{ "paper_id": "P17-1017", "header": { "generated_with": "S2ORC 1.0.0", "date_generated": "2023-01-19T08:15:12.956590Z" }, "title": "Creating Training Corpora for NLG Micro-Planning", "authors": [ { "first": "Claire", "middle": [], "last": "Gardent", "suffix": "", "affiliation": { "laboratory": "", "institution": "LORIA", "location": { "postCode": "7503, F-54500", "settlement": "Vandoeuvre-l\u00e8s-Nancy", "region": "UMR", "country": "France" } }, "email": "claire.gardent@loria.fr" }, { "first": "Anastasia", "middle": [], "last": "Shimorina", "suffix": "", "affiliation": { "laboratory": "", "institution": "LORIA", "location": { "postCode": "7503, F-54500", "settlement": "Vandoeuvre-l\u00e8s-Nancy", "region": "UMR", "country": "France" } }, "email": "anastasia.shimorina@loria.fr" }, { "first": "Shashi", "middle": [], "last": "Narayan", "suffix": "", "affiliation": { "laboratory": "", "institution": "University of Edinburgh", "location": { "addrLine": "10 Crichton Street", "postCode": "EH8 9AB", "settlement": "Edinburgh", "country": "UK" } }, "email": "shashi.narayan@ed.ac.uk" }, { "first": "Laura", "middle": [], "last": "Perez-Beltrachini", "suffix": "", "affiliation": { "laboratory": "", "institution": "University of Edinburgh", "location": { "addrLine": "10 Crichton Street", "postCode": "EH8 9AB", "settlement": "Edinburgh", "country": "UK" } }, "email": "" } ], "year": "", "venue": null, "identifiers": {}, "abstract": "In this paper, we present a novel framework for semi-automatically creating linguistically challenging microplanning data-to-text corpora from existing Knowledge Bases. Because our method pairs data of varying size and shape with texts ranging from simple clauses to short texts, a dataset created using this framework provides a challenging benchmark for microplanning. Another feature of this framework is that it can be applied to any large scale knowledge base and can therefore be used to train and learn KB verbalisers. We apply our framework to DBpedia data and compare the resulting dataset with Wen et al. (2016)'s. We show that while Wen et al.'s dataset is more than twice larger than ours, it is less diverse both in terms of input and in terms of text. We thus propose our corpus generation framework as a novel method for creating challenging data sets from which NLG models can be learned which are capable of handling the complex interactions occurring during in micro-planning between lexicalisation, aggregation, surface realisation, referring expression generation and sentence segmentation. To encourage researchers to take up this challenge, we recently made available a dataset created using this framework in the context of the WEBNLG shared task.", "pdf_parse": { "paper_id": "P17-1017", "_pdf_hash": "", "abstract": [ { "text": "In this paper, we present a novel framework for semi-automatically creating linguistically challenging microplanning data-to-text corpora from existing Knowledge Bases. Because our method pairs data of varying size and shape with texts ranging from simple clauses to short texts, a dataset created using this framework provides a challenging benchmark for microplanning. Another feature of this framework is that it can be applied to any large scale knowledge base and can therefore be used to train and learn KB verbalisers. We apply our framework to DBpedia data and compare the resulting dataset with Wen et al. (2016)'s. We show that while Wen et al.'s dataset is more than twice larger than ours, it is less diverse both in terms of input and in terms of text. We thus propose our corpus generation framework as a novel method for creating challenging data sets from which NLG models can be learned which are capable of handling the complex interactions occurring during in micro-planning between lexicalisation, aggregation, surface realisation, referring expression generation and sentence segmentation. To encourage researchers to take up this challenge, we recently made available a dataset created using this framework in the context of the WEBNLG shared task.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Abstract", "sec_num": null } ], "body_text": [ { "text": "To train Natural Language Generation (NLG) systems, various input-text corpora have been developed which associate (numerical, formal, linguistic) input with text. As discussed in detail in Sec-tion 2, these corpora can be classified into three main types namely, (i) domain specific corpora, (ii) benchmarks constructed from \"Expert\" Linguistic Annotations and (iii) crowdsourced benchmarks. 1 In this paper, we focus on how to create datato-text corpora which can support the learning of micro-planners i.e., data-to-text generation systems that can handle the complex interactions occurring between lexicalisation (mapping data to words), aggregation (exploiting linguistic constructs such as ellipsis and coordination to avoid repetition), surface realisation (using the appropriate syntactic constructs to build sentences), sentence segmentation and referring expression generation.", "cite_spans": [ { "start": 393, "end": 394, "text": "1", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "We start by reviewing the main existing types of NLG benchmarks and we argue for a crowdsourcing approach in which (i) data units are automatically built from an existing Knowledge Base (KB) and (ii) text is crowdsourced from the data (Section 2). We then propose a generic framework for semi-automatically creating training corpora for NLG (Section 3) from existing knowledge bases. In Section 4, we apply this framework to DBpedia data and we compare the resulting dataset with the dataset of Wen et al. (2016) using various metrics to evaluate the linguistic and computational adequacy of both datasets. By applying these metrics, we show that while Wen et al.'s dataset is more than twice larger than ours, it is less diverse both in terms of input and in terms of text. We also com-pare the performance of a sequence-to-sequence model (Vinyals et al., 2015) on both datasets to estimate the complexity of the learning task induced by each dataset. We show that the performance of this neural model is much lower on the new data set than on the existing ones. We thus propose our corpus generation framework as a novel method for creating challenging data sets from which NLG models can be learned which are capable of generating complex texts from KB data.", "cite_spans": [ { "start": 840, "end": 862, "text": "(Vinyals et al., 2015)", "ref_id": "BIBREF15" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "Domain specific benchmarks. Several domain specific data-text corpora have been built by researchers to train and evaluate NLG systems. In the sports domain, Chen and Mooney (2008) constructed a dataset mapping soccer games events to text which consists of 1,539 data-text pairs and a vocabulary of 214 words. For weather forecast generation, the dataset of Liang et al. (2009) includes 29,528 data-text pairs with a vocabulary of 345 words. For the air travel domain, Ratnaparkhi (2000) created a dataset consisting of 5,426 datatext pairs with a richer vocabulary (927 words) and in the biology domain, the KBGen shared task (Banik et al., 2013) made available 284 data-text pairs where the data was extracted from an existing knowledge base and the text was authored by biology experts.", "cite_spans": [ { "start": 358, "end": 377, "text": "Liang et al. (2009)", "ref_id": "BIBREF6" }, { "start": 469, "end": 487, "text": "Ratnaparkhi (2000)", "ref_id": "BIBREF13" }, { "start": 627, "end": 647, "text": "(Banik et al., 2013)", "ref_id": "BIBREF1" } ], "ref_spans": [], "eq_spans": [], "section": "NLG Benchmarks", "sec_num": "2" }, { "text": "An important limitation of these datasets is that, because they are domain specific, systems learned from them are restricted to generating domain specific, often strongly stereotyped text (e.g., weather forecast or soccer game commentator reports). Arguably, training corpora for NLG should support the learning of more generic systems capable of handling a much wider range of linguistic interactions than is present in stereotyped texts. By nature however, domain specific corpora restrict the lexical and often the syntactic coverage of the texts to be produced and thereby indirectly limit the expressivity of the generators trained on them.