Building RDF Content for Data-to-Text Generation

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 1493–1502, Osaka, Japan, December 11-17 2016. Building RDF Content for Data-to-Text Generation Laura Perez-Beltrachini CNRS/LORIA Nancy (France) laura.perez@loria.fr Rania Mohamed Sayed Universit´e de Lorraine Nancy (France) rania.mohamed.sayed@gmail.com Claire Gardent CNRS/LORIA Nancy (France) claire.gardent@loria.fr Abstract In Natural Language Generation (NLG), one important limitation is the lack of common bench- marks on which to train, evaluate and compare data-to-text generators. In this paper, we make one step in that direction and introduce a method for automatically creating an arbitrary large repertoire of data units that could serve as input for generation. Using both automated metrics and a human evaluation, we show that the data units produced by our method are both diverse and coherent. 1 Introduction In Natural Language Generation, one important limitation is the lack of common benchmarks on which to train, evaluate and compare data-to-text generators. In this paper, we make one step in that direction and introduce a method to automatically create an arbitrary large repertoire of data units which could serve as input for data-to-text generation. We focus on generation from RDFS data where the communicative goal is to describe entities of various categories (e.g., astronauts or monuments). RDF data consists of (subject property object) triples (e.g., (Alan Bean occupation Test pilot)) – as illustrated in Figure 1, RDF data can be represented by a graph in which edges are labelled with properties and vertices with subject and object resources. To construct a corpus of RDF data units which could serve as input for NLG, we introduce a content selection method which, given some DBPedia entity, retrieves DBPedia subgraphs that encode relevant and coherent knowledge about that entity. Our approach differs from previous work on content selection in that it leverages the categorial infor- mation provided by large scale knowledge bases about entities of a given ontological type. Based on this ontological knowledge, we learn two types of category-specific bigram models: one model (S-Model) for bigrams occurring in sibling triples (triples with a share subject) and one model (C-Model) for bigrams occurring in chained triples (the object of one triple is the subject of the other). The intuition is that these two models capture different types of coherence, namely, topic-based coherence for the S-Model and discourse-based coherence for the C-Model. Using these bigram models of RDF properties, we formulate the content selection task as an Inte- ger Linear Programming problem and select for a given entity of category C, subgraphs with maximal probability that is, subgraphs which contain properties that are true of that entity, that are typical of that category and that support the generation of a coherent text. We evaluate the impact of our n-gram models on content selection (how well do they help support the selection of a coherent and diverse set of data units?) using quantitative metrics, a human evaluation and a qualitative analysis. 2 Related Work Our approach has similarity with approaches on entity summarisation, content planning from DBpe- …

Content truncated to 500 words