{ "paper_id": "U19-1007", "header": { "generated_with": "S2ORC 1.0.0", "date_generated": "2023-01-19T03:07:43.374777Z" }, "title": "Feature-guided Neural Model Training for Supervised Document Representation Learning", "authors": [ { "first": "Aili", "middle": [], "last": "Shen", "suffix": "", "affiliation": { "laboratory": "", "institution": "The University of Melbourne Victoria", "location": { "country": "Australia" } }, "email": "ailis@student.unimelb.edu.au" }, { "first": "Bahar", "middle": [], "last": "Salehi", "suffix": "", "affiliation": { "laboratory": "", "institution": "The University of Melbourne Victoria", "location": { "country": "Australia" } }, "email": "baharsalehi@gmail.com" }, { "first": "Jianzhong", "middle": [], "last": "Qi", "suffix": "", "affiliation": { "laboratory": "", "institution": "The University of Melbourne Victoria", "location": { "country": "Australia" } }, "email": "jianzhong.qi@unimelb.edu.au" }, { "first": "Timothy", "middle": [], "last": "Baldwin", "suffix": "", "affiliation": { "laboratory": "", "institution": "The University of Melbourne Victoria", "location": { "country": "Australia" } }, "email": "" } ], "year": "", "venue": null, "identifiers": {}, "abstract": "With the advent of neural models, there has been a rapid move away from feature engineering, or at best, simplistically combining handcrafted features with learned representations as side information. We propose a method that uses hand-crafted features to guide learning by explicitly attending to feature indicators when learning the relationship between the input and target variables. In experiments over two different tasks-quality assessment of Wikipedia articles and popularity prediction of online petitions-we demonstrate that the proposed method yields neural models that consistently outperform those that simply use hand-crafted features as side information.", "pdf_parse": { "paper_id": "U19-1007", "_pdf_hash": "", "abstract": [ { "text": "With the advent of neural models, there has been a rapid move away from feature engineering, or at best, simplistically combining handcrafted features with learned representations as side information. We propose a method that uses hand-crafted features to guide learning by explicitly attending to feature indicators when learning the relationship between the input and target variables. In experiments over two different tasks-quality assessment of Wikipedia articles and popularity prediction of online petitions-we demonstrate that the proposed method yields neural models that consistently outperform those that simply use hand-crafted features as side information.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Abstract", "sec_num": null } ], "body_text": [ { "text": "Text classification/regression is a fundamental problem in natural language processing. Traditional methods make use of hand-crafted features, such as the length of a document, to represent a document. A classifier/regressor is built on top of such features to learn a model (Wang and Manning, 2012; Warncke-Wang et al., 2013 Dang and Ignat, 2016) . Recently, neural models such as LSTMs (Hochreiter and Schmidhuber, 1997) and convolutional neural networks (CNNs: Kim (2014) ; Kalchbrenner et al. (2014) ) have become the de facto for text classification/regression tasks, with one oft-cited advantage being that they are able to learn implicit features as part of the representation learning.", "cite_spans": [ { "start": 275, "end": 299, "text": "(Wang and Manning, 2012;", "ref_id": "BIBREF12" }, { "start": 300, "end": 325, "text": "Warncke-Wang et al., 2013", "ref_id": "BIBREF14" }, { "start": 326, "end": 347, "text": "Dang and Ignat, 2016)", "ref_id": "BIBREF0" }, { "start": 388, "end": 422, "text": "(Hochreiter and Schmidhuber, 1997)", "ref_id": "BIBREF3" }, { "start": 464, "end": 474, "text": "Kim (2014)", "ref_id": "BIBREF5" }, { "start": 477, "end": 503, "text": "Kalchbrenner et al. (2014)", "ref_id": "BIBREF4" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction and Background", "sec_num": "1" }, { "text": "Studies employing neural models either eschew hand-crafted features or simplistically use handcrafted features as side information. For example, Dang and Ignat (2017) propose to use a bidirectional LSTM (\"bi-LSTM\") to classify Wikipedia articles by their quality classes, and Shen et al. (2017) concatenate structural features (e.g., article length) and readability scores with bi-LSTM-learned document representations for the same task. Subramanian et al. (2018) hand-engineer a set of features (e.g., the ratio of indefinite and definite articles), and concatenate them with CNNlearned document representations to predict the popularity of online petitions. Wu et al. (2018) explore the utility of hand-crafted features in NER by concatenating these features with character representations learned via an CNN and word embeddings. These