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "NLG Benchmarks", "sec_num": "2" }, { "text": "Benchmarks constructed from \"expert\" linguistic annotations. NLG benchmarks have also been proposed where the input data is either derived from dependency parse trees (SR'11 task, Belz et al. 2011) or constructed through manual annotation (AMR Corpus, Banarescu et al. 2012) . Contrary to the domain-specific data sets just mentioned, these corpora have a wider coverage and are large enough for training systems that can generate linguistically sophisticated text.", "cite_spans": [ { "start": 180, "end": 197, "text": "Belz et al. 2011)", "ref_id": "BIBREF2" }, { "start": 252, "end": 274, "text": "Banarescu et al. 2012)", "ref_id": "BIBREF0" } ], "ref_spans": [], "eq_spans": [], "section": "NLG Benchmarks", "sec_num": "2" }, { "text": "One main drawback of these benchmarks however is that their construction required massive manual annotation of text with complex linguistic structures (parse trees for the SR task and Abstract Meaning Representation for the AMR corpus). Moreover because these structures are complex, the annotation must be done by experts. It cannot be delegated to the crowd. In short, the creation of such benchmark is costly both in terms of time and in terms of expertise.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "NLG Benchmarks", "sec_num": "2" }, { "text": "Another drawback is that, because the input representation derived from a text is relatively close to its surface form 2 , the NLG task is mostly restricted to surface realisation (mapping input to sentences). That is, these benchmarks give very limited support for learning models that can handle the interactions between micro-planning subtasks.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "NLG Benchmarks", "sec_num": "2" }, { "text": "Crowdsourced benchmarks. More recently, data-to-text benchmarks have also been created by associating data units with text using crowdsourcing.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "NLG Benchmarks", "sec_num": "2" }, { "text": "Wen et al. (2016) first created data by enumerating all possible combinations of 14 dialog act types (e.g., request, inform) and attribute-value pairs present in four small-size, hand-written ontologies about TVs, laptops, restaurants and hotels. They then use crowdsourcing to associate each data unit with a text. The resulting dataset is both large and varied (4 domains) and was successfully exploited to train neural and imitation learning data-to-text generator (Wen et al., 2016; Lampouras and Vlachos, 2016) . Similarly, Novikova and Rieser (2016) described a framework for collecting data-text pairs using automatic quality control measures and evaluating how the type of the input representations (text vs pictures) impacts the quality of crowdsourced text.", "cite_spans": [ { "start": 468, "end": 486, "text": "(Wen et al., 2016;", "ref_id": "BIBREF16" }, { "start": 487, "end": 515, "text": "Lampouras and Vlachos, 2016)", "ref_id": "BIBREF4" } ], "ref_spans": [], "eq_spans": [], "section": "NLG Benchmarks", "sec_num": "2" }, { "text": "The crowdsourcing approach to creating inputtext corpora has several advantages.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "NLG Benchmarks", "sec_num": "2" }, { "text": "First, it is low cost in that the data is produced automatically and the text is authored by a crowdworker. This is in stark contrast with the previous approach where expert linguists are required to align text with data.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "NLG Benchmarks", "sec_num": "2" }, { "text": "Second, because the text is crowd-sourced from the data (rather than the other way round), there is an adequate match between text and data both semantically (the text expresses the information contained in the data) and computationally (the data is sufficiently different from the text to require the learning of complex generation operations such as sentence segmentation, aggregation and referring expression generation).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "NLG Benchmarks", "sec_num": "2" }, { "text": "Third, by exploiting small hand-written ontologies to quickly construct meaningful artificial data, the crowdsourcing approach allows for the easy creation of a large dataset with data units of various size and bearing on different domains. This, in turn, allows for better linguistic coverage and for NLG tasks of various complexity since typically, inputs of larger size increases the need for complex microplanning operations.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "NLG Benchmarks", "sec_num": "2" }, { "text": "Creating Data-to-Text, Micro-Planning Benchmarks", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "The WebNLG Framework for", "sec_num": "3" }, { "text": "While as just noted, the crowdsourcing approach presented by Wen et al. 2016has several advantages, it also has a number of shortcomings. One important drawback is that it builds on artificial rather than \"real\" data i.e., data that would be extracted from an existing knowledge base. As a result, the training corpora built using this method cannot be used to train KB verbalisers i.e., generation systems that can verbalise KB fragments.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "The WebNLG Framework for", "sec_num": "3" }, { "text": "Another limitation concerns the shape of the input data. Wen et al.'s data can be viewed as trees of depth one (a set of attributes-value pairs describing a single entity e.g., a restaurant or a laptop). As illustrated in Figure 1 however, there is a strong correlation between the shape of the input and the syntactic structure of the corresponding sentence. The path structure T 1 where B is shared by two predicates (mission and operator) will favour the use of a participial or a passive subject relative clause. In contrast, the branching structure T 2 will favour the use of a new clause with a pronominal subject or a coordinated VP. More generally, allowing for trees of deeper depth is necessary to indirectly promote the introduction in the benchmark of a more varied set of syntactic constructs to be learned by generators.", "cite_spans": [], "ref_spans": [ { "start": 222, "end": 230, "text": "Figure 1", "ref_id": "FIGREF0" } ], "eq_spans": [], "section": "The WebNLG Framework for", "sec_num": "3" }, { "text": "To address these issues, we introduce a novel method for creating data-to-text corpora from large knowledge bases such as DBPedia. Our T1 A B C mission operator S1.1 A participated in mission B operated by C.