By virtue of the comprehensive annotation of SafeConv, we benchmark three powerful models for the mitigation of conversational unsafe behavior, including a checker to detect unsafe utterances, a tagger to extract unsafe spans, and a rewriter to convert an unsafe response to a safe version. Moreover, we explore the huge benefits brought by combining the models for explaining the emergence of unsafe behavior and detoxifying chatbots. Experiments show that the detected unsafe behavior could be well explained with unsafe spans and popular chatbots could be detoxified by a huge extent. The dataset is available at https://github.com/mianzhang/SafeConv. |
2023.acl.xml | Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) | 2023 | https://aclanthology.org/2023.acl-long.262 | null | null | others | null | Reward Gaming in Conditional Text Generation | To align conditional text generation model outputs with desired behaviors, there has been an increasing focus on training the model using reinforcement learning (RL) with reward functions learned from human annotations. Under this framework, we identify three common cases where high rewards are incorrectly assigned to undesirable patterns: noise-induced spurious correlation, naturally occurring spurious correlation, and covariate shift. We show that even though learned metrics achieve high performance on the distribution of the data used to train the reward function, the undesirable patterns may be amplified during RL training of the text generation model. While there has been discussion about reward gaming in the RL or safety community, in this discussion piece, we would like to highlight reward gaming in the natural language generation (NLG) community using concrete conditional text generation examples and discuss potential fixes and areas for future work. |
2023.acl.xml | Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) | 2023 | https://aclanthology.org/2023.acl-long.757 | null | 1 | toxicity, bias | null | CHBias: Bias Evaluation and Mitigation of Chinese Conversational Language Models | redWarning: This paper contains content that may be offensive or upsetting.Pretrained conversational agents have been exposed to safety issues, exhibiting a range of stereotypical human biases such as gender bias. However, there are still limited bias categories in current research, and most of them only focus on English. In this paper, we introduce a new Chinese dataset, CHBias, for bias evaluation and mitigation of Chinese conversational language models.Apart from those previous well-explored bias categories, CHBias includes under-explored bias categories, such as ageism and appearance biases, which received less attention. We evaluate two popular pretrained Chinese conversational models, CDial-GPT and EVA2.0, using CHBias. Furthermore, to mitigate different biases, we apply several debiasing methods to the Chinese pretrained models. Experimental results show that these Chinese pretrained models are potentially risky for generating texts that contain social biases, and debiasing methods using the proposed dataset can make response generation less biased while preserving the models’ conversational capabilities. |
2023.acl.xml | Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) | 2023 | https://aclanthology.org/2023.acl-short.72 | null | null | others | null | Using contradictions improves question answering systems | This work examines the use of contradiction in natural language inference (NLI) for question answering (QA). Typically, NLI systems help answer questions by determining if a potential answer is entailed (supported) by some background context. But is it useful to also determine if an answer contradicts the context? We test this in two settings, multiple choice and extractive QA, and find that systems that incorporate contradiction can do slightly better than entailment-only systems on certain datasets. However, the best performances come from using contradiction, entailment, and QA model confidence scores together. This has implications for the deployment of QA systems in domains such as medicine and science where safety is an issue. |
2023.cl.xml | Computational Linguistics, Volume 49, Issue 1 - March 2023 | 2023 | https://aclanthology.org/2023.cl-4.2 | x | 0 | hallucination, factuality | null | Measuring Attribution in Natural Language Generation Models | Large neural models have brought a new challenge to natural language generation (NLG): It has become imperative to ensure the safety and reliability of the output of models that generate freely. To this end, we present an evaluation framework, Attributable to Identified Sources (AIS), stipulating that NLG output pertaining to the external world is to be verified against an independent, provided source. We define AIS and a two-stage annotation pipeline for allowing annotators to evaluate model output according to annotation guidelines. We successfully validate this approach on generation datasets spanning three tasks (two conversational QA datasets, a summarization dataset, and a table-to-text dataset). We provide full annotation guidelines in the appendices and publicly release the annotated data at https://github.com/google-research-datasets/AIS. |
2023.sigdial.xml | Proceedings of the 24th Annual Meeting of the Special Interest Group on Discourse and Dialogue | 2023 | https://aclanthology.org/2023.sigdial-1.22 | null | null | others | null | Multi-party Goal Tracking with LLMs: Comparing Pre-training, Fine-tuning, and Prompt Engineering | This paper evaluates the extent to which current LLMs can capture task-oriented multi-party conversations (MPCs). We have recorded and transcribed 29 MPCs between patients, their companions, and a social robot in a hospital. We then annotated this corpus for multi-party goal-tracking and intent-slot recognition. People share goals, answer each other’s goals, and provide other people’s goals in MPCs - none of which occur in dyadic interactions. To understand user goals in MPCs, we compared three methods in zero-shot and few-shot settings: we fine-tuned T5, created pre-training tasks to train DialogLM using LED, and employed prompt engineering techniques with GPT-3.5-turbo, to determine which approach can complete this novel task with limited data. GPT-3.5-turbo significantly outperformed the others in a few-shot setting. The ‘reasoning’ style prompt, when given 7% of the corpus as example annotated conversations, was the best performing method. It correctly annotated 62.32% of the goal tracking MPCs, and 69.57% of the intent-slot recognition MPCs. A ‘story’ style prompt increased model hallucination, which could be detrimental if deployed in safety-critical settings. We conclude that multi-party conversations still challenge state-of-the-art LLMs. |
2023.ranlp.xml | Proceedings of the 14th International Conference on Recent Advances in Natural Language Processing | 2023 | https://aclanthology.org/2023.ranlp-1.41 | null | 0 | general safety, LLM alignment | null | Deep Learning Approaches to Detecting Safeguarding Concerns in Schoolchildren’s Online Conversations | For school teachers and Designated Safeguarding Leads (DSLs), computers and other school-owned communication devices are both indispensable and deeply worrisome. For their education, children require access to the Internet, as well as a standard institutional ICT infrastructure, including e-mail and other forms of online communication technology. Given the sheer volume of data being generated and shared on a daily basis within schools, most teachers and DSLs can no longer monitor the safety and wellbeing of their students without the use of specialist safeguarding software. In this paper, we experiment with the use of state-of-the-art neural network models on the modelling of a dataset of almost 9,000 anonymised child-generated chat messages on the Microsoft Teams platform. The data was manually classified into eight fine-grained classes of safeguarding concerns (or false alarms) that a monitoring program would be interested in, and these were further split into two binary classes: true positives (real safeguarding concerns) and false positives (false alarms). For the fine grained classification, our models achieved a macro F1 score of 73.56, while for the binary classification, we achieved a macro F1 score of 87.32. This first experiment into the use of Deep Learning for detecting safeguarding concerns represents an important step towards achieving high-accuracy and reliable monitoring information for busy teachers and safeguarding leads. |
2023.icon.xml | Proceedings of the 20th International Conference on Natural Language Processing (ICON) | 2023 | https://aclanthology.org/2023.icon-1.20 | null | 0 | others | null | KITLM: Domain-Specific Knowledge InTegration into Language Models for Question Answering | Large language models (LLMs) have demon- strated remarkable performance in a wide range of natural language tasks. However, as these models continue to grow in size, they face sig- nificant challenges in terms of computational costs. Additionally, LLMs often lack efficient domain-specific understanding, which is par- ticularly crucial in specialized fields such as aviation and healthcare. To boost the domain- specific understanding, we propose, KITLM 1 , a novel knowledge base integration approach into language model through relevant informa- tion infusion. By integrating pertinent knowl- edge, not only the performance of the lan- guage model is greatly enhanced, but the model size requirement is also significantly reduced while achieving comparable performance. Our proposed knowledge-infused model surpasses the performance of both GPT-3.5-turbo and the state-of-the-art knowledge infusion method, SKILL, achieving over 1.5 times improvement in exact match scores on the MetaQA. KITLM showed a similar performance boost in the avi- ation domain with AeroQA. The drastic perfor- mance improvement of KITLM over the exist- ing methods can be attributed to the infusion of relevant knowledge while mitigating noise. In addition, we release two curated datasets to accelerate knowledge infusion research in specialized fields: a) AeroQA, a new bench- mark dataset designed for multi-hop question- answering within the aviation domain, and b) Aviation Corpus, a dataset constructed from unstructured text extracted from the National Transportation Safety Board reports. Our re- search contributes to advancing the field of domain-specific language understanding and showcases the potential of knowledge infusion techniques in improving the performance. |