representations are then fed into a bi-LSTM to identify named entities (with the help of a CRF) and re-construct the hand-crafted features in the output simultaneously, which is achieved by combining an auto-encoder loss with the NER loss.", "cite_spans": [ { "start": 145, "end": 166, "text": "Dang and Ignat (2017)", "ref_id": "BIBREF1" }, { "start": 276, "end": 294, "text": "Shen et al. (2017)", "ref_id": "BIBREF8" }, { "start": 438, "end": 463, "text": "Subramanian et al. (2018)", "ref_id": "BIBREF10" }, { "start": 660, "end": 676, "text": "Wu et al. (2018)", "ref_id": "BIBREF15" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction and Background", "sec_num": "1" }, { "text": "The motivation underlying this work is that when hand-crafted features are represented by numerical vectors and concatenated with neural network representations, there is no information on what kind of feature each value represents. To make better use of hand-crafted features, we propose a feature-guided neural training method that guides the network to map feature indicators onto (explicit or implicit) features in the document. We evaluate the effectiveness of the proposed method over two datasets for two different tasks: (1) quality assessment of Wikipedia articles, and (2) popularity prediction of online petitions. Taking stateof-the-art approaches for the respective tasks, we achieve consistent improvements when using our model.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction and Background", "sec_num": "1" }, { "text": "The closest work to our approach is the labelguided model training of Wang et al. (2018) . They embed words and labels in the same embedding space, and compute a label-based attention score between a word and all possible labels, which is used to weight word embeddings in obtaining document representations. Our work differs in two aspects: (1) Wang et al. (2018) capture direct associations between labels and words, while we use the proxy of (potentially much higher-level) handcrafted features to guide network learning; and (2) our method does not rely on the target variable being closely related to the semantics of a document, leading to better generalisability. Figure 1 is an illustration of our proposed approach, in the context of a stacked bi-LSTM, where two bi-LSTMs are applied to obtain the sentence and document level representations, respectively. Note that our method is not limited to LSTMs, as we show in Section 3.", "cite_spans": [ { "start": 70, "end": 88, "text": "Wang et al. (2018)", "ref_id": "BIBREF11" } ], "ref_spans": [ { "start": 671, "end": 679, "text": "Figure 1", "ref_id": null } ], "eq_spans": [], "section": "Introduction and Background", "sec_num": "1" }, { "text": "A hand-crafted feature can consist of multiple feature indicators. For example, having level 3+ headings is one of the features in quality assessment of Wikipedia articles. This hand-crafted feature consists of two feature indicators (tokens) {\"===\", \"====\"}. We embed the document and feature indicators into a shared space. Then, as indicated in Figure 1 , for each feature indicator, we compute cosine similarity between the feature indicator and word embeddings, followed by average-pooling to obtain a sentence score:", "cite_spans": [], "ref_spans": [ { "start": 348, "end": 356, "text": "Figure 1", "ref_id": null } ], "eq_spans": [], "section": "Methodology", "sec_num": "2" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "score = 1 N N j=1 FV j F 2 V j 2 .", "eq_num": "(1)" } ], "section": "Methodology", "sec_num": "2" }, { "text": "Here, F and V j are embeddings of the feature indicator and the jth word in a sentence, respectively; F 2 and V j 2 are the 2 norms of F and V j , respectively; and N is the number of words in a sentence. All scores based on the feature indicator are concatenated with sentence representations Z 1 , which are learned through a bi-LSTM layer (f 1 ). Then, another bi-LSTM layer (f 2 ) is applied to the concatenated sentence representations to obtain document representation Z 2 , which is followed by a dense layer (f 3 ) to compute y. The score computed in Equation 1 is for a single-token feature indicator. If a hand-crafted feature consists of multiple feature indicators (tokens, e.g., {\"===\", \"====\"}), the score becomes:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Methodology", "sec_num": "2" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "score = 1 M M i=1 1 N N j=1 F i V j F i 2 V j 2 .", "eq_num": "(2)" } ], "section": "Methodology", "sec_num": "2" }, { "text": "Here, M is the number of feature indicators in a hand-crafted feature, and F i is the word embedding of the ith feature indicator in the feature. For example, the feature level 3+ headings is one of structural features described in Shen et al.