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "The WebNLG Framework for", "sec_num": "3" }, { "text": "A participated in mission B which was operated by C.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "S1.2", "sec_num": null }, { "text": "A D E o cc u p at io n b ir th P la ce S2.1 A was born in E. She worked as an engineer.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "T2", "sec_num": null }, { "text": "A was born in E and worked as an engineer. method combines (i) a content selection module designed to extract varied, relevant and coherent data units from DBPedia with (ii) a crowdsourcing process for associating data units with human authored texts that correctly capture their meaning. Example 1 shows a data/text unit created by our method using DBPedia as input KB.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "S2.2", "sec_num": null }, { "text": "( 1) Our method has the following features. First, it can be used to create a data-to-text corpus from any knowledge base where entities are categorised and there is a large number of entities belonging to the same category. As noted above, this means that the resulting corpus can be used to train KB verbalisers i.e., generators that are able to verbalise fragments of existing knowledge bases. It could be used for instance, to verbalise fragments of e.g., MusicBrainz 3 , FOAF 4 or Linked-GeoData. 5 Second, as crowdworkers are required to enter text that matches the data and a majority vote validation process is used to eliminate mis-matched pairs, there is a direct match between text and data. This allows for a clear focus on the non content selection part of generation known as microplanning.", "cite_spans": [ { "start": 502, "end": 503, "text": "5", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "S2.2", "sec_num": null }, { "text": "Third, because data of increasing size is matched with texts ranging from simple clauses to short texts consisting of several sentences, the resulting benchmark is appropriate for exercising the main subtasks of microplanning. For instance, in Example (1) above, given the input shown in (1a), generating (1b) involves lexicalising the occupation property as the phrase worked as (lexicalisation); using PP coordination (born in San Antonio on 1942-08-26) to avoid repeating the word born (aggregation); and verbalising the three triples using a single complex sentence including an apposition, a PP coordination and a transitive verb construction (sentence segmentation and surface realisation).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "S2.2", "sec_num": null }, { "text": "To illustrate the functioning of our benchmark creation framework, we apply it to DBPedia. DBPedia is a multilingual knowledge base that was built from various kinds of structured information contained in Wikipedia (Mendes et al., 2012) . This data is stored as RDF (Resource Description Format) triples of the form (subject, property, object) where the subject is a URI (Uniform Resource Identifier), the property is a binary relation and the object is either a URI or a literal value such as a string, a date or a number. We use an English version of the DBPedia knowledge base which encompasses 6.2M entities, 739 classes, 1,099 properties with reference values and 1,596 properties with typed literal values. 6", "cite_spans": [ { "start": 205, "end": 236, "text": "Wikipedia (Mendes et al., 2012)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "DBPedia", "sec_num": "3.1" }, { "text": "To create data units, we adapted the procedure outlined by Perez-Beltrachini et al. 2016and sketched in Figure 2 . This method can be summarised as follows.", "cite_spans": [], "ref_spans": [ { "start": 104, "end": 112, "text": "Figure 2", "ref_id": "FIGREF1" } ], "eq_spans": [], "section": "Selecting Content", "sec_num": "3.2" }, { "text": "First, DBPedia category graphs are extracted from DBPedia by retrieving up to 500 entity graphs for entities of the same category. 7 For example, we build a category graph for the Astronaut category by collecting, graphs of depth five for 500 entities of types astronaut.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Selecting Content", "sec_num": "3.2" }, { "text": "Next, category graphs are used to learn bi-gram models of DBPedia properties which specify the probability of two properties co-occuring together. Three types of bi-gram models are extracted from category graphs using the SRILM toolkit (Stolcke, 2002) : one model (S-Model) for bigrams occurring in sibling triples (triples with a shared subject); one model (C-Model) for bigrams occurring in chained triples (the object of one triple is the subject of the other); and one model (M-Model) which is a linear interpolation of the sibling and the chain model. The intuition is that these sib-ling and chain models capture different types of coherence, namely, topic-based coherence for the S-Model and discourse-based coherence for the C-Model.", "cite_spans": [ { "start": 236, "end": 251, "text": "(Stolcke, 2002)", "ref_id": "BIBREF14" } ], "ref_spans": [], "eq_spans": [], "section": "Selecting Content", "sec_num": "3.2" }, { "text": "Finally, the content selection task is formulated as an Integer Linear Programming (ILP) problem to select, for a given entity of category C and its entity graph G e , subtrees of G e with maximal bigram probability and varying size (between 1 and 7 RDF triples). We applied this content selection procedure to the DBPedia categories Astronaut (A), Building (B), Monument (M), University (U), Sports team (S) and Written work (W), using the three bi-gram models (S-Model, C-Model, M-Model) and making the number of triples required by the ILP constraint to occur in the output solutions vary between 1 and 7. The results are shown in Table 1 . An input is a set of triples produced by the content selection module. The number of input (#Inputs) is thus the number of distinct sets of triples produced by this module. In contrast, input patterns are inputs where subject and object have been abstracted over. That is, the number of input patterns (#I. Patterns) is the number of distinct sets of properties present in the set of inputs. The number of properties (#Properties) is the number of distinct RDF properties occurring in the dataset. Similarly, the number of entities (#Entities) is the number of distinct RDF subjects and objects occurring in each given dataset.", "cite_spans": [], "ref_spans": [ { "start": 634, "end": 641, "text": "Table 1", "ref_id": "TABREF2" } ], "eq_spans": [], "section": "Selecting Content", "sec_num": "3.2" }, { "text": "We associate data with text using the Crowdflower platform. 8 We do this in four main steps as follows.", "cite_spans": [ { "start": 60, "end": 61, "text": "8", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Associating Content with Text", "sec_num": "3.3" }, { "text": "1. Clarifying properties. One difficulty when collecting texts verbalising sets of DBPedia triples is that the meaning of DBPedia properties may be unclear. We therefore first manually clarified for each category being worked on, those properties which have no obvious lexicalisations (e.g., crew1up was replaced by commander).