2023.yrrsds.xml | Proceedings of the 19th Annual Meeting of the Young Reseachers' Roundtable on Spoken Dialogue Systems | 2023 | https://aclanthology.org/2023.yrrsds-1.2 | x | 0 | general safety, LLM alignment | null | Safety and Robustness in Conversational AI | In this position paper, I will present the research interests in my PostDoc on safety and robustness specific to conversational AI, including then relevant overlap from my PhD. |
2023.yrrsds.xml | Proceedings of the 19th Annual Meeting of the Young Reseachers' Roundtable on Spoken Dialogue Systems | 2023 | https://aclanthology.org/2023.yrrsds-1.4 | x | 0 | general safety, LLM alignment | null | Advancing Spoken Dialog Systems for Manufacturing: From Conceptual Architecture and Taxonomy to Real Case Applications and Future Directions | This research encompasses a comprehensive exploration of Spoken Dialogue Systems (SDSs) in the manufacturing sector. It begins by establishing a conceptual architecture and taxonomy to guide the design and selection of SDS elements. Real case applications, including worker safety and cybersecurity support, validate the research findings and highlight areas for improvement. Looking ahead, the study delves into the potential of Large Language Models (LLMs) and multi-modal applications. Emphasizing the importance of extreme personalization, the study highlights the need to cater to the diverse qualifications and preferences of workers. Additionally, it investigates the integration of SDSs with other sensory modalities, such as images, videos, and augmented or virtual reality scenarios, to enhance the user experience and productivity. The research also addresses crucial considerations related to knowledge base optimization. It examines semantic variations of words across different application contexts, the continuous updating of procedures and data, and the adaptability of SDSs to diverse dialects and linguistic abilities, particularly in low-schooling personnel scenarios. Privacy, industrial protection, and ethical concerns in the era of LLMs and external players like OpenAI are given due attention. The study explores the boundaries of knowledge that conversational systems should possess, advocating for transparency, explainability, and responsible data handling practices. |
2023.emnlp.xml | Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing | 2023 | https://aclanthology.org/2023.emnlp-main.84 | x | 0 | toxicity, bias | null | Unveiling the Implicit Toxicity in Large Language Models | The open-endedness of large language models (LLMs) combined with their impressive capabilities may lead to new safety issues when being exploited for malicious use. While recent studies primarily focus on probing toxic outputs that can be easily detected with existing toxicity classifiers, we show that LLMs can generate diverse implicit toxic outputs that are exceptionally difficult to detect via simply zero-shot prompting. Moreover, we propose a reinforcement learning (RL) based attacking method to further induce the implicit toxicity in LLMs. Specifically, we optimize the language model with a reward that prefers implicit toxic outputs to explicit toxic and non-toxic ones. Experiments on five widely-adopted toxicity classifiers demonstrate that the attack success rate can be significantly improved through RL fine-tuning. For instance, the RL-finetuned LLaMA-13B model achieves an attack success rate of 90.04% on BAD and 62.85% on Davinci003. Our findings suggest that LLMs pose a significant threat in generating undetectable implicit toxic outputs. We further show that fine-tuning toxicity classifiers on the annotated examples from our attacking method can effectively enhance their ability to detect LLM-generated implicit toxic language. |
2023.emnlp.xml | Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing | 2023 | https://aclanthology.org/2023.emnlp-main.159 | null | 0 | general safety, LLM alignment | null | Did You Mean...? Confidence-based Trade-offs in Semantic Parsing | We illustrate how a calibrated model can help balance common trade-offs in task-oriented parsing. In a simulated annotator-in-the-loop experiment, we show that well-calibrated confidence scores allow us to balance cost with annotator load, improving accuracy with a small number of interactions. We then examine how confidence scores can help optimize the trade-off between usability and safety. We show that confidence-based thresholding can substantially reduce the number of incorrect low-confidence programs executed; however, this comes at a cost to usability. We propose the DidYouMean system which better balances usability and safety by rephrasing low-confidence inputs. |
2023.emnlp.xml | Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing | 2023 | https://aclanthology.org/2023.emnlp-main.539 | x | 0 | policy | null | Regulation and NLP (RegNLP): Taming Large Language Models | The scientific innovation in Natural Language Processing (NLP) and more broadly in artificial intelligence (AI) is at its fastest pace to date. As large language models (LLMs) unleash a new era of automation, important debates emerge regarding the benefits and risks of their development, deployment and use. Currently, these debates have been dominated by often polarized narratives mainly led by the AI Safety and AI Ethics movements. This polarization, often amplified by social media, is swaying political agendas on AI regulation and governance and posing issues of regulatory capture. Capture occurs when the regulator advances the interests of the industry it is supposed to regulate, or of special interest groups rather than pursuing the general public interest. Meanwhile in NLP research, attention has been increasingly paid to the discussion of regulating risks and harms. This often happens without systematic methodologies or sufficient rooting in the disciplines that inspire an extended scope of NLP research, jeopardizing the scientific integrity of these endeavors. Regulation studies are a rich source of knowledge on how to systematically deal with risk and uncertainty, as well as with scientific evidence, to evaluate and compare regulatory options. This resource has largely remained untapped so far. In this paper, we argue how NLP research on these topics can benefit from proximity to regulatory studies and adjacent fields. We do so by discussing basic tenets of regulation, and risk and uncertainty, and by highlighting the shortcomings of current NLP discussions dealing with risk assessment. Finally, we advocate for the development of a new multidisciplinary research space on regulation and NLP (RegNLP), focused on connecting scientific knowledge to regulatory processes based on systematic methodologies. |
2023.emnlp.xml | Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing | 2023 | https://aclanthology.org/2023.emnlp-main.673 | null | 1 | hallucination, factuality | German, English | FaMeSumm: Investigating and Improving Faithfulness of Medical Summarization | Summaries of medical text shall be faithful by being consistent and factual with source inputs, which is an important but understudied topic for safety and efficiency in healthcare. In this paper, we investigate and improve faithfulness in summarization on a broad range of medical summarization tasks. Our investigation reveals that current summarization models often produce unfaithful outputs for medical input text. We then introduce FaMeSumm, a framework to improve faithfulness by fine-tuning pre-trained language models based on medical knowledge. FaMeSumm performs contrastive learning on designed sets of faithful and unfaithful summaries, and it incorporates medical terms and their contexts to encourage faithful generation of medical terms. We conduct comprehensive experiments on three datasets in two languages: health question and radiology report summarization datasets in English, and a patient-doctor dialogue dataset in Chinese. Results demonstrate that FaMeSumm is flexible and effective by delivering consistent improvements over mainstream language models such as BART, T5, mT5, and PEGASUS, yielding state-of-the-art performances on metrics for faithfulness and general quality. Human evaluation by doctors also shows that FaMeSumm generates more faithful outputs. Our code is available at https://github.com/psunlpgroup/FaMeSumm. |
2023.emnlp.xml | Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing | 2023 | https://aclanthology.org/2023.emnlp-main.959 | null | 1 | toxicity, bias | Bengali | BanglaAbuseMeme: A Dataset for Bengali Abusive Meme Classification | The dramatic increase in the use of social media platforms for information sharing has also fueled a steep growth in online abuse. A simple yet effective way of abusing individuals or communities is by creating memes, which often integrate an image with a short piece of text layered on top of it. Such harmful elements are in rampant use and are a threat to online safety. Hence it is necessary to develop efficient models to detect and flag abusive memes. The problem becomes more challenging in a low-resource setting (e.g., Bengali memes, i.e., images with Bengali text embedded on it) because of the absence of benchmark datasets on which AI models could be trained. In this paper we bridge this gap by building a Bengali meme dataset. To setup an effective benchmark we implement several baseline models for classifying abusive memes using this dataset. We observe that multimodal models that use both textual and visual information outperform unimodal models. Our best-performing model achieves a macro F1 score of 70.51. Finally, we perform a qualitative error analysis of the misclassified memes of the best-performing text-based, image-based and multimodal models. |
2023.emnlp.xml | Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing | 2023 | https://aclanthology.org/2023.emnlp-tutorial.5 | x | 0 | general safety, LLM alignment | null | Mitigating Societal Harms in Large Language Models | Numerous recent studies have highlighted societal harms that can be caused by language technologies deployed in the wild. While several surveys, tutorials, and workshops have discussed the risks of harms in specific contexts – e.g., detecting and mitigating gender bias in NLP models – no prior work has developed a unified typology of technical approaches for mitigating harms of language generation models. Our tutorial is based on a survey we recently wrote that proposes such a typology. We will provide an overview of potential social issues in language generation, including toxicity, social biases, misinformation, factual inconsistency, and privacy violations. Our primary focus will be on how to systematically identify risks, and how eliminate them at various stages of model development, from data collection, to model development, to inference/language generation. Through this tutorial, we aim to equip NLP researchers and engineers with a suite of practical tools for mitigating safety risks from pretrained language generation models. |