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Methodology", "sec_num": "2" }, { "text": "word embedding V feature embedding F score Z 1 f 1 concat Z 2 f 2 y f 3", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Methodology", "sec_num": "2" }, { "text": "Figure 1: Illustration of our proposed method. Here, \u2297 denotes cosine similarity between the feature indicator and word embeddings; f 1 and f 2 denote bi-LSTM layers; f 3 denotes a dense layer; Z 1 and Z 2 denote sentence and document representations, respectively; V and F are the document input and feature indicators, respectively; and y is the target output.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Methodology", "sec_num": "2" }, { "text": "(2017), which consists of two feature indicators {\"===\", \"====\"}. To obtain the similarity score for the feature level 3+ headings, we first compute the similarity score between each feature indicator \"===\"/\"====\" and word embeddings in the sentence, then apply average-pooling to obtain the similarity score for each feature indicator. We obtain the similarity score of level 3+ headings by averaging similarity scores among the feature indicators at the sentence level. Finally, the feature score is concatenated with the sentence representation, which is fed into a latter layer.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Methodology", "sec_num": "2" }, { "text": "While we don't experiment with this in this paper, it is also possible to first average feature indicator embeddings and then compute the sentence score by Equation 1. This way, we can efficiently reduce the computation of similarity scores for hand-crafted features with a large number of feature indicators. In this paper, we use Equation 2, as the maximum number of feature indicators in a given hand-crated feature is less than 1, 000, and less than 10 in most cases.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Methodology", "sec_num": "2" }, { "text": "To test the effectiveness of our proposed method, we experiment with a Wikipedia document quality assessment task (Shen et al., 2019) , and online petition signature prediction task (Subramanian et al., 2018) , as detailed below. The reasons we chose these particular tasks are as follows. First, extensive domain-specific feature engineering had taken place in each case, that we could use as the basis of our feature indicators. Second, strong neural benchmarks have been established, based on extensive experimentation with both neural and non-neural models. Our experiments in this paper are based on the state-of-the-art.", "cite_spans": [ { "start": 114, "end": 133, "text": "(Shen et al., 2019)", "ref_id": "BIBREF9" }, { "start": 182, "end": 208, "text": "(Subramanian et al., 2018)", "ref_id": "BIBREF10" } ], "ref_spans": [], "eq_spans": [], "section": "Experiments", "sec_num": "3" }, { "text": "We aim to explore the relative gains of our proposed method relative to the current state-of-theart for the task, which in both cases is not based on contextualised embeddings. For BERT (Devlin et al., 2019) or other contextualised encoders, the same word in different contexts will end with different embeddings, leading to localized representations of feature indicators. As such, the proposed method is not directly applicable to models such as BERT, and novel research would be required to adapt the method to such models.", "cite_spans": [ { "start": 186, "end": 207, "text": "(Devlin et al., 2019)", "ref_id": "BIBREF2" } ], "ref_spans": [], "eq_spans": [], "section": "Experiments", "sec_num": "3" }, { "text": "Dataset The Wikipedia dataset (Shen et al., 2019) consists of 29,794 English Wikipedia articles and their corresponding quality labels: Featured Article, Good Article, B-class Article, Cclass Article, Start Article, and Stub Article, in descending order of document quality. The dataset is class-balanced and partitioned into training, development, and test splits (8:1:1). Documents are relatively long, and processed in a hierarchical manner, by constructing sentence representations, and composing these into a document representation.", "cite_spans": [ { "start": 30, "end": 49, "text": "(Shen et al., 2019)", "ref_id": "BIBREF9" } ], "ref_spans": [], "eq_spans": [], "section": "Wikipedia Document Quality Assessment", "sec_num": "3.1" }, { "text": "Following Dang and Ignat (2017) and Shen et al. (2019), we formulate the quality assessment of Wikipedia articles as a multi-class classification problem, and all models are trained to minimise cross-entropy loss. We report average accuracy and standard deviation over 10 runs.", "cite_spans": [ { "start": 10, "end": 31, "text": "Dang and Ignat (2017)", "ref_id": "BIBREF1" } ], "ref_spans": [], "eq_spans": [], "section": "Wikipedia Document Quality Assessment", "sec_num": "3.1" }, { "text": "Hand-crafted features used to guide network learning here include: (1) references indicators;", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Wikipedia