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Associating Content with Text", "sec_num": "3.3" }, { "text": "Next, we collected three verbalisations for data units of size one, i.e. single triples consisting of a subject, a property and an object. For each such input, crowdworkers were asked to produce a sentence verbalising its content. We used both a priori automatic checks to prevent spamming and a posteriori manual checks to remove incorrect verbalisations. We also monitored crowdworkers as they entered their input and banned those who tried to circumvent our instructions and validators. The automatic checks comprise 12 custom javascript validators implemented in the CrowdFlower platform to block contributor answers which fail to meet requirements such as the minimal time a contributor should stay on page, the minimal length of the text produced, the minimal match of tokens between a triple and its verbalisation and various format restrictions used to detect invalid input. The exact match between a triple and its verbalisation was also prohibited. In addition, after data collection was completed, we manually checked each data-text pair and eliminated from the data set any pair where the text either did not match the information conveyed by the triple or was not a well-formed English sentence.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Getting verbalisations for single triples.", "sec_num": "2." }, { "text": "3. Getting verbalisations for input containing more than one triple. The verbalisations collected for single triples were used to construct input with bigger size. Thus, for input with a number of triples more than one, the crowd was asked to merge the sentences corresponding to each triple (obtained in step 2) into a natural sounding text. In such a way, we diminish the risk of having misinterpretations of the original semantics of a data unit. Contributors were also encouraged to change the order, and the wording of sentences, while writing their texts. For each data unit, we collected three verbalisations.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Getting verbalisations for single triples.", "sec_num": "2." }, { "text": "The verbalisations obtained in Step 3 were verified through crowdsourcing. Each verbalisation collected in Step 3 was displayed to CrowdFlower contributors together with the corresponding set of triples. Then the crowd was asked to assess its fluency, semantic adequacy, and grammaticality. Those criteria were checked by asking the follow- Table 2 : Text statistics from crowdsourcing for triple sets of varying sizes (min/max/avg).", "cite_spans": [], "ref_spans": [ { "start": 341, "end": 348, "text": "Table 2", "ref_id": null } ], "eq_spans": [], "section": "Verifying the quality of the collected texts.", "sec_num": "4." }, { "text": "ing three questions:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Verifying the quality of the collected texts.", "sec_num": "4." }, { "text": "Does the text sound fluent and natural? Does the text contain all and only the information from the data? Is the text good English (no spelling or grammatical mistakes)?", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Verifying the quality of the collected texts.", "sec_num": "4." }, { "text": "We collected five answers per verbalisation. A verbalisation was considered bad, if it received three negative answers in at least one criterion. After the verification step, the total corpus loss was of 8.7%. An example of rejected verbalisation can be found in Example (2). The verbalisation was dropped due to the lack of fluency (awkward lexicalisation of the property club).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Verifying the quality of the collected texts.", "sec_num": "4." }, { "text": "(2) (AEK Athens F.C. manager Gus Poyet) (Gus Poyet club Chelsea F.C.) AEK Athens F.C. are managed by Gus Poyet, who is in Chelsea F.C. Table 2 shows some statistics about the texts obtained using our crowdsourcing procedure for triple sets of size one to seven.", "cite_spans": [], "ref_spans": [ { "start": 135, "end": 142, "text": "Table 2", "ref_id": null } ], "eq_spans": [], "section": "Verifying the quality of the collected texts.", "sec_num": "4." }, { "text": "We now compare a dataset created using our dataset creation framework (henceforth WEBNLG) with the dataset of Wen et al. (2016) 9(henceforth, RNNLG). Example 3 shows a sample data-text pair taken from the RNNLG dataset.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Comparing Benchmarks", "sec_num": "4" }, { "text": "(3) Dialog Moves recommend(name=caerus 33;type=television; screensizerange=medium;family=t5;hasusbport=true) The caerus 33 is a medium television in the T5 family that's USB-enabled.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Comparing Benchmarks", "sec_num": "4" }, { "text": "As should be clear from the discussion in Section 2 and 3, both datasets are similar in that, in both cases, data is built from ontological information and text is crowdsourced from the data. An important difference between the two datasets is that, while the RNNLG data was constructed by enumerating possible combinations of dialog act types and attribute-value pairs, the WEBNLG data is created using a sophisticated content selection procedure geared at producing sets of data 9 https://github.com/shawnwun/RNNLG units that are relevant for a given ontological category and that are varied in terms of size, shape and content. We now investigate the impact of this difference on the two datasets (WEBNLG and RNNLG). To assess the degree to which both datasets support the generation of linguistically varied text requiring complex micro-planning operations, we examine a number of data and text related metrics. We also compare the results of an out-of-the-box sequence-to-sequence model as a way to estimate the complexity of the learning task induced by each dataset. ", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Comparing Benchmarks", "sec_num": "4" }, { "text": "Terminology. The attributes in the RNNLG dataset can be viewed as binary relations between the object talked about (a restaurant, a laptop, a TV or a hotel) and a value. Similarly, in the WEBNLGdataset, DBpedia RDF properties relate a subject entity to an object which can be either an entity or a datatype value. In what follows, we refer to both as attributes. Table 3 shows several statistics which indicate that, while the RNNLG dataset is larger than WEBNLG, WEBNLG is much more diverse in terms of attributes, input patterns and input shapes.", "cite_spans": [], "ref_spans": [ { "start": 363, "end": 370, "text": "Table 3", "ref_id": "TABREF5" } ], "eq_spans": [], "section": "Data Comparison", "sec_num": "4.1" }, { "text": "As illustrated in Example (4) below, different attributes can be lexicalised using different parts of speech. A dataset with a larger number of attributes is therefore more likely to induce texts with greater syntactic variety.