2023.emnlp.xml | Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing | 2023 | https://aclanthology.org/2023.emnlp-demo.20 | x | 0 | others | null | Gentopia.AI: A Collaborative Platform for Tool-Augmented LLMs | Augmented Language Models (ALMs) empower large language models with the ability to use tools, transforming them into intelligent agents for real-world interactions. However, most existing frameworks for ALMs, to varying degrees, are deficient in the following critical features: flexible customization, collaborative democratization, and holistic evaluation. This paper proposes Gentopia, a lightweight and extensible framework for ALMs. Gentopia allows the flexible customization of agents through simple configurations, seamlessly integrating various language models, task formats, prompting modules, and plugins into a unified paradigm. Furthermore, we establish Gentpool, a public platform enabling the registration and sharing of user-customized agents. Agents registered in Gentpool are composable such that they can be assembled together for agent collaboration, advancing the democratization of artificial intelligence. To ensure high-quality agents, Gentbench, an integral component of Gentpool, is designed to thoroughly evaluate user-customized agents across diverse aspects such as safety, robustness, efficiency, etc. We release Gentopia on Github and will continuously move forward. |
2023.emnlp.xml | Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing | 2023 | https://aclanthology.org/2023.emnlp-industry.37 | null | 0 | jailbreaking attacks | null | AART: AI-Assisted Red-Teaming with Diverse Data Generation for New LLM-powered Applications | Adversarially testing large language models (LLMs) is crucial for their safe and responsible deployment in practice. We introduce an AI-assisted approach for automated generation of adversarial evaluation datasets to test the safety of LLM generations on new downstream applications. We call it AART AI-assisted Red-Teaming - an automated alternative to current manual red-teaming efforts. AART offers a data generation and augmentation pipeline of reusable and customizable recipes that reduce significantly human effort and enable integration of adversarial testing earlier in new product development. AART generates evaluation datasets with high diversity of content characteristics critical for effective adversarial testing (e.g. sensitive and harmful concepts, specific to a wide range of cultural and geographic regions and application scenarios). The data generation is steered by AI-assisted recipes to define, scope and prioritize diversity within a new application context. This feeds into a structured LLM-generation process that scales up evaluation priorities. This provides transparency of developers evaluation intentions and enables quick adaptation to new use cases and newly discovered model weaknesses. Compared to some of the state-of-the-art tools AART shows promising results in terms of concept coverage and data quality. |
2023.emnlp.xml | Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing | 2023 | https://aclanthology.org/2023.emnlp-industry.49 | null | 1 | general safety, LLM alignment | Japanese | Towards Safer Operations: An Expert-involved Dataset of High-Pressure Gas Incidents for Preventing Future Failures | This paper introduces a new IncidentAI dataset for safety prevention. Different from prior corpora that usually contain a single task, our dataset comprises three tasks: named entity recognition, cause-effect extraction, and information retrieval. The dataset is annotated by domain experts who have at least six years of practical experience as high-pressure gas conservation managers. We validate the contribution of the dataset in the scenario of safety prevention. Preliminary results on the three tasks show that NLP techniques are beneficial for analyzing incident reports to prevent future failures. The dataset facilitates future research in NLP and incident management communities. The access to the dataset is also provided (The IncidentAI dataset is available at: https://github.com/Cinnamon/incident-ai-dataset). |
2023.emnlp.xml | Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing | 2023 | https://aclanthology.org/2023.emnlp-industry.76 | x | 0 | toxicity, bias | null | DELPHI: Data for Evaluating LLMs’ Performance in Handling Controversial Issues | Controversy is a reflection of our zeitgeist, and an important aspect to any discourse. The rise of large language models (LLMs) as conversational systems has increased public reliance on these systems for answers to their various questions. Consequently, it is crucial to systematically examine how these models respond to questions that pertaining to ongoing debates. However, few such datasets exist in providing human-annotated labels reflecting the contemporary discussions. To foster research in this area, we propose a novel construction of a controversial questions dataset, expanding upon the publicly released Quora Question Pairs Dataset. This dataset presents challenges concerning knowledge recency, safety, fairness, and bias. We evaluate different LLMs using a subset of this dataset, illuminating how they handle controversial issues and the stances they adopt. This research ultimately contributes to our understanding of LLMs’ interaction with controversial issues, paving the way for improvements in their comprehension and handling of complex societal debates. |
2024.tacl.xml | Transactions of the Association for Computational Linguistics, Volume 12 | 2024 | https://aclanthology.org/2024.tacl-1.64 | x | 0 | hallucination, factuality | null | Investigating Hallucinations in Pruned Large Language Models for Abstractive Summarization | Despite the remarkable performance of generative large language models (LLMs) on abstractive summarization, they face two significant challenges: their considerable size and tendency to hallucinate. Hallucinations are concerning because they erode reliability and raise safety issues. Pruning is a technique that reduces model size by removing redundant weights, enabling more efficient sparse inference. Pruned models yield downstream task performance comparable to the original, making them ideal alternatives when operating on a limited budget. However, the effect that pruning has upon hallucinations in abstractive summarization with LLMs has yet to be explored. In this paper, we provide an extensive empirical study across five summarization datasets, two state-of-the-art pruning methods, and five instruction-tuned LLMs. Surprisingly, we find that hallucinations are less prevalent from pruned LLMs than the original models. Our analysis suggests that pruned models tend to depend more on the source document for summary generation. This leads to a higher lexical overlap between the generated summary and the source document, which could be a reason for the reduction in hallucination risk.1 |
2024.nlp4science.xml | Proceedings of the 1st Workshop on NLP for Science (NLP4Science) | 2024 | https://aclanthology.org/2024.nlp4science-1.22 | x | 0 | general safety, LLM alignment | null | CogErgLLM: Exploring Large Language Model Systems Design Perspective Using Cognitive Ergonomics | Integrating cognitive ergonomics with LLMs is crucial for improving safety, reliability, and user satisfaction in human-AI interactions. Current LLM designs often lack this integration, resulting in systems that may not fully align with human cognitive capabilities and limitations. This oversight exacerbates biases in LLM outputs and leads to suboptimal user experiences due to inconsistent application of user-centered design principles. Researchers are increasingly leveraging NLP, particularly LLMs, to model and understand human behavior across social sciences, psychology, psychiatry, health, and neuroscience. Our position paper explores the need to integrate cognitive ergonomics into LLM design, providing a comprehensive framework and practical guidelines for ethical development. By addressing these challenges, we aim to advance safer, more reliable, and ethically sound human-AI interactions. |
2024.blackboxnlp.xml | Proceedings of the 7th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP | 2024 | https://aclanthology.org/2024.blackboxnlp-1.19 | null | 0 | others | null | Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2 | Sparse autoencoders (SAEs) are an unsupervised method for learning a sparse decomposition of a neural network’s latent representations into seemingly interpretable features.Despite recent excitement about their potential, research applications outside of industry are limited by the high cost of training a comprehensive suite of SAEs.In this work, we introduce Gemma Scope, an open suite of JumpReLU SAEs trained on all layers and sub-layers of Gemma 2 2B and 9B and select layers of Gemma 2 27B base models.We primarily train SAEs on the Gemma 2 pre-trained models, but additionally release SAEs trained on instruction-tuned Gemma 2 9B for comparison.We evaluate the quality of each SAE on standard metrics and release these results.We hope that by releasing these SAE weights, we can help make more ambitious safety and interpretability research easier for the community. Weights and a tutorial can be found at https://huggingface.co/google/gemma-scope and an interactive demo can be found at https://neuronpedia.org/gemma-scope. |