Document Quality Assessment", "sec_num": "3.1" }, { "text": "(2) links to other Wikipedia pages indicators;", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Wikipedia Document Quality Assessment", "sec_num": "3.1" }, { "text": "(3) citation templates indicators; (4) non-citation templates indicators; (5) categories linked in the article indicators; (6) image indicators; (7) infobox indicators; (8) level 2 headings indicators; and (9) level 3+ heading indicators. These features are from Dang and Ignat (2016) and Shen et al. (2017) . Hand-crafted features in Side-information are based on counting the number of appearances of such feature indicators.", "cite_spans": [ { "start": 263, "end": 284, "text": "Dang and Ignat (2016)", "ref_id": "BIBREF0" }, { "start": 289, "end": 307, "text": "Shen et al. (2017)", "ref_id": "BIBREF8" } ], "ref_spans": [], "eq_spans": [], "section": "Wikipedia Document Quality Assessment", "sec_num": "3.1" }, { "text": "We apply our proposed approach (\"Feature-guided\") over the four models detailed below.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Model configuration", "sec_num": null }, { "text": "In each case, we contrast with two baselines: (1) Vanilla, makes no use of hand-crafted features; and 2Side-information which uses the handcrafted features as side information, by concatenating them with learned representations in the penultimate layer.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Model configuration", "sec_num": null }, { "text": "with width 2, 3, and 4 (32 for each width size) to word embeddings within a sentence, and a tanh activation function to each; pass the output of the filters through a bi-LSTM.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "CNN BILSTM: apply convolution kernels", "sec_num": "1." }, { "text": "2. AVERAGE BILSTM (Shen et al., 2017) : average word embeddings to get the sentence representation, and run a bi-LSTM over the sequence of sentence representations.", "cite_spans": [ { "start": 18, "end": 37, "text": "(Shen et al., 2017)", "ref_id": "BIBREF8" } ], "ref_spans": [], "eq_spans": [], "section": "CNN BILSTM: apply convolution kernels", "sec_num": "1." }, { "text": "3. STACKED BILSTM: feed the word embeddings in a sentence through a bi-LSTM, and the output through a max-pooling layer; finally, apply another bi-LSTM over the sentence representations.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "CNN BILSTM: apply convolution kernels", "sec_num": "1." }, { "text": "4. STACKED BILSTM ATT (Yang et al., 2016) : use a hierarchical STACKED BILSTM, except that an attention mechanism with a context size of 100 is applied to the output of each bi-LSTM to weight words/sentences based on their importance in the sentence/document.", "cite_spans": [ { "start": 22, "end": 41, "text": "(Yang et al., 2016)", "ref_id": "BIBREF16" } ], "ref_spans": [], "eq_spans": [], "section": "CNN BILSTM: apply convolution kernels", "sec_num": "1." }, { "text": "A max-pooling layer is applied to the output of the bi-LSTM at the sentence level for all models except STACKED BILSTM ATT to get the document level representation, which is followed by two dense layers, one with a ReLU activation and one without any activation function. For all models, dropout layers are applied at both the sentence and document levels with a rate of 0.5 during training. For both CNN BILSTM and AVERAGE BILSTM, the bi-LSTM cell size is set to 256. For STACKED BILSTM and STACKED BILSTM ATT, the cell size is set to 32 and 256 for the sentence and document level bi-LSTM, respectively.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "CNN BILSTM: apply convolution kernels", "sec_num": "1." }, { "text": "We use 50-dimensional pre-trained word embeddings from GloVe (Pennington et al., 2014) . 1 For OOV words, the word embeddings are randomly initialised based on sampling from a uniform distribution U (\u22121, 1). All word embeddings 1 We fine-tuned hyper-parameters over the development set for quality predictions of Wikipedia articles. 50dimensional embeddings were chosen because Vanilla performs the best under this setting (meaning the baseline without features is as strong as possible). Table 1 : accuracy over Wikipedia dataset. The best result is in bold, and marked with \" \u2020\" if the improvement is statistically significant (based on a one-tailed Wilcoxon signed-rank test; p < 0.05).", "cite_spans": [ { "start": 61, "end": 86, "text": "(Pennington et al., 2014)", "ref_id": "BIBREF7" }, { "start": 228, "end": 229, "text": "1", "ref_id": null } ], "ref_spans": [ { "start": 489, "end": 496, "text": "Table 1", "ref_id": null } ], "eq_spans": [], "section": "CNN BILSTM: apply convolution kernels", "sec_num": "1." }, { "text": "are updated in the training process. We use a mini-batch size of 128 and a learning rate of 0.01. We train each model for 50 epochs. To prevent over-fitting, early stopping