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Number of attributes.", "sec_num": null }, { "text": "(4) Verb: X title Y / X served as Y Relational noun: X nationality Y / X's nationality is Y Preposition: X country Y / X is in Y Adjective: X nationality USA / X is American As shown in Table 3 , WEBNLG has a more diverse attribute set than RNNLG both in absolute (172 attributes in WEBNLG against 108 in RNNLG) and in relative terms (RNNLG is a little more than twice as large as WEBNLG).", "cite_spans": [], "ref_spans": [ { "start": 186, "end": 193, "text": "Table 3", "ref_id": "TABREF5" } ], "eq_spans": [], "section": "Number of attributes.", "sec_num": null }, { "text": "Number of input patterns. Since attributes may give rise to lexicalisation with different parts of speech, the sets of attributes present in an input (input pattern) 10 indirectly determine the syntactic realisation of the corresponding text. Hence a higher number of input patterns will favour a higher number of syntactic realisations. This is exemplified in Example (5) where two inputs with the same number of attributes give rise to texts with different syntactic forms. While in Example (5a), the attribute set {country, location, start-Date} is realised by a passive (is located), an apposition (Australia) and a deverbal nominal (its construction), in Example (5b), the attribute set {almaMater, birthPlace, selection} induced a passive (was born) and two VP coordinations (graduated and joined). Again, despite the much larger size of the RNNLG dataset, the number of input patterns in both datasets is almost the same. That is, the relative variety in input patterns is higher in WEBNLG.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Number of attributes.", "sec_num": null }, { "text": "The ratio between number of inputs and the number of input patterns has an important impact both in terms of linguistic diversity and in terms of learning complexity. A large ratio indicates a \"repetitive dataset\" where the same pattern is instantiated a high number of times. While this facilitates learning, this also reduces linguistic coverage (less combinations of structures can be learned) and may induce over-fitting. Note that because datasets are typically delexicalised when training NLG models (cf. e.g., Wen et al. 2015 and Lampouras and Vlachos 2016), at training time, different instantiations of the same input pattern reduce to identical input.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Number of input / Number of input patterns.", "sec_num": null }, { "text": "The two datasets markedly differ on this ratio which is five times lower in WEBNLG. While in WEBNLG, the same pattern is instantiated in average 2.40 times, it is instantiated 10.31 times in average in RNNLG. From a learning perspective, this means that the RNNLG dataset facilitates learning but also makes it harder to assess how well systems trained on it can generalise to handle unseen input.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Number of input / Number of input patterns.", "sec_num": null }, { "text": "Input shape. As mentioned in Section 3, in the RNNLG dataset, all inputs can be viewed as trees of depth one while in the WEBNLG dataset, input may have various shapes. As a result, RNNLG texts will be restricted to syntactic forms which permit expressing such multiple predications of the same entity e.g., subject relative clause, VP and sentence coordination etc. In contrast, the trees extracted by the WEBNLG content selection procedure may be of depth five and therefore allow for further syntactic constructs such as object relative clause and passive participles (cf. Figure 1) .", "cite_spans": [], "ref_spans": [ { "start": 576, "end": 585, "text": "Figure 1)", "ref_id": "FIGREF0" } ], "eq_spans": [], "section": "Number of input / Number of input patterns.", "sec_num": null }, { "text": "We can show this empirically as well that WEBNLG is far more diverse than RNNLG in terms of input shapes. The RNNLG dataset has only 6 distinct shapes and all of them are of depth 1, i.e., all (attribute, value) pairs in an input are siblings to each other. In contrast, the WEBNLG dataset has 58 distinct shapes, out of which only 7 shapes are with depth 1, all others have depth more than 1 and they cover 49.6% of all inputs. Table 4 gives some statistics about the texts contained in each dataset. As illustrated by the contrast between Examples (6) and 7above, text length (number of tokens per text) and the number of sentences per text are strong indicators of the complexity of the generation task. We use the Stanford Part-Of-Speech Tagger and Parser version 3.5.2 (dated 2015-04-20, Manning et al. 2014) to tokenize and to perform sentence segmentation on text. As shown in Table 4 , WEBNLG's texts are longer both in terms of tokens and in terms of number of sentences per text. Another difference between the two datasets is that WEBNLG contains a higher number of text per input thereby providing a better basis for learning paraphrases. The size and the content of the vocabulary is another important factor in ensuring the learning of wide coverage generators. While a large vocabulary makes the learning problem harder, it also allows for larger coverage. WEBNLG exhibits a higher corrected type-token ratio (CTTR), which indicates greater lexical variety, and higher lexical sophistication (LS). Lexical sophistication measures the proportion of relatively unusual or advanced word types in the text. In practice, LS is the proportion of lexical word types (lemma) which are not in the list of 2,000 most frequent words generated from the British National Corpus 11 . Type-token ratio (TTR) is a measure of diversity defined as the ratio of the number of word types to the number of words in a text. To address the fact that this ratio tends to decrease with the size of the corpus, corrected TTR can be used to control for corpus size. It is defined as T / \u221a 2N , where T is the number of types and N the number of tokens.", "cite_spans": [ { "start": 793, "end": 813, "text": "Manning et al. 2014)", "ref_id": "BIBREF8" } ], "ref_spans": [ { "start": 429, "end": 436, "text": "Table 4", "ref_id": "TABREF7" }, { "start": 884, "end": 891, "text": "Table 4", "ref_id": "TABREF7" } ], "eq_spans": [], "section": "Number of input / Number of input patterns.", "sec_num": null }, { "text": "Overall, the results shown in Table 4 indicate that WEBNLG texts are both lexically more diverse (higher corrected type/token ratio) and more sophisticated (higher proportion of unfrequent words) than RNNLG's. They also show a proportionately larger vocabulary for WEBNLG (2,992 types for 290,479 tokens in WEBNLG against 3,524 types for 531,871 tokens in RNNLG).", "cite_spans": [], "ref_spans": [ { "start": 30, "end": 37, "text": "Table 4", "ref_id": "TABREF7" } ], "eq_spans": [], "section": "Text Comparison", "sec_num": "4.2" }, { "text": "Richer and more varied datasets are harder to learn from. As a proof-of-concept study of the comparative difficulty of the two datasets with respect to machine learning, we compare the performance of a sequence-to-sequence model for generation on both datasets.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Neural Generation", "sec_num": "4.3" }, { "text": "We use the multi-layered sequence-to-sequence model with attention mechanism described in (Vinyals et al., 2015 ). 