2024.blackboxnlp.xml | Proceedings of the 7th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP | 2024 | https://aclanthology.org/2024.blackboxnlp-1.26 | x | 0 | jailbreaking attacks | null | Pruning for Protection: Increasing Jailbreak Resistance in Aligned LLMs Without Fine-Tuning | This paper investigates the impact of model compression on the way Large Language Models (LLMs) process prompts, particularly concerning jailbreak resistance. We show that moderate WANDA pruning can enhance resistance to jailbreaking attacks without fine-tuning, while maintaining performance on standard benchmarks. To systematically evaluate this safety enhancement, we introduce a dataset of 225 harmful tasks across five categories. Our analysis of LLaMA-2 Chat, Vicuna 1.3, and Mistral Instruct v0.2 reveals that pruning benefits correlate with initial model safety levels. We interpret these results by examining changes in attention patterns and perplexity shifts, demonstrating that pruned models exhibit sharper attention and increased sensitivity to artificial jailbreak constructs. We extend our evaluation to the AdvBench harmful behavior tasks and the GCG attack method. We find that LLaMA-2 is much safer on AdvBench prompts than on our dataset when evaluated with manual jailbreak attempts, and that pruning is effective against both automated attacks and manual jailbreaking on Advbench. |
2024.blackboxnlp.xml | Proceedings of the 7th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP | 2024 | https://aclanthology.org/2024.blackboxnlp-1.29 | x | 0 | others | null | Investigating Layer Importance in Large Language Models | Large language models (LLMs) have gained increasing attention due to their prominent ability to understand and process texts. Nevertheless, LLMs largely remain opaque. The lack of understanding of LLMs has obstructed the deployment in safety-critical scenarios and hindered the development of better models. In this study, we advance the understanding of LLM by investigating the significance of individual layers in LLMs. We propose an efficient sampling method to faithfully evaluate the importance of layers using Shapley values, a widely used explanation framework in feature attribution and data valuation. In addition, we conduct layer ablation experiments to assess the performance degradation resulting from the exclusion of specific layers. Our findings reveal the existence of cornerstone layers, wherein certain early layers can exhibit a dominant contribution over others. Removing one cornerstone layer leads to a drastic collapse of the model performance, often reducing it to random guessing. Conversely, removing non-cornerstone layers results in only marginal performance changes. This study identifies cornerstone layers in LLMs and underscores their critical role for future research. |
2024.findings.xml | Findings of the Association for Computational Linguistics: EACL 2024 | 2024 | https://aclanthology.org/2024.findings-emnlp.64 | null | 1 | general safety, LLM alignment | English, Chinese | SAFETY-J: Evaluating Safety with Critique | The deployment of Large Language Models (LLMs) in content generation raises significant safety concerns, particularly regarding the transparency and interpretability of content evaluations. Current methods, primarily focused on binary safety classifications, lack mechanisms for detailed critique, limiting their utility for model improvement and user trust. To address these limitations, we introduce SAFETY-J, a bilingual generative safety evaluator for English and Chinese with critique-based judgment. SAFETY-J utilizes a robust training dataset that includes diverse dialogues and augmented query-response pairs to assess safety across various scenarios comprehensively. We establish an automated meta-evaluation benchmark that objectively assesses the quality of critiques with minimal human intervention, facilitating scalable and continuous improvement. Additionally, SAFETY-Jemploys an iterative preference learning technique to dynamically refine safety assessments based on meta-evaluations and critiques. Our evaluations demonstrate that SAFETY-J provides more nuanced and accurate safety evaluations, thereby enhancing both critique quality and predictive reliability in complex content scenarios. To facilitate further research and application, we have released SAFETY-J’s training protocols, datasets, and code at https://github.com/GAIR-NLP/Safety-J. |
2024.findings.xml | Findings of the Association for Computational Linguistics: EACL 2024 | 2024 | https://aclanthology.org/2024.findings-emnlp.79 | x | 0 | general safety, LLM alignment | null | R-Judge: Benchmarking Safety Risk Awareness for LLM Agents | Large language models (LLMs) have exhibited great potential in autonomously completing tasks across real-world applications. Despite this, these LLM agents introduce unexpected safety risks when operating in interactive environments. Instead of centering on the harmlessness of LLM-generated content in most prior studies, this work addresses the imperative need for benchmarking the behavioral safety of LLM agents within diverse environments. We introduce R-Judge, a benchmark crafted to evaluate the proficiency of LLMs in judging and identifying safety risks given agent interaction records. R-Judge comprises 569 records of multi-turn agent interaction, encompassing 27 key risk scenarios among 5 application categories and 10 risk types. It is of high-quality curation with annotated safety labels and risk descriptions. Evaluation of 11 LLMs on R-Judge shows considerable room for enhancing the risk awareness of LLMs: The best-performing model, GPT-4o, achieves 74.42% while no other models significantly exceed the random. Moreover, we reveal that risk awareness in open agent scenarios is a multi-dimensional capability involving knowledge and reasoning, thus challenging for LLMs. With further experiments, we find that fine-tuning on safety judgment significantly improve model performance while straightforward prompting mechanisms fail. R-Judge is publicly available at Annoymous. |
2024.findings.xml | Findings of the Association for Computational Linguistics: EACL 2024 | 2024 | https://aclanthology.org/2024.findings-emnlp.96 | null | 1 | general safety, LLM alignment | Catalan, French, Indonesian, Portuguese, Spanish, Chichewa, Swahili, Tamil, Telugu, Finnish, Hungarian, Chinese, Japanese, Korean, Vietnamese | Exploring Multilingual Concepts of Human Values in Large Language Models: Is Value Alignment Consistent, Transferable and Controllable across Languages? | Prior research has revealed that certain abstract concepts are linearly represented as directions in the representation space of LLMs, predominantly centered around English. In this paper, we extend this investigation to a multilingual context, with a specific focus on human values-related concepts (i.e., value concepts) due to their significance for AI safety. Through our comprehensive exploration covering 7 types of human values, 16 languages and 3 LLM series with distinct multilinguality (e.g., monolingual, bilingual and multilingual), we first empirically confirm the presence of value concepts within LLMs in a multilingual format. Further analysis on the cross-lingual characteristics of these concepts reveals 3 traits arising from language resource disparities: cross-lingual inconsistency, distorted linguistic relationships, and unidirectional cross-lingual transfer between high- and low-resource languages, all in terms of value concepts. Moreover, we validate the feasibility of cross-lingual control over value alignment capabilities of LLMs, leveraging the dominant language as a source language. Ultimately, recognizing the significant impact of LLMs’ multilinguality on our results, we consolidate our findings and provide prudent suggestions on the composition of multilingual data for LLMs pre-training. |
2024.findings.xml | Findings of the Association for Computational Linguistics: EACL 2024 | 2024 | https://aclanthology.org/2024.findings-emnlp.118 | x | 0 | others | null | Conditional Language Policy: A General Framework For Steerable Multi-Objective Finetuning | Reward-based finetuning is crucial for aligning language policies with intended behaviors (*e.g.*, creativity and safety). A key challenge is to develop steerable language models that trade-off multiple (conflicting) objectives in a flexible and efficient manner. This paper presents Conditional Language Policy (CLP), a general framework for finetuning language models on multiple objectives. Building on techniques from multi-task training and parameter-efficient finetuning, CLP learn steerable models that effectively trade-off conflicting objectives at *inference time*. Notably, this does not require training or maintaining multiple models to achieve different trade-offs between the objectives. Through extensive experiments and ablations on two summarization datasets, we show that CLP learns steerable language models that outperform and Pareto-dominate the existing approaches for multi-objective |
2024.findings.xml | Findings of the Association for Computational Linguistics: EACL 2024 | 2024 | 1 | null | 0 | general safety, LLM alignment | null | Can AI Relate: Testing Large Language Model Response for Mental Health Support | Large language models (LLMs) are already being piloted for clinical use in hospital systems like NYU Langone, Dana-Farber and the NHS. A proposed deployment use case is psychotherapy, where a LLM-powered chatbot can treat a patient undergoing a mental health crisis. Deployment of LLMs for mental health response could hypothetically broaden access to psychotherapy and provide new possibilities for personalizing care. However, recent high-profile failures, like damaging dieting advice offered by the Tessa chatbot to patients with eating disorders, have led to doubt about their reliability in high-stakes and safety-critical settings.In this work, we develop an evaluation framework for determining whether LLM response is a viable and ethical path forward for the automation of mental health treatment. Our framework measures equity in empathy and adherence of LLM responses to motivational interviewing theory. Using human evaluation with trained clinicians and automatic quality-of-care metrics grounded in psychology research, we compare the responses provided by peer-to-peer responders to those provided by a state-of-the-art LLM.We show that LLMs like GPT-4 use implicit and explicit cues to infer patient demographics like race. We then show that there are statistically significant discrepancies between patient subgroups: Responses to Black posters consistently have lower empathy than for any other demographic group (2%-13% lower than the control group). Promisingly, we do find that the manner in which responses are generated significantly impacts the quality of the response. We conclude by proposing safety guidelines for the potential deployment of LLMs for mental health response. |