is adopted. All hyperparameters are set empirically over the development data, and the models are optimised using Adam (Kingma and Ba, 2015) .", "cite_spans": [ { "start": 292, "end": 313, "text": "(Kingma and Ba, 2015)", "ref_id": "BIBREF6" } ], "ref_spans": [], "eq_spans": [], "section": "CNN BILSTM: apply convolution kernels", "sec_num": "1." }, { "text": "Results The experimental results are presented in Table 1 .", "cite_spans": [], "ref_spans": [ { "start": 50, "end": 57, "text": "Table 1", "ref_id": null } ], "eq_spans": [], "section": "CNN BILSTM: apply convolution kernels", "sec_num": "1." }, { "text": "We can see that our method outperforms both Vanilla and Side-information across all four network architectures, at a level of statistical significance for 3 out of the 4 models. It is worth noting that the performance of STACKED BILSTM ATT is worse than that of STACKED BILSTM for both Vanilla and Side-information, due to attention in STACKED BILSTM ATT not being as effective as max-pooling on this task. However, the performance of STACKED BILSTM ATT is not worsened by incorporating the attention for our method, indicating that our feature-guided learning can guide the network to learn better.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "CNN BILSTM: apply convolution kernels", "sec_num": "1." }, { "text": "Dataset The online petitions dataset (Subramanian et al., 2018) consists of 10,950 UK petitions and their corresponding signature counts. Following Subramanian et al. (2018) , we chronologically split the data into training, dev, and test (8:1:1).", "cite_spans": [ { "start": 37, "end": 63, "text": "(Subramanian et al., 2018)", "ref_id": "BIBREF10" }, { "start": 148, "end": 173, "text": "Subramanian et al. (2018)", "ref_id": "BIBREF10" } ], "ref_spans": [], "eq_spans": [], "section": "Online Petition Signature Prediction", "sec_num": "3.2" }, { "text": "We formulate signature count prediction as a regression problem, and all models are trained to minimise mean squared error. For evaluation, average mean absolute error (\"MAE\") and mean absolute percentage error (\"MAPE\") over 10 runs are reported. Here, MAPE is calculated as 100 n n 1", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Online Petition Signature Prediction", "sec_num": "3.2" }, { "text": "|\u0177 i \u2212 y i | y i", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Online Petition Signature Prediction", "sec_num": "3.2" }, { "text": ", where\u0177 i and y i are the predicted and ground-truth signature counts.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Online Petition Signature Prediction", "sec_num": "3.2" }, { "text": "Hand-crafted features used to guide network learning here include: (1) indefinite vs. definite articles;", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Online Petition Signature Prediction", "sec_num": "3.2" }, { "text": "(2) P1 singular and plural, P2, and P3 singular and plural pronouns; (3) subjective, positive, and negative words; and (4) biased words. These features are from Subramanian et al. (2018 Hand-crafted features in Side-information are based on counting the number of appearances of such feature words.", "cite_spans": [ { "start": 161, "end": 185, "text": "Subramanian et al. (2018", "ref_id": "BIBREF10" } ], "ref_spans": [], "eq_spans": [], "section": "Online Petition Signature Prediction", "sec_num": "3.2" }, { "text": "Model configuration We again compare our proposed Feature-guided approach with Side-information and Vanilla, in the form of a state-of-the-art CNN with convolution kernels of width 1, 2, and 3 (100 kernels for each width size) over word embeddings, with a ReLU applied to each. The outputs are passed through two dense layers, one with a tanh activation function and one with an ELU activation function, to obtain the final output. A dropout layer is applied to the output of the convolution filters at a rate of 0.5 during training. We use a mini-batch size of 32, and a learning rate of 1e \u2212 4. All other hyper-parameters are the same as in the Wikipedia setting, except that the early stopping is based on MAE.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Online Petition Signature Prediction", "sec_num": "3.2" }, { "text": "Results Table 2 summarises our results. We observe that our approach benefits Vanilla and Side-information once again, at a level of significance in terms of both MAE and MAPE. It is worth noting that Side-information performs worse than Vanilla, as it over-fits to features only present in the training data. In comparison, our method improves the model performance even in this case, as it identifies words semantically related to the feature indicators. For example, the word increase, not in the predefined list of positive words, has a high similarity score (> 0.8) with positive words help, hope, give, and allow.