12 The model was trained with 3-layer LSTMs with 512 units each with a batch size of 64 and a learning rate of 0.5.", "cite_spans": [ { "start": 90, "end": 111, "text": "(Vinyals et al., 2015", "ref_id": "BIBREF15" }, { "start": 115, "end": 117, "text": "12", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Neural Generation", "sec_num": "4.3" }, { "text": "To allow for a fair comparison, we use a similar amount of data (13K data-text pairs) for both datasets. As RNNLG is bigger in size than WEBNLG, we constructed a balanced sample of RNNLG which included equal number of instances per category (tv, laptop, etc). We use a 3:1:1 ratio for training, developement and testing. The training was done in two delexicalisation modes: fully and name only. In case of fully delexicalisation, all entities were replaced by their generic terms, whereas in name only mode only subjects were modified in that way. For instance, the triple (FC K\u00f6ln manager Peter St\u00f6ger) was delexicalised as (SportsTeam manager Manager) in the first mode, and as (SportsTeam manager Peter St\u00f6ger) in the second mode. The delexicalisation in sentences was done using the exact match between entities and tokens. For training, we use all the available vocabulary. Input and output vocabulary sizes are reported in Table 5 . Table 5 shows the perplexity results. In both modes, RNNLG yielded lower scores than WEBNLG. This is inline with the observations made above concerning the higher data diversity, larger vocabulary and more complex texts of WEBNLG. Similary, the BLEU score of the generated sentences (Papineni et al., 2002) is lower for WEBNLG suggesting again a dataset that is more complex and therefore more difficult to learn from. ", "cite_spans": [ { "start": 1222, "end": 1245, "text": "(Papineni et al., 2002)", "ref_id": "BIBREF11" } ], "ref_spans": [ { "start": 929, "end": 936, "text": "Table 5", "ref_id": "TABREF9" }, { "start": 939, "end": 946, "text": "Table 5", "ref_id": "TABREF9" } ], "eq_spans": [], "section": "Neural Generation", "sec_num": "4.3" }, { "text": "We presented a framework for building NLG datato-text training corpora from existing knowledge bases.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusion", "sec_num": "5" }, { "text": "One feature of our framework is that datasets created using this framework can be used for training and testing KB verbalisers an in particular, verbalisers for RDF knowledge bases. Following the development of the semantic web, many large scale datasets are encoded in the RDF language (e.g., MusicBrainz, FOAF, LinkedGeoData) and official institutions 13 increasingly publish their data in this format. In this context, our framework is useful both for creating training data from RDF KB verbalisers and to increase the number of datasets available for training and testing NLG.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusion", "sec_num": "5" }, { "text": "Another important feature of our framework is that it permits creating semantically and linguistically diverse datasets which should support the learning of lexically and syntactically, wide coverage micro-planners. We applied our framework to DBpedia data and showed that although twice smaller than the largest corpora currently available for training data-to-text microplanners, the resulting dataset is more semantically and linguistically diverse. Despite the disparity in size, the number of attributes is comparable in the two datasets. The ratio between input and input patterns is five times lower in our dataset thereby making learning harder but also diminishing the risk of overfitting and providing for wider linguistic coverage. Conversely, the ratio of text per input is twice higher thereby providing better support for learning paraphrases. 13 See http://museum-api.pbworks.com for examples.", "cite_spans": [ { "start": 858, "end": 860, "text": "13", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Conclusion", "sec_num": "5" }, { "text": "We have recently released a first version of the WebNLG dataset in the context of a shared task on micro-planning 14 . This new dataset consists of 21,855 data/text pairs with a total of 8,372 distinct data input.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusion", "sec_num": "5" }, { "text": "The input describes entities belonging to 9 distinct DBpedia categories namely, Astronaut, University, Monument, Building, ComicsCharacter, Food, Airport, SportsTeam and WrittenWork. The WebNLG data is licensed under the following license: CC Attribution-Noncommercial-Share Alike 4.0 International and can be downloaded at http://talc1.loria.fr/ webnlg/stories/challenge.html.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusion", "sec_num": "5" }, { "text": "Recently, several sequence-to-sequence models have been proposed for generation. Our experiments suggest that these are not optimal when it comes to generate linguistically complex texts from rich data. More generally, they indicate that the data-to-text corpora built by our framework are challenging for such models. We hope that the WEBNLG dataset which we have made available for the WEBNLG shared task will drive the deep learning community to take up this new challenge and work on the development of neural generators that can handle the generation of KB verbalisers and of linguistically rich texts.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusion", "sec_num": "5" }, { "text": "We ignore here(Lebret et al., 2016)'s dataset which was created fully automatically from Wikipedia by associating infoboxes with text because this dataset fails to ensure an adequate match between data and text. We manually examined 50 input/output pairs randomly extracted from this dataset and did not find a single example where data and text matched. As such, this dataset is ill-suited for training microplanners. Moreover, since its texts contain both missing and additional information, it cannot be used to train joint models for content selection and micro-planning either.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "For instance, the input structures made available by the shallow track of the SR task contain all the lemmas present in the corresponding text. In this case, the generation task is limited to determining (i) the linear ordering and (ii) the full form of the word in the input.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "https://musicbrainz.org/ 4 http://www.foaf-project.org/ 5 http://linkedgeodata.org/", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "http://wiki.dbpedia.org/ dbpedia-dataset-version-2015-107 An entity graph for some entity e is a graph obtained by traversing the DBPedia graph starting in e and stopping at depth five.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "http://www.crowdflower.com", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "Recall from section 3 that input patterns are inputs where subjects and objects have been remove thus, in essence, an input pattern is the set of all the attributes occurring in a given input.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "We compute LS and CTTR using the Lexical Complexity Analyzer developed byLu (2012).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "We used the TensorFlow code available at https://github.com/tensorflow/models/ tree/master/tutorials/rnn/translate. Alternatively, we could have used the implementation of Wen et al. (2016) which is optimised for generation. However the code is geared toward dialog acts and modifying it to handle RDF triples is non trivial. Since the comparison aims at examining the relative performance of the same neural network on the two datasets, we used the tensor flow implementation instead.