2024.findings.xml | Findings of the Association for Computational Linguistics: EACL 2024 | 2024 | https://aclanthology.org/2024.findings-emnlp.139 | x | 0 | jailbreaking attacks | null | How Alignment and Jailbreak Work: Explain LLM Safety through Intermediate Hidden States | Large language models (LLMs) rely on safety alignment to avoid responding to malicious user inputs. Unfortunately, jailbreak can circumvent safety guardrails, resulting in LLMs generating harmful content and raising concerns about LLM safety. Due to language models with intensive parameters often regarded as black boxes, the mechanisms of alignment and jailbreak are challenging to elucidate. In this paper, we employ weak classifiers to explain LLM safety through the intermediate hidden states. We first confirm that LLMs learn ethical concepts during pre-training rather than alignment and can identify malicious and normal inputs in the early layers. Alignment actually associates the early concepts with emotion guesses in the middle layers and then refines them to the specific reject tokens for safe generations. Jailbreak disturbs the transformation of early unethical classification into negative emotions. We conduct experiments on models from 7B to 70B across various model families to prove our conclusion. Overall, our paper indicates the intrinsical mechanism of LLM safety and how jailbreaks circumvent safety guardrails, offering a new perspective on LLM safety and reducing concerns. |
2024.findings.xml | Findings of the Association for Computational Linguistics: EACL 2024 | 2024 | https://aclanthology.org/2024.findings-emnlp.188 | x | 0 | jailbreaking attacks | null | Be a Multitude to Itself: A Prompt Evolution Framework for Red Teaming | Large Language Models (LLMs) have gained increasing attention for their remarkable capacity, alongside concerns about safety arising from their potential to produce harmful content. Red teaming aims to find prompts that could elicit harmful responses from LLMs, and is essential to discover and mitigate safety risks before real-world deployment. However, manual red teaming is both time-consuming and expensive, rendering it unscalable. In this paper, we propose RTPE, a scalable evolution framework to evolve red teaming prompts across both breadth and depth dimensions, facilitating the automatic generation of numerous high-quality and diverse red teaming prompts. Specifically, in-breadth evolving employs a novel enhanced in-context learning method to create a multitude of quality prompts, whereas in-depth evolving applies customized transformation operations to enhance both content and form of prompts, thereby increasing diversity. Extensive experiments demonstrate that RTPE surpasses existing representative automatic red teaming methods on both attack success rate and diversity. In addition, based on 4,800 red teaming prompts created by RTPE, we further provide a systematic analysis of 8 representative LLMs across 8 sensitive topics. |
2024.findings.xml | Findings of the Association for Computational Linguistics: EACL 2024 | 2024 | https://aclanthology.org/2024.findings-emnlp.238 | x | 0 | general safety, LLM alignment | null | Self-Evolution Fine-Tuning for Policy Optimization | The alignment of large language models (LLMs) is crucial not only for unlocking their potential in specific tasks but also for ensuring that responses meet human expectations and adhere to safety and ethical principles. To address the challenges of current alignment methodologies, we introduce self-evolution fine-tuning (SEFT) for LLM alignment, aiming to eliminate the need for annotated samples while retaining the stability and efficiency of SFT. SEFT first trains an adaptive reviser to elevate low-quality responses while maintaining high-quality ones. The reviser then gradually guides the policy’s optimization by fine-tuning it with enhanced responses. The method excels in utilizing unlimited unannotated data to optimize policies via supervised fine-tuning. Our experiments on AlpacaEval and MT-Bench demonstrate the effectiveness of SEFT and its advantages over existing alignment techniques. |
2024.findings.xml | Findings of the Association for Computational Linguistics: EACL 2024 | 2024 | https://aclanthology.org/2024.findings-emnlp.293 | x | 0 | jailbreaking attacks | null | Defending Large Language Models Against Jailbreak Attacks via Layer-specific Editing | Large language models (LLMs) are increasingly being adopted in a wide range of real-world applications. Despite their impressive performance, recent studies have shown that LLMs are vulnerable to deliberately crafted adversarial prompts even when aligned via Reinforcement Learning from Human Feedback or supervised fine-tuning. While existing defense methods focus on either detecting harmful prompts or reducing the likelihood of harmful responses through various means, defending LLMs against jailbreak attacks based on the inner mechanisms of LLMs remains largely unexplored. In this work, we investigate how LLMs respond to harmful prompts and propose a novel defense method termed Layer-specific Editing (LED) to enhance the resilience of LLMs against jailbreak attacks. Through LED, we reveal that several critical safety layers exist among the early layers of LLMs. We then show that realigning these safety layers (and some selected additional layers) with the decoded safe response from identified toxic layers can significantly improve the alignment of LLMs against jailbreak attacks. Extensive experiments across various LLMs (e.g., Llama2, Mistral) show the effectiveness of LED, which effectively defends against jailbreak attacks while maintaining performance on benign prompts. Our code is available at https://github.com/ledllm/ledllm. |
2024.findings.xml | Findings of the Association for Computational Linguistics: EACL 2024 | 2024 | https://aclanthology.org/2024.findings-emnlp.296 | x | 0 | toxicity, bias | null | Rater Cohesion and Quality from a Vicarious Perspective | Human feedback is essential for building human-centered AI systems across domains where disagreement is prevalent, such as AI safety, content moderation, or sentiment analysis. Many disagreements, particularly in politically charged settings, arise because raters have opposing values or beliefs. Vicarious annotation is a method for breaking down disagreement by asking raters how they think others would annotate the data. In this paper, we explore the use of vicarious annotation with analytical methods for moderating rater disagreement. We employ rater cohesion metrics to study the potential influence of political affiliations and demographic backgrounds on raters’ perceptions of offense. Additionally, we utilize CrowdTruth’s rater quality metrics, which consider the demographics of the raters, to score the raters and their annotations. We study how the rater quality metrics influence the in-group and cross-group rater cohesion across the personal and vicarious levels. |
2024.findings.xml | Findings of the Association for Computational Linguistics: EACL 2024 | 2024 | https://aclanthology.org/2024.findings-emnlp.301 | null | 0 | jailbreaking attacks | null | Immunization against harmful fine-tuning attacks | Large Language Models (LLMs) are often trained with safety guards intended to prevent harmful text generation. However, such safety training can be removed by fine-tuning the LLM on harmful datasets. While this emerging threat (harmful fine-tuning attacks) has been characterized by previous work, there is little understanding of how we should proceed in constructing and validating defenses against these attacks especially in the case where defenders would not have control of the fine-tuning process. We introduce a formal framework based on the training budget of an attacker which we call “Immunization” conditions. Using a formal characterisation of the harmful fine-tuning problem, we provide a thorough description of what a successful defense must comprise of and establish a set of guidelines on how rigorous defense research that gives us confidence should proceed. |
2024.findings.xml | Findings of the Association for Computational Linguistics: EACL 2024 | 2024 | https://aclanthology.org/2024.findings-emnlp.331 | null | 0 | toxicity, bias | null | Gender Bias in Decision-Making with Large Language Models: A Study of Relationship Conflicts | Large language models (LLMs) acquire beliefs about gender from training data and can therefore generate text with stereotypical gender attitudes. Prior studies have demonstrated model generations favor one gender or exhibit stereotypes about gender, but have not investigated the complex dynamics that can influence model reasoning and decision-making involving gender. We study gender equity within LLMs through a decision-making lens with a new dataset, DeMET Prompts, containing scenarios related to intimate, romantic relationships. We explore nine relationship configurations through name pairs across three name lists (men, women, neutral). We investigate equity in the context of gender roles through numerous lenses: typical and gender-neutral names, with and without model safety enhancements, same and mixed-gender relationships, and egalitarian versus traditional scenarios across various topics. While all models exhibit the same biases (women favored, then those with gender-neutral names, and lastly men), safety guardrails reduce bias. In addition, models tend to circumvent traditional male dominance stereotypes and side with “traditionally female” individuals more often, suggesting relationships are viewed as a female domain by the models. |
2024.findings.xml | Findings of the Association for Computational Linguistics: EACL 2024 | 2024 | https://aclanthology.org/2024.findings-emnlp.574 | x | 0 | privacy | null | Can Textual Unlearning Solve Cross-Modality Safety Alignment? | Recent studies reveal that integrating new modalities into large language models (LLMs), such as vision-language models (VLMs), creates a new attack surface that bypasses existing safety training techniques like supervised fine-tuning (SFT) and reinforcement learning with human feedback (RLHF). While further SFT and RLHF-based safety training can be conducted in multi-modal settings, collecting multi-modal training datasets poses a significant challenge. Inspired by the structural design of recent multi-modal models, where all input modalities are ultimately fused into the language space, we explore whether unlearning solely in the textual domain can be effective for cross-modality safety alignment. Our empirical evaluation across seven datasets demonstrates promising transferability — textual unlearning in VLMs significantly reduces the Attack Success Rate (ASR) to less than 8% and in some cases, even as low as nearly 2% for both text-based and vision-text-based attacks, alongside preserving the utility. Moreover, our experiments show that unlearning with a multi-modal dataset offers no potential benefits but incurs significantly increased computational demands. |
