Creating Training Corpora for NLG Micro-Planners

Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 179–188 Vancouver, Canada, July 30 - August 4, 2017. c⃝2017 Association for Computational Linguistics https://doi.org/10.18653/v1/P17-1017 Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 179–188 Vancouver, Canada, July 30 - August 4, 2017. c⃝2017 Association for Computational Linguistics https://doi.org/10.18653/v1/P17-1017 Creating Training Corpora for NLG Micro-Planning Claire Gardent Anastasia Shimorina CNRS, LORIA, UMR 7503 Vandoeuvre-l`es-Nancy, F-54500, France {claire.gardent,anastasia.shimorina}@loria.fr Shashi Narayan Laura Perez-Beltrachini School of Informatics, University of Edinburgh 10 Crichton Street, Edinburgh, EH8 9AB, UK {shashi.narayan,lperez}@ed.ac.uk Abstract In this paper, we present a novel framework for semi-automatically cre- ating linguistically challenging micro- planning data-to-text corpora from ex- isting Knowledge Bases. Because our method pairs data of varying size and shape with texts ranging from simple clauses to short texts, a dataset created us- ing 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 creat- ing challenging data sets from which NLG models can be learned which are capable of handling the complex interactions oc- curring during in micro-planning between lexicalisation, aggregation, surface reali- sation, referring expression generation and sentence segmentation. To encourage re- searchers to take up this challenge, we re- cently made available a dataset created us- ing this framework in the context of the WEBNLG shared task. 1 Introduction