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "The test data for the WEBNLG challenge will be released on August 18th, 2017 and preliminary results will be presented and discussed at INLG 2017, https:// eventos.citius.usc.es/inlg2017/index.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null } ], "back_matter": [ { "text": "The research presented in this paper was partially supported by the French National Research Agency within the framework of the WebNLG Project (ANR-14-CE24-0033). The third author is supported by the H2020 project SUMMA (under grant agreement 688139).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Acknowledgments", "sec_num": null } ], "bib_entries": { "BIBREF0": { "ref_id": "b0", "title": "Abstract meaning representation (AMR) 1.0 specification", "authors": [ { "first": "Laura", "middle": [], "last": "Banarescu", "suffix": "" }, { "first": "Claire", "middle": [], "last": "Bonial", "suffix": "" }, { "first": "Shu", "middle": [], "last": "Cai", "suffix": "" }, { "first": "Madalina", "middle": [], "last": "Georgescu", "suffix": "" }, { "first": "Kira", "middle": [], "last": "Griffitt", "suffix": "" }, { "first": "Ulf", "middle": [], "last": "Hermjakob", "suffix": "" } ], "year": 2012, "venue": "Proceedings of EMNLP", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Laura Banarescu, Claire Bonial, Shu Cai, Madalina Georgescu, Kira Griffitt, Ulf Hermjakob, Kevin Knight, Philipp Koehn, Martha Palmer, and Nathan Schneider. 2012. Abstract meaning representa- tion (AMR) 1.0 specification. In Proceedings of EMNLP.", "links": null }, "BIBREF1": { "ref_id": "b1", "title": "The KBGen challenge", "authors": [ { "first": "Eva", "middle": [], "last": "Banik", "suffix": "" }, { "first": "Claire", "middle": [], "last": "Gardent", "suffix": "" }, { "first": "Eric", "middle": [], "last": "Kow", "suffix": "" } ], "year": 2013, "venue": "Proceedings of ENLG", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Eva Banik, Claire Gardent, and Eric Kow. 2013. The KBGen challenge. In Proceedings of ENLG.", "links": null }, "BIBREF2": { "ref_id": "b2", "title": "The first surface realisation shared task: Overview and evaluation results", "authors": [ { "first": "Anja", "middle": [], "last": "Belz", "suffix": "" }, { "first": "Michael", "middle": [], "last": "White", "suffix": "" }, { "first": "Dominic", "middle": [], "last": "Espinosa", "suffix": "" }, { "first": "Eric", "middle": [], "last": "Kow", "suffix": "" }, { "first": "Deirdre", "middle": [], "last": "Hogan", "suffix": "" }, { "first": "Amanda", "middle": [], "last": "Stent", "suffix": "" } ], "year": 2011, "venue": "Proceedings of ENLG", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Anja Belz, Michael White, Dominic Espinosa, Eric Kow, Deirdre Hogan, and Amanda Stent. 2011. The first surface realisation shared task: Overview and evaluation results. In Proceedings of ENLG.", "links": null }, "BIBREF3": { "ref_id": "b3", "title": "Learning to sportscast: A test of grounded language acquisition", "authors": [ { "first": "L", "middle": [], "last": "David", "suffix": "" }, { "first": "Raymond J", "middle": [], "last": "Chen", "suffix": "" }, { "first": "", "middle": [], "last": "Mooney", "suffix": "" } ], "year": 2008, "venue": "Proceedings of ICML", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "David L Chen and Raymond J Mooney. 2008. Learn- ing to sportscast: A test of grounded language ac- quisition. In Proceedings of ICML.", "links": null }, "BIBREF4": { "ref_id": "b4", "title": "Imitation learning for language generation from unaligned data", "authors": [ { "first": "Gerasimos", "middle": [], "last": "Lampouras", "suffix": "" }, { "first": "Andreas", "middle": [], "last": "Vlachos", "suffix": "" } ], "year": 2016, "venue": "Proceedings of COLING", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Gerasimos Lampouras and Andreas Vlachos. 2016. Imitation learning for language generation from un- aligned data. In Proceedings of COLING.", "links": null }, "BIBREF5": { "ref_id": "b5", "title": "Neural Text Generation from Structured Data with Application to the Biography Domain", "authors": [ { "first": "R\u00e9mi", "middle": [], "last": "Lebret", "suffix": "" }, { "first": "David", "middle": [], "last": "Grangier", "suffix": "" }, { "first": "Michael", "middle": [], "last": "Auli", "suffix": "" } ], "year": 2016, "venue": "Proceedings of EMNLP", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "R\u00e9mi Lebret, David Grangier, and Michael Auli. 2016. Neural Text Generation from Structured Data with Application to the Biography Domain. In Proceed- ings of EMNLP.", "links": null }, "BIBREF6": { "ref_id": "b6", "title": "Learning Semantic Correspondences with Less Supervision", "authors": [ { "first": "Percy", "middle": [], "last": "Liang", "suffix": "" }, { "first": "Dan", "middle": [], "last": "Michael I Jordan", "suffix": "" }, { "first": "", "middle": [], "last": "Klein", "suffix": "" } ], "year": 2009, "venue": "Proceedings of ACL-IJCNLP", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Percy Liang, Michael I Jordan, and Dan Klein. 2009. Learning Semantic Correspondences with Less Su- pervision. In Proceedings of ACL-IJCNLP.", "links": null }, "BIBREF7": { "ref_id": "b7", "title": "The relationship of lexical richness to the quality of ESL learners' oral narratives", "authors": [ { "first": "Xiaofei", "middle": [], "last": "Lu", "suffix": "" } ], "year": 2012, "venue": "The Modern Language Journal", "volume": "96", "issue": "2", "pages": "190--208", "other_ids": {}, "num": null, "urls": [], "raw_text": "Xiaofei Lu. 2012. The relationship of lexical richness to the quality of ESL learners' oral narratives. The Modern Language Journal 96(2):190-208.", "links": null }, "BIBREF8": { "ref_id": "b8", "title": "The Stanford CoreNLP natural language processing toolkit", "authors": [ { "first": "Christopher", "middle": [ "D" ], "last": "Manning", "suffix": "" }, { "first": "Mihai", "middle": [], "last": "Surdeanu", "suffix": "" }, { "first": "John", "middle": [], "last": "Bauer", "suffix": "" }, { "first": "Jenny", "middle": [], "last": "Finkel", "suffix": "" }, { "first": "Steven", "middle": [ "J" ], "last": "Bethard", "suffix": "" }, { "first": "David", "middle": [], "last": "Mc-Closky", "suffix": "" } ], "year": 2014, "venue": "Proceedings of ACL:System Demonstrations", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Christopher D. Manning, Mihai Surdeanu, John Bauer, Jenny Finkel, Steven J. Bethard, and David Mc- Closky. 2014. The Stanford CoreNLP natural language processing toolkit. In Proceedings of ACL:System Demonstrations.", "links": null }, "BIBREF9": { "ref_id": "b9", "title": "DBpedia: A Multilingual Cross-domain Knowledge Base", "authors": [ { "first": "N", "middle": [], "last": "Pablo", "suffix": "" }, { "first": "Max", "middle": [], "last": "Mendes", "suffix": "" }, { "first": "Christian", "middle": [], "last": "Jakob", "suffix": "" }, { "first": "", "middle": [], "last": "Bizer", "suffix": "" } ], "year": 2012, "venue": "Proceedings of LREC", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Pablo N Mendes, Max Jakob, and Christian Bizer. 