2024.findings.xml | Findings of the Association for Computational Linguistics: EACL 2024 | 2024 | https://aclanthology.org/2024.findings-emnlp.585 | x | 0 | general safety, LLM alignment | null | TrustAgent: Towards Safe and Trustworthy LLM-based Agents | The rise of LLM-based agents shows great potential to revolutionize task planning, capturing significant attention. Given that these agents will be integrated into high-stake domains, ensuring their reliability and safety is crucial. This paper presents an Agent-Constitution-based agent framework, TrustAgent, with a particular focus on improving the LLM-based agent safety. The proposed framework ensures strict adherence to the Agent Constitution through three strategic components: pre-planning strategy which injects safety knowledge to the model before plan generation, in-planning strategy which enhances safety during plan generation, and post-planning strategy which ensures safety by post-planning inspection. Our experimental results demonstrate that the proposed framework can effectively enhance an LLM agent’s safety across multiple domains by identifying and mitigating potential dangers during the planning. Further analysis reveals that the framework not only improves safety but also enhances the helpfulness of the agent. Additionally, we highlight the importance of the LLM reasoning ability in adhering to the Constitution. This paper sheds light on how to ensure the safe integration of LLM-based agents into human-centric environments. Data and code are available at https://anonymous.4open.science/r/TrustAgent-06DC. |
2024.findings.xml | Findings of the Association for Computational Linguistics: EACL 2024 | 2024 | https://aclanthology.org/2024.findings-emnlp.610 | null | 1 | jailbreaking attacks | English, Chinese | ShieldLM: Empowering LLMs as Aligned, Customizable and Explainable Safety Detectors | The safety of Large Language Models (LLMs) has gained increasing attention in recent years, but there still lacks a comprehensive approach for detecting safety issues within LLMs’ responses in an aligned, customizable and explainable manner. In this paper, we propose ShieldLM, an LLM-based safety detector, which aligns with common safety standards, supports customizable detection rules, and provides explanations for its decisions. To train ShieldLM, we compile a large bilingual dataset comprising 14,387 query-response pairs, annotating the safety of responses based on various safety standards. Through extensive experiments, we demonstrate that ShieldLM surpasses strong baselines across four test sets, showcasing remarkable customizability and explainability. Besides performing well on standard detection datasets, ShieldLM has also been shown to be effective as a safety evaluator for advanced LLMs. ShieldLM is released at https://github.com/thu-coai/ShieldLM to support accurate and explainable safety detection under various safety standards. |
2024.findings.xml | Findings of the Association for Computational Linguistics: EACL 2024 | 2024 | https://aclanthology.org/2024.findings-emnlp.620 | null | 0 | general safety, LLM alignment | null | Interpretable Preferences via Multi-Objective Reward Modeling and Mixture-of-Experts | Reinforcement learning from human feedback (RLHF) has emerged as the primary method for aligning large language models (LLMs) with human preferences. The RLHF process typically starts by training a reward model (RM) using human preference data. Conventional RMs are trained on pairwise responses to the same user request, with relative ratings indicating which response humans prefer. The trained RM serves as a proxy for human preferences. However, due to the black-box nature of RMs, their outputs lack interpretability, as humans cannot intuitively understand why an RM thinks a response is good or not. As RMs act as human preference proxies, it is desirable for them to be human-interpretable to ensure that their internal decision processes are consistent with human preferences and to prevent reward hacking in LLM alignment. To build RMs with interpretable preferences, we propose a two-stage approach: i) train an Absolute-Rating Multi-Objective Reward Model (ArmoRM) with multi-dimensional absolute-rating data, each dimension corresponding to a human-interpretable objective (e.g., honesty, verbosity, safety); ii) employ a Mixture-of-Experts (MoE) strategy with a gating network that automatically selects the most suitable reward objectives based on the context. We efficiently trained an ArmoRM with Llama-3 8B and a gating network consisting of a shallow MLP on top of the ArmoRM. Our trained model, ArmoRM-Llama3-8B, obtains state-of-the-art performance on RewardBench, a benchmark evaluating RMs for language modeling. Notably, the performance of our model surpasses the LLM-as-a-judge method with GPT-4 judges by a margin, and approaches the performance of the much larger Nemotron-4 340B reward model. |
2024.findings.xml | Findings of the Association for Computational Linguistics: EACL 2024 | 2024 | https://aclanthology.org/2024.findings-emnlp.629 | x | 0 | toxicity, bias | null | Seeing Through VisualBERT: A Causal Adventure on Memetic Landscapes | Detecting offensive memes is crucial, yet standard deep neural network systems often remain opaque. Various input attribution-based methods attempt to interpret their behavior, but they face challenges with implicitly offensive memes and non-causal attributions. To address these issues, we propose a framework based on a Structural Causal Model (SCM). In this framework, VisualBERT is trained to predict the class of an input meme based on both meme input and causal concepts, allowing for transparent interpretation. Our qualitative evaluation demonstrates the framework’s effectiveness in understanding model behavior, particularly in determining whether the model was right due to the right reason, and in identifying reasons behind misclassification. Additionally, quantitative analysis assesses the significance of proposed modelling choices, such as de-confounding, adversarial learning, and dynamic routing, and compares them with input attribution methods. Surprisingly, we find that input attribution methods do not guarantee causality within our framework, raising questions about their reliability in safety-critical applications. The project page is at: https://newcodevelop.github.io/causality_adventure/ |
2024.findings.xml | Findings of the Association for Computational Linguistics: EACL 2024 | 2024 | https://aclanthology.org/2024.findings-emnlp.630 | null | 1 | general safety, LLM alignment | null | Cross-Lingual Unlearning of Selective Knowledge in Multilingual Language Models | Pretrained language models memorize vast amounts of information, including private and copyrighted data, raising significant safety concerns. Retraining these models after excluding sensitive data is prohibitively expensive, making machine unlearning a viable, cost-effective alternative. Previous research has focused on machine unlearning for monolingual models, but we find that unlearning in one language does not necessarily transfer to others. This vulnerability makes models susceptible to low-resource language attacks, where sensitive information remains accessible in less dominant languages. This paper presents a pioneering approach to machine unlearning for multilingual language models, selectively erasing information across different languages while maintaining overall performance. Specifically, our method employs an adaptive unlearning scheme that assigns language-dependent weights to address different language performances of multilingual language models. Empirical results demonstrate the effectiveness of our framework compared to existing unlearning baselines, setting a new standard for secure and adaptable multilingual language models. |
2024.findings.xml | Findings of the Association for Computational Linguistics: EACL 2024 | 2024 | https://aclanthology.org/2024.findings-emnlp.674 | null | 0 | general safety, LLM alignment | null | Ask the experts: sourcing a high-quality nutrition counseling dataset through Human-AI collaboration | Large Language Models (LLMs) are being employed by end-users for various tasks, including sensitive ones such as health counseling, disregarding potential safety concerns. It is thus necessary to understand how adequately LLMs perform in such domains. We conduct a case study on ChatGPT in nutrition counseling, a popular use-case where the model supports a user with their dietary struggles. We crowd-source real-world diet-related struggles, then work with nutrition experts to generate supportive text using ChatGPT. Finally, experts evaluate the safety and text quality of ChatGPT’s output. The result is the HAI-coaching dataset, containing ~2.4K crowdsourced dietary struggles and ~97K corresponding ChatGPT-generated and expert-annotated supportive texts. We analyse ChatGPT’s performance, discovering potentially harmful behaviours, especially for sensitive topics like mental health. Finally, we use HAI-coaching to test open LLMs on various downstream tasks, showing that even the latest models struggle to achieve good performance. HAI-coaching is available at https://github.com/uccollab/ hai-coaching/ |
2024.findings.xml | Findings of the Association for Computational Linguistics: EACL 2024 | 2024 | https://aclanthology.org/2024.findings-emnlp.713 | x | 0 | toxicity, bias | null | CantTalkAboutThis: Aligning Language Models to Stay on Topic in Dialogues | Recent advancements in instruction-tuning datasets have predominantly focused on specific tasks like mathematical or logical reasoning. There has been a notable gap in data designed for aligning language models to maintain topic relevance in conversations - a critical aspect for deploying chatbots to production. We introduce the CantTalkAboutThis dataset to help language models remain focused on the subject at hand during task-oriented interactions. It consists of synthetic dialogues on a wide range of conversation topics from different domains. These dialogues are interspersed with distractor turns that intentionally divert the chatbot from the predefined topic. Fine-tuning language models on this dataset helps make them resilient to deviating from the assigned role and improves their ability to maintain topical coherence compared to general-purpose instruction-tuned LLMs like gpt-4-turbo and Mixtral-Instruct. Additionally, preliminary observations suggest that training models on this dataset also enhance their performance on fine-grained instruction following tasks, including safety alignment. |