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Online Petition Signature Prediction", "sec_num": "3.2" }, { "text": "We proposed a method to guide network learning by attending to feature indicators associated with hand-crafted features. Experimental results over two tasks (quality assessment of Wikipedia articles, and popularity prediction of online petitions) show that our approach consistently outperforms two baselines, across a range of neural architectures. For future work, we are interested in exploring hand-crafted features from external sources, such as editor comments of a Wikipedia article.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Online Petition Signature Prediction", "sec_num": "3.2" } ], "back_matter": [], "bib_entries": { "BIBREF0": { "ref_id": "b0", "title": "Measuring quality of collaboratively edited documents: the case of Wikipedia", "authors": [ { "first": "Quang-Vinh", "middle": [], "last": "Dang", "suffix": "" }, { "first": "Claudia-Lavinia", "middle": [], "last": "Ignat", "suffix": "" } ], "year": 2016, "venue": "Proceedings of the 2nd IEEE International Conference on Collaboration and Internet Computing", "volume": "", "issue": "", "pages": "266--275", "other_ids": {}, "num": null, "urls": [], "raw_text": "Quang-Vinh Dang and Claudia-Lavinia Ignat. 2016. Measuring quality of collaboratively edited docu- ments: the case of Wikipedia. In Proceedings of the 2nd IEEE International Conference on Collabo- ration and Internet Computing, pages 266-275.", "links": null }, "BIBREF1": { "ref_id": "b1", "title": "An end-to-end learning solution for assessing the quality of Wikipedia articles", "authors": [ { "first": "Vinh", "middle": [], "last": "Quang", "suffix": "" }, { "first": "Claudia-Lavinia", "middle": [], "last": "Dang", "suffix": "" }, { "first": "", "middle": [], "last": "Ignat", "suffix": "" } ], "year": 2017, "venue": "Proceedings of the 13th International Symposium on Open Collaboration", "volume": "4", "issue": "", "pages": "1--4", "other_ids": {}, "num": null, "urls": [], "raw_text": "Quang Vinh Dang and Claudia-Lavinia Ignat. 2017. An end-to-end learning solution for assessing the quality of Wikipedia articles. In Proceedings of the 13th International Symposium on Open Collabora- tion, pages 4:1-4:10.", "links": null }, "BIBREF2": { "ref_id": "b2", "title": "BERT: Pre-training of deep bidirectional transformers for language understanding", "authors": [ { "first": "Jacob", "middle": [], "last": "Devlin", "suffix": "" }, { "first": "Ming-Wei", "middle": [], "last": "Chang", "suffix": "" }, { "first": "Kenton", "middle": [], "last": "Lee", "suffix": "" }, { "first": "Kristina", "middle": [], "last": "Toutanova", "suffix": "" } ], "year": 2019, "venue": "Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, (NAACL-HLT 2019)", "volume": "", "issue": "", "pages": "4171--4186", "other_ids": {}, "num": null, "urls": [], "raw_text": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language under- standing. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Tech- nologies, (NAACL-HLT 2019), pages 4171-4186.", "links": null }, "BIBREF3": { "ref_id": "b3", "title": "Long short-term memory", "authors": [ { "first": "Sepp", "middle": [], "last": "Hochreiter", "suffix": "" }, { "first": "J\u00fcrgen", "middle": [], "last": "Schmidhuber", "suffix": "" } ], "year": 1997, "venue": "Neural Computation", "volume": "9", "issue": "8", "pages": "1735--1780", "other_ids": {}, "num": null, "urls": [], "raw_text": "Sepp Hochreiter and J\u00fcrgen Schmidhuber. 1997. Long short-term memory. Neural Computation, 9(8):1735-1780.", "links": null }, "BIBREF4": { "ref_id": "b4", "title": "A convolutional neural network for modelling sentences", "authors": [ { "first": "Nal", "middle": [], "last": "Kalchbrenner", "suffix": "" }, { "first": "Edward", "middle": [], "last": "Grefenstette", "suffix": "" }, { "first": "Phil", "middle": [], "last": "Blunsom", "suffix": "" } ], "year": 2014, "venue": "Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (ACL 2014)", "volume": "", "issue": "", "pages": "655--665", "other_ids": {}, "num": null, "urls": [], "raw_text": "Nal Kalchbrenner, Edward Grefenstette, and Phil Blun- som. 2014. A convolutional neural network for modelling sentences. In Proceedings of the 52nd Annual Meeting of the Association for Computa- tional Linguistics (ACL 2014), pages 655-665.", "links": null }, "BIBREF5": { "ref_id": "b5", "title": "Convolutional neural networks for sentence classification", "authors": [ { "first": "Yoon", "middle": [], "last": "Kim", "suffix": "" } ], "year": 2014, "venue": "Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing", "volume": "", "issue": "", "pages": "1746--1751", "other_ids": {}, "num": null, "urls": [], "raw_text": "Yoon Kim. 2014. Convolutional neural networks for sentence classification. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP 2014), pages 1746- 1751.", "links": null }, "BIBREF6": { "ref_id": "b6", "title": "Adam: A method for stochastic optimization", "authors": [ { "first": "P", "middle": [], "last": "Diederik", "suffix": "" }, { "first": "Jimmy", "middle": [], "last": "Kingma", "suffix": "" }, { "first": "", "middle": [], "last": "Ba", "suffix": "" } ], "year": 2015, "venue": "The 3rd International Conference on Learning Representations", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Diederik P. Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. In The 3rd International Conference on Learning Representa- tions, (ICLR 2015).", "links": null }, "BIBREF7": { "ref_id": "b7", "title": "GloVe: Global vectors for word representation", "authors": [ { "first": "Jeffrey", "middle": [], "last": "Pennington", "suffix": "" }, { "first": "Richard", "middle": [], "last": "Socher", "suffix": "" }, { "first": "Christopher", "middle": [ "D" ], "last": "Manning", "suffix": "" } ], "year": 2014, "venue": "Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing", "volume": "", "issue": "", "pages": "1532--1543", "other_ids": {}, "num": null, "urls": [], "raw_text": "Jeffrey Pennington, Richard Socher, and Christo- pher D. Manning. 2014. GloVe: Global vectors for word representation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP 2018), pages 1532- 1543.", "links": null }, "BIBREF8": { "ref_id": "b8", "title": "A hybrid model for quality assessment of Wikipedia articles", "authors": [ { "first": "Aili", "middle": [], "last": "Shen", "suffix": "" }, { "first": "Jianzhong", "middle": [], "last": "Qi", "suffix": "" }, { "first": "Timothy", "middle": [], "last": "Baldwin", "suffix": "" } ], "year": 2017, "venue": "Proceedings of the Australasian Language Technology Association Workshop", "volume": "", "issue": "", "pages": "43--52", "other_ids": {}, "num": null, "urls": [], "raw_text": "Aili Shen, Jianzhong Qi, and Timothy Baldwin. 2017. A hybrid model for quality assessment of Wikipedia articles. In Proceedings of the Australasian Lan- guage Technology Association Workshop, pages 43- 52.", "links": null }, "BIBREF9": { "ref_id": "b9", "title": "A joint model for multimodal document quality assessment", "authors": [ { "first": "Aili", "middle": [], "last": "Shen", "suffix": "" }, { "first": "Bahar", "middle": [], "last": "Salehi", "suffix": "" }, { "first": "Timothy", "middle": [], "last": "Baldwin", "suffix": "" }, { "first": "Jianzhong", "middle": [], "last": "Qi", "suffix": "" } ], "year": 2019, "venue": "Proceedings of the 19th ACM/IEEE-CS on Joint Conference on Digital Libraries (JCDL 2019)", "volume": "", "issue": "", "pages": "107--110", "other_ids": {}, "num": null, "urls": [], "raw_text": "Aili Shen, Bahar Salehi, Timothy Baldwin, and Jianzhong Qi. 2019. A joint model for multimodal document quality assessment. In Proceedings of the 19th ACM/IEEE-CS on Joint Conference on Digital Libraries (JCDL 2019), pages 107-110.", "links": null }, "BIBREF10": { "ref_id": "b10", "title": "Content-based popularity prediction of online petitions using a deep regression model", "authors": [ { "first": "Shivashankar", "middle": [], "last": "Subramanian", "suffix": "" }, { "first": "Timothy", "middle": [], "last": "Baldwin", "suffix": "" }, { "first": "Trevor", "middle": [], "last": "Cohn", "suffix": "" } ], "year": 2018, "venue": "Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics", "volume": "", "issue": "", "pages": "182--188", "other_ids": {}, "num": null, "urls": [], "raw_text": "Shivashankar Subramanian, Timothy Baldwin, and Trevor Cohn. 2018. Content-based popularity pre- diction of online petitions using a deep regression model. In Proceedings of the 56th Annual Meet- ing of the Association for Computational Linguis- tics, (ACL 2018), pages 182-188.", "links": null }, "BIBREF11": { "ref_id": "b11", "title": "Joint embedding of words and labels for text classification", "authors": [ { "first": "Guoyin", "middle": [], "last": "Wang", "suffix": "" }, { "first": "Chunyuan", "middle": [], "last": "Li", "suffix": "" }, { "first": "Wenlin", "middle": [], "last": "Wang", "suffix": "" }, { "first": "Yizhe", "middle": [], "last": "Zhang", "suffix": "" }, { "first": "Dinghan", "middle": [], "last": "Shen", "suffix": "" }, { "first": "Xinyuan", "middle": [], "last": "Zhang", "suffix": "" }, { "first": "Ricardo", "middle": [], "last": "Henao", "suffix": "" }, { "first": "Lawrence", "middle": [], "last": "Carin", "suffix": "" } ], "year": 2018, "venue": "Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (ACL 2018)", "volume": "", "issue": "", "pages": "2321--2331", "other_ids": {}, "num": null, "urls": [], "raw_text": "Guoyin Wang, Chunyuan Li, Wenlin Wang, Yizhe Zhang, Dinghan Shen, Xinyuan Zhang, Ricardo Henao, and Lawrence Carin. 