2012. DBpedia: A Multilingual Cross-domain Knowledge Base. In Proceedings of LREC.", "links": null }, "BIBREF10": { "ref_id": "b10", "title": "The aNALoGuE challenge: Non aligned language generation", "authors": [ { "first": "Jekaterina", "middle": [], "last": "Novikova", "suffix": "" }, { "first": "Verena", "middle": [], "last": "Rieser", "suffix": "" } ], "year": 2016, "venue": "Proceedings of INLG", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Jekaterina Novikova and Verena Rieser. 2016. The aNALoGuE challenge: Non aligned language gen- eration. In Proceedings of INLG.", "links": null }, "BIBREF11": { "ref_id": "b11", "title": "Bleu: A method for automatic evaluation of machine translation", "authors": [ { "first": "Kishore", "middle": [], "last": "Papineni", "suffix": "" }, { "first": "Salim", "middle": [], "last": "Roukos", "suffix": "" }, { "first": "Todd", "middle": [], "last": "Ward", "suffix": "" }, { "first": "Wei", "middle": [], "last": "Zhu", "suffix": "" } ], "year": 2002, "venue": "Proceedings of ACL", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Kishore Papineni, Salim Roukos, Todd Ward, and Wei jing Zhu. 2002. Bleu: A method for automatic eval- uation of machine translation. In Proceedings of ACL.", "links": null }, "BIBREF12": { "ref_id": "b12", "title": "Building RDF content for Data-to-Text generation", "authors": [], "year": 2016, "venue": "Proceedings of COL-ING", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Laura Perez-Beltrachini, Rania Mohamed Sayed, and Claire Gardent. 2016. Building RDF content for Data-to-Text generation. In Proceedings of COL- ING.", "links": null }, "BIBREF13": { "ref_id": "b13", "title": "Trainable methods for surface natural language generation", "authors": [ { "first": "Adwait", "middle": [], "last": "Ratnaparkhi", "suffix": "" } ], "year": 2000, "venue": "Proceedings of NAACL", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Adwait Ratnaparkhi. 2000. Trainable methods for sur- face natural language generation. In Proceedings of NAACL.", "links": null }, "BIBREF14": { "ref_id": "b14", "title": "SRILM -An extensible language modeling toolkit", "authors": [ { "first": "Andreas", "middle": [], "last": "Stolcke", "suffix": "" } ], "year": 2002, "venue": "Proceedings of ICSLP", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Andreas Stolcke. 2002. SRILM -An extensible lan- guage modeling toolkit. In Proceedings of ICSLP.", "links": null }, "BIBREF15": { "ref_id": "b15", "title": "Grammar as a foreign language", "authors": [ { "first": "Oriol", "middle": [], "last": "Vinyals", "suffix": "" }, { "first": "\u0141ukasz", "middle": [], "last": "Kaiser", "suffix": "" }, { "first": "Terry", "middle": [], "last": "Koo", "suffix": "" }, { "first": "Slav", "middle": [], "last": "Petrov", "suffix": "" }, { "first": "Ilya", "middle": [], "last": "Sutskever", "suffix": "" }, { "first": "Geoffrey", "middle": [], "last": "Hinton", "suffix": "" } ], "year": 2015, "venue": "Proceedings of NIPS", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Oriol Vinyals, \u0141ukasz Kaiser, Terry Koo, Slav Petrov, Ilya Sutskever, and Geoffrey Hinton. 2015. Gram- mar as a foreign language. In Proceedings of NIPS.", "links": null }, "BIBREF16": { "ref_id": "b16", "title": "Multi-domain neural network language generation for spoken dialogue systems", "authors": [ { "first": "Milica", "middle": [], "last": "Tsung-Hsien Wen", "suffix": "" }, { "first": "Nikola", "middle": [], "last": "Ga\u0161i\u0107", "suffix": "" }, { "first": "Lina", "middle": [ "M" ], "last": "Mrk\u0161i\u0107", "suffix": "" }, { "first": "Pei-Hao", "middle": [], "last": "Rojas-Barahona", "suffix": "" }, { "first": "David", "middle": [], "last": "Su", "suffix": "" }, { "first": "Steve", "middle": [], "last": "Vandyke", "suffix": "" }, { "first": "", "middle": [], "last": "Young", "suffix": "" } ], "year": 2016, "venue": "Proceedings of NAACL-HLT", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Tsung-Hsien Wen, Milica Ga\u0161i\u0107, Nikola Mrk\u0161i\u0107, Lina M. Rojas-Barahona, Pei-Hao Su, David Vandyke, and Steve Young. 2016. Multi-domain neural network language generation for spoken di- alogue systems. In Proceedings of NAACL-HLT.", "links": null }, "BIBREF17": { "ref_id": "b17", "title": "Semantically conditioned LSTM-based natural language generation for spoken dialogue systems", "authors": [ { "first": "Milica", "middle": [], "last": "Tsung-Hsien Wen", "suffix": "" }, { "first": "Nikola", "middle": [], "last": "Ga\u0161i\u0107", "suffix": "" }, { "first": "Pei-Hao", "middle": [], "last": "Mrk\u0161i\u0107", "suffix": "" }, { "first": "David", "middle": [], "last": "Su", "suffix": "" }, { "first": "Steve", "middle": [], "last": "Vandyke", "suffix": "" }, { "first": "", "middle": [], "last": "Young", "suffix": "" } ], "year": 2015, "venue": "Proceedings of EMNLP", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Tsung-Hsien Wen, Milica Ga\u0161i\u0107, Nikola Mrk\u0161i\u0107, Pei- Hao Su, David Vandyke, and Steve Young. 2015. Semantically conditioned LSTM-based natural lan- guage generation for spoken dialogue systems. In Proceedings of EMNLP.", "links": null } }, "ref_entries": { "FIGREF0": { "uris": null, "type_str": "figure", "num": null, "text": "Input shape and linguistic structures (A = Susan Helms, B = STS 78, C = NASA, D = engineer, E = Charlotte, North Carolina)." }, "FIGREF1": { "uris": null, "type_str": "figure", "num": null, "text": "Extracting data units from DBPedia." }, "FIGREF2": { "uris": null, "type_str": "figure", "num": null, "text": "(5) a. ('108 St Georges Terrace location Perth', 'Perth country Australia', '108 St Georges Terrace start-Date 1981') country, location, startDate 108 St. Georges Terrace is located in Perth, Australia. Its construction began in 1981. passive, apposition, deverbal nominal b. ('William Anders selection 1963', 'William Anders birthPlace British Hong Kong', 'William Anders almaMater \"AFIT, M.S. 1962\"') almaMater, birthPlace, selection William Anders was born in British Hong Kong, graduated from AFIT in 1962, and joined NASA in 1963. passive, VP coordination, VP coordination" }, "FIGREF3": { "uris": null, "type_str": "figure", "num": null, "text": "(Alan Bean birthDate \"1932-03-15\") Alan Bean was born on March 15, 1932." }, "FIGREF4": { "uris": null, "type_str": "figure", "num": null, "text": "('Alan Bean nationality United States', 'Alan Bean birthDate \"1932-03-15\"', 'Alan Bean almaMater \"UT Austin, B.S. 1955\"', 'Alan Bean birthPlace Wheeler, Texas', 'Alan Bean selection 1963') Alan Bean was an American astronaut, born on March 15, 1932 in Wheeler, Texas. He received a Bachelor of Science degree at the University of Texas at Austin in 1955 and was chosen by NASA in 1963." }, "TABREF2": { "html": null, "text": "", "num": null, "type_str": "table", "content": "
: Data statistics from content selec-
tion (A:Astronaut, B:Building, M:Monument,
U:University, W:Written work, S:Sports team).
" }, "TABREF5": { "html": null, "text": "Comparing WEBNLG and RNNLG datasets. Attributes are properties in RDF triples or slots in dialog acts.", "num": null, "type_str": "table", "content": "" }, "TABREF7": { "html": null, "text": "Text statistics from WEBNLG and RNNLG.", "num": null, "type_str": "table", "content": "
" }, "TABREF9": { "html": null, "text": "Vocabulary sizes of input, output (number of tokens). Perplexity and BLEU scores.", "num": null, "type_str": "table", "content": "
" } } } }