2024.findings.xml | Findings of the Association for Computational Linguistics: EACL 2024 | 2024 | https://aclanthology.org/2024.findings-emnlp.762 | null | 0 | general safety, LLM alignment | null | Model Merging and Safety Alignment: One Bad Model Spoils the Bunch | Merging Large Language Models (LLMs) is a cost-effective technique for combining multiple expert LLMs into a single versatile model, retaining the expertise of the original ones. However, current approaches often overlook the importance of safety alignment during merging, leading to highly misaligned models. This work investigates the effects of model merging on alignment. We evaluate several popular model merging techniques, demonstrating that existing methods do not only transfer domain expertise but also propagate misalignment. We propose a simple two-step approach to address this problem: (i) generating synthetic safety and domain-specific data, and (ii) incorporating these generated data into the optimization process of existing data-aware model merging techniques. This allows us to treat alignment as a skill that can be maximized in the resulting merged LLM. Our experiments illustrate the effectiveness of integrating alignment-related data during merging, resulting in models that excel in both domain expertise and alignment. |
2024.findings.xml | Findings of the Association for Computational Linguistics: EACL 2024 | 2024 | https://aclanthology.org/2024.findings-emnlp.784 | null | 1 | toxicity, bias | Arabic, Czech, German, English, Spanish, French, Hindi, Indonesian, Italian, Japanese, Korean, Dutch, Polish, Portuguese, Russian, Swedish, Chinese | Preference Tuning For Toxicity Mitigation Generalizes Across Languages | Detoxifying multilingual Large Language Models (LLMs) has become crucial due to their increasing global use. In this work, we explore zero-shot cross-lingual generalization of preference tuning in detoxifying LLMs. Unlike previous studies that show limited cross-lingual generalization for other safety tasks, we demonstrate that Direct Preference Optimization (DPO) training with only English data can significantly reduce toxicity in multilingual open-ended generations. For example, the probability of mGPT-1.3B generating toxic continuations drops from 46.8% to 3.9% across 17 different languages after training. Our results also extend to other multilingual LLMs, such as BLOOM, Llama3, and Aya-23. Using mechanistic interpretability tools like causal intervention and activation analysis, we identified the dual multilinguality property of MLP layers in LLMs, which explains the cross-lingual generalization of DPO. Finally, we show that bilingual sentence retrieval can predict the cross-lingual transferability of DPO preference tuning. |
2024.findings.xml | Findings of the Association for Computational Linguistics: EACL 2024 | 2024 | https://aclanthology.org/2024.findings-emnlp.808 | x | 0 | general safety, LLM alignment | null | Aligners: Decoupling LLMs and Alignment | Large Language Models (LLMs) need to be aligned with human expectations to ensure their safety and utility in most applications. Alignment is challenging, costly, and needs to be repeated for every LLM and alignment criterion. We propose to decouple LLMs and alignment by training *aligner* models that can be used to align any LLM for a given criteria on an as-needed basis, thus also reducing the potential negative impacts of alignment on performance. Our recipe for training the aligner models solely relies on synthetic data generated with a (prompted) LLM and can be easily adjusted for a variety of alignment criteria. We use the same synthetic data to train *inspectors*, binary miss-alignment classification models to guide a *squad* of multiple aligners. Our empirical results demonstrate consistent improvements when applying aligner squad to various LLMs, including chat-aligned models, across several instruction-following and red-teaming datasets. |
2024.findings.xml | Findings of the Association for Computational Linguistics: EACL 2024 | 2024 | https://aclanthology.org/2024.findings-emnlp.813 | null | 0 | jailbreaking attacks | null | DrAttack: Prompt Decomposition and Reconstruction Makes Powerful LLMs Jailbreakers | Safety-aligned Large Language Models (LLMs) are still vulnerable to some manual and automated jailbreak attacks, which adversarially trigger LLMs to output harmful content. However, existing jailbreaking methods usually view a harmful prompt as a whole but they are not effective at reducing LLMs’ attention on combinations of words with malice, which well-aligned LLMs can easily reject. This paper discovers that decomposing a malicious prompt into separated sub-prompts can effectively reduce LLMs’ attention on harmful words by presenting them to LLMs in a fragmented form, thereby addressing these limitations and improving attack effectiveness. We introduce an automatic prompt Decomposition and Reconstruction framework for jailbreaking Attack (DrAttack). DrAttack consists of three key components: (a) ‘Decomposition’ of the original prompt into sub-prompts, (b) ‘Reconstruction’ of these sub-prompts implicitly by In-Context Learning with semantically similar but benign reassembling example, and (c) ‘Synonym Search’ of sub-prompts, aiming to find sub-prompts’ synonyms that maintain the original intent while jailbreaking LLMs. An extensive empirical study across multiple open-source and closed-source LLMs demonstrates that, with fewer queries, DrAttack obtains a substantial gain of success rate on powerful LLMs over prior SOTA attackers. Notably, the success rate of 80% on GPT-4 surpassed previous art by 65%. Code and data are made publicly available at https://turningpoint-ai.github.io/DrAttack/. |
2024.findings.xml | Findings of the Association for Computational Linguistics: EACL 2024 | 2024 | https://aclanthology.org/2024.findings-emnlp.870 | x | 0 | hallucination, factuality | null | LLMs for Generating and Evaluating Counterfactuals: A Comprehensive Study | As NLP models become more complex, understanding their decisions becomes more crucial. Counterfactuals (CFs), where minimal changes to inputs flip a model’s prediction, offer a way to explain these models. While Large Language Models (LLMs) have shown remarkable performance in NLP tasks, their efficacy in generating high-quality CFs remains uncertain. This work fills this gap by investigating how well LLMs generate CFs for three tasks. We conduct a comprehensive comparison of several common LLMs, and evaluate their CFs, assessing both intrinsic metrics, and the impact of these CFs on data augmentation. Moreover, we analyze differences between human and LLM-generated CFs, providing insights for future research directions. Our results show that LLMs generate fluent CFs, but struggle to keep the induced changes minimal. Generating CFs for Sentiment Analysis (SA) is less challenging than NLI and Hate Speech (HS) where LLMs show weaknesses in generating CFs that flip the original label. This also reflects on the data augmentation performance, where we observe a large gap between augmenting with human and LLM CFs. Furthermore, we evaluate LLMs’ ability to assess CFs in a mislabelled data setting, and show that they have a strong bias towards agreeing with the provided labels. GPT4 is more robust against this bias, but it shows strong preference to its own generations. Our analysis suggests that safety training is causing GPT4 to prefer its generations, since these generations do not contain harmful content. Our findings reveal several limitations and point to potential future work directions. |
2024.findings.xml | Findings of the Association for Computational Linguistics: EACL 2024 | 2024 | https://aclanthology.org/2024.findings-emnlp.874 | null | 0 | hallucination, factuality | null | CED: Comparing Embedding Differences for Detecting Out-of-Distribution and Hallucinated Text | Detecting out-of-distribution (OOD) samples is crucial for ensuring the safety and robustness of models deployed in real-world scenarios. While most studies on OOD detection focus on fine-tuned models trained on in-distribution (ID) data, detecting OOD in pre-trained models is also important due to computational limitations and the widespread use of open-source pre-trained models. However, in the same domain shift setting, the OOD detection performance of pre-trained models is insufficient because both ID and OOD samples originate from the same domain, leading to a high overlap in their embeddings. To address this issue, we introduce a new method called CED, a training-free OOD detection technique designed to enhance the distinction between ID and OOD datasets. We theoretically validate that specific auxiliary and oracle samples that satisfy certain conditions improve this distinction. Motivated by our theoretical analysis, CED enhances the differentiation by utilizing these specially designed auxiliary and oracle samples. As a result, CED significantly improves the ability of pre-trained models to distinguish between ID and OOD samples in text classification and hallucination detection tasks. Furthermore, we verify that CED is a plug-and-play method compatible with various backbone networks, such as RoBERTa, Llama, and OpenAI Embedding. |