2018. Joint embedding of words and labels for text classification. In Pro- ceedings of the 56th Annual Meeting of the Asso- ciation for Computational Linguistics (ACL 2018), pages 2321-2331.", "links": null }, "BIBREF12": { "ref_id": "b12", "title": "Baselines and bigrams: Simple, good sentiment and topic classification", "authors": [ { "first": "I", "middle": [], "last": "Sida", "suffix": "" }, { "first": "Christopher", "middle": [ "D" ], "last": "Wang", "suffix": "" }, { "first": "", "middle": [], "last": "Manning", "suffix": "" } ], "year": 2012, "venue": "Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics (ACL 2012)", "volume": "", "issue": "", "pages": "90--94", "other_ids": {}, "num": null, "urls": [], "raw_text": "Sida I. Wang and Christopher D. Manning. 2012. Base- lines and bigrams: Simple, good sentiment and topic classification. In Proceedings of the 50th Annual Meeting of the Association for Computational Lin- guistics (ACL 2012), pages 90-94.", "links": null }, "BIBREF13": { "ref_id": "b13", "title": "The success and failure of quality improvement projects in peer production communities", "authors": [ { "first": "Morten", "middle": [], "last": "Warncke-Wang", "suffix": "" }, { "first": "R", "middle": [], "last": "Vladislav", "suffix": "" }, { "first": "Brent", "middle": [], "last": "Ayukaev", "suffix": "" }, { "first": "Loren", "middle": [], "last": "Hecht", "suffix": "" }, { "first": "", "middle": [], "last": "Terveen", "suffix": "" } ], "year": 2015, "venue": "Proceedings of the 18th ACM Conference on Computer-Supported Cooperative Work and Social Computing", "volume": "", "issue": "", "pages": "743--756", "other_ids": {}, "num": null, "urls": [], "raw_text": "Morten Warncke-Wang, Vladislav R. Ayukaev, Brent Hecht, and Loren Terveen. 2015. The success and failure of quality improvement projects in peer pro- duction communities. In Proceedings of the 18th ACM Conference on Computer-Supported Coopera- tive Work and Social Computing, pages 743-756.", "links": null }, "BIBREF14": { "ref_id": "b14", "title": "Tell me more: An actionable quality model for Wikipedia", "authors": [ { "first": "Morten", "middle": [], "last": "Warncke-Wang", "suffix": "" }, { "first": "Dan", "middle": [], "last": "Cosley", "suffix": "" }, { "first": "John", "middle": [], "last": "Riedl", "suffix": "" } ], "year": 2013, "venue": "Proceedings of the 9th International Symposium on Open Collaboration", "volume": "8", "issue": "", "pages": "1--8", "other_ids": {}, "num": null, "urls": [], "raw_text": "Morten Warncke-Wang, Dan Cosley, and John Riedl. 2013. Tell me more: An actionable quality model for Wikipedia. In Proceedings of the 9th Inter- national Symposium on Open Collaboration, pages 8:1-8:10.", "links": null }, "BIBREF15": { "ref_id": "b15", "title": "Evaluating the utility of hand-crafted features in sequence labelling", "authors": [ { "first": "Minghao", "middle": [], "last": "Wu", "suffix": "" }, { "first": "Fei", "middle": [], "last": "Liu", "suffix": "" }, { "first": "Trevor", "middle": [], "last": "Cohn", "suffix": "" } ], "year": 2018, "venue": "Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Minghao Wu, Fei Liu, and Trevor Cohn. 2018. Evalu- ating the utility of hand-crafted features in sequence labelling. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP 2018), page 28502856.", "links": null }, "BIBREF16": { "ref_id": "b16", "title": "Hierarchical attention networks for document classification", "authors": [ { "first": "Zichao", "middle": [], "last": "Yang", "suffix": "" }, { "first": "Diyi", "middle": [], "last": "Yang", "suffix": "" }, { "first": "Chris", "middle": [], "last": "Dyer", "suffix": "" }, { "first": "Xiaodong", "middle": [], "last": "He", "suffix": "" }, { "first": "Alex", "middle": [], "last": "Smola", "suffix": "" }, { "first": "Eduard", "middle": [], "last": "Hovy", "suffix": "" } ], "year": 2016, "venue": "Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NACCL-HLT 2016)", "volume": "", "issue": "", "pages": "1480--1489", "other_ids": {}, "num": null, "urls": [], "raw_text": "Zichao Yang, Diyi Yang, Chris Dyer, Xiaodong He, Alex Smola, and Eduard Hovy. 2016. Hierarchi- cal attention networks for document classification. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computa- tional Linguistics: Human Language Technologies (NACCL-HLT 2016), pages 1480-1489.", "links": null } }, "ref_entries": { "TABREF0": { "type_str": "table", "num": null, "content": "
Vanilla57.12\u00b10.58%57.91\u00b10.81%57.60\u00b10.65%56.70\u00b11.21%
Side-information 57.24\u00b10.47%59.04\u00b10.33%57.97\u00b10.74%57.44\u00b10.62%
Feature-guided58.10\u00b10.50% \u202059.90\u00b10.45% \u202058.30\u00b10.71%58.30\u00b10.65% \u2020
", "text": "ModelCNN BILSTM AVERAGE BILSTM STACKED BILSTM STACKED BILSTM ATT", "html": null }, "TABREF2": { "type_str": "table", "num": null, "content": "
: Results over online petitions. The best result is
indicated in bold, and marked with \" \u2020\" if the improve-
ment is statistically significant (based on a one-tailed
Wilcoxon signed-rank test; p < 0.05).
", "text": "", "html": null } } } }