2024.findings.xml | Findings of the Association for Computational Linguistics: EACL 2024 | 2024 | https://aclanthology.org/2024.findings-emnlp.901 | null | 0 | general safety, LLM alignment | null | Beyond Perplexity: Multi-dimensional Safety Evaluation of LLM Compression | Increasingly, model compression techniques enable large language models (LLMs) to be deployed in real-world applications. As a result of this momentum towards local deployment, compressed LLMs will interact with a large population. Prior work on compression typically prioritize preserving perplexity, which is directly analogous to training loss. The impact of compression method on other critical aspects of model behavior—particularly safety—requires systematic assessment. To this end, we investigate the impact of model compression along four dimensions: (1) degeneration harm, i.e., bias and toxicity in generation; (2) representational harm, i.e., biases in discriminative tasks; (3) dialect bias; and (4) language modeling and downstream task performance. We examine a wide spectrum of LLM compression techniques, including unstructured pruning, semi-structured pruning, and quantization. Our analysis reveals that compression can lead to unexpected consequences. Although compression may unintentionally alleviate LLMs’ degeneration harm, it can still exacerbate representational harm. Furthermore, increasing compression produces a divergent impact on different protected groups. Finally, different compression methods have drastically different safety impacts: for example, quantization mostly preserves bias while pruning degrades quickly. Our findings underscore the importance of integrating safety assessments into the development of compressed LLMs to ensure their reliability across real-world applications. |
2024.findings.xml | Findings of the Association for Computational Linguistics: EACL 2024 | 2024 | https://aclanthology.org/2024.findings-emnlp.915 | x | 0 | general safety, LLM alignment | null | Merge to Learn: Efficiently Adding Skills to Language Models with Model Merging | Adapting general-purpose language models to new skills is currently an expensive process that must be repeated as new instruction datasets targeting new skills are created, or can cause the models to forget older skills. In this work, we investigate the effectiveness of adding new skills to preexisting models by training on the new skills in isolation and later merging with the general model (e.g. using task vectors). In experiments focusing on scientific literature understanding, safety, and coding, we find that the parallel-train-then-merge procedure, which is significantly cheaper than retraining the models on updated data mixtures, is often comparably effective. Our experiments also show that parallel training is especially well-suited for enabling safety features in LMs relative to continued finetuning and retraining, as it dramatically improves model compliance with safe prompts while preserving its ability to refuse dangerous or harmful prompts. |
2024.findings.xml | Findings of the Association for Computational Linguistics: EACL 2024 | 2024 | https://aclanthology.org/2024.findings-emnlp.928 | null | 0 | toxicity, bias | null | Evaluating Gender Bias of LLMs in Making Morality Judgements | Large Language Models (LLMs) have shown remarkable capabilities in a multitude of Natural Language Processing (NLP) tasks. However, these models are still not immune to limitations such as social biases, especially gender bias. This work investigates whether current closed and open-source LLMs possess gender bias, especially when asked to give moral opinions. To evaluate these models, we curate and introduce a new dataset GenMO (Gender-bias in Morality Opinions) comprising parallel short stories featuring male and female characters respectively. Specifically, we test models from the GPT family (GPT-3.5-turbo, GPT-3.5-turbo-instruct, GPT-4-turbo), Llama 3 and 3.1 families (8B/70B), Mistral-7B and Claude 3 families (Sonnet and Opus). Surprisingly, despite employing safety checks, all production-standard models we tested display significant gender bias with GPT-3.5-turbo giving biased opinions in 24% of the samples. Additionally, all models consistently favour female characters, with GPT showing bias in 68-85% of cases and Llama 3 in around 81-85% instances. Additionally, our study investigates the impact of model parameters on gender bias and explores real-world situations where LLMs reveal biases in moral decision-making. |
2024.findings.xml | Findings of the Association for Computational Linguistics: EACL 2024 | 2024 | https://aclanthology.org/2024.findings-emnlp.990 | x | 0 | jailbreaking attacks | null | Improving Adversarial Robustness in Vision-Language Models with Architecture and Prompt Design | Vision-Language Models (VLMs) have seen a significant increase in both research interest and real-world applications across various domains, including healthcare, autonomous systems, and security. However, their growing prevalence demands higher reliability and safety including robustness to adversarial attacks. We systematically examine the possibility of incorporating adversarial robustness through various model design choices. We explore the effects of different vision encoders, the resolutions of vision encoders, and the size and type of language models. Additionally, we introduce novel, cost-effective approaches to enhance robustness through prompt engineering. By simply suggesting the possibility of adversarial perturbations or rephrasing questions, we demonstrate substantial improvements in model robustness against strong image-based attacks such as Auto-PGD. Our findings provide important guidelines for developing more robust VLMs, particularly for deployment in safety-critical environments where reliability and security are paramount. These insights are crucial for advancing the field of VLMs, ensuring they can be safely and effectively utilized in a wide range of applications. |
2024.customnlp4u.xml | Proceedings of the 1st Workshop on Customizable NLP: Progress and Challenges in Customizing NLP for a Domain, Application, Group, or Individual (CustomNLP4U) | 2024 | https://aclanthology.org/2024.customnlp4u-1.10 | x | 0 | general safety, LLM alignment | null | Active Learning for Robust and Representative LLM Generation in Safety-Critical Scenarios | Ensuring robust safety measures across a wide range of scenarios is crucial for user-facing systems. While Large Language Models (LLMs) can generate valuable data for safety measures, they often exhibit distributional biases, focusing on common scenarios and neglecting rare but critical cases. This can undermine the effectiveness of safety protocols developed using such data. To address this, we propose a novel framework that integrates active learning with clustering to guide LLM generation, enhancing their representativeness and robustness in safety scenarios. We demonstrate the effectiveness of our approach by constructing a dataset of 5.4K potential safety violations through an iterative process involving LLM generation and an active learner model’s feedback. Our results show that the proposed framework produces a more representative set of safety scenarios without requiring prior knowledge of the underlying data distribution. Additionally, data acquired through our method improves the accuracy and F1 score of both the active learner model as well models outside the scope of active learning process, highlighting its broad applicability. |
2024.wnu.xml | Proceedings of the The 6th Workshop on Narrative Understanding | 2024 | https://aclanthology.org/2024.wnu-1.7 | x | 0 | privacy | null | Is It Safe to Tell Your Story? Towards Achieving Privacy for Sensitive Narratives | Evolving tools for narrative analysis present an opportunity to identify common structure in stories that are socially important to tell, such as stories of survival from domestic abuse. A greater structural understanding of such stories could lead to stronger protections against de-anonymization, as well as future tools to help survivors navigate the complex trade-offs inherent in trying to tell their stories safely. In this work we explore narrative patterns within a small set of domestic violence stories, identifying many similarities. We then propose a method to assess the safety of sharing a story based on a distance feature vector. |
2024.emnlp.xml | Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing | 2024 | https://aclanthology.org/2024.emnlp-main.41 | x | 0 | general safety, LLM alignment | null | FLIRT: Feedback Loop In-context Red Teaming | Warning: this paper contains content that may be inappropriate or offensive.As generative models become available for public use in various applications, testing and analyzing vulnerabilities of these models has become a priority. In this work, we propose an automatic red teaming framework that evaluates a given black-box model and exposes its vulnerabilities against unsafe and inappropriate content generation. Our framework uses in-context learning in a feedback loop to red team models and trigger them into unsafe content generation. In particular, taking text-to-image models as target models, we explore different feedback mechanisms to automatically learn effective and diverse adversarial prompts. Our experiments demonstrate that even with enhanced safety features, Stable Diffusion (SD) models are vulnerable to our adversarial prompts, raising concerns on their robustness in practical uses. Furthermore, we demonstrate that the proposed framework is effective for red teaming text-to-text models. |
2024.emnlp.xml | Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing | 2024 | https://aclanthology.org/2024.emnlp-main.108 | x | 0 | others | null | Evaluating Psychological Safety of Large Language Models | In this work, we designed unbiased prompts to systematically evaluate the psychological safety of large language models (LLMs). First, we tested five different LLMs by using two personality tests: Short Dark Triad (SD-3) and Big Five Inventory (BFI). All models scored higher than the human average on SD-3, suggesting a relatively darker personality pattern. Despite being instruction fine-tuned with safety metrics to reduce toxicity, InstructGPT, GPT-3.5, and GPT-4 still showed dark personality patterns; these models scored higher than self-supervised GPT-3 on the Machiavellianism and narcissism traits on SD-3. Then, we evaluated the LLMs in the GPT series by using well-being tests to study the impact of fine-tuning with more training data. We observed a continuous increase in the well-being scores of GPT models. Following these observations, we showed that fine-tuning Llama-2-chat-7B with responses from BFI using direct preference optimization could effectively reduce the psychological toxicity of the model. Based on the findings, we recommended the application of systematic and comprehensive psychological metrics to further evaluate and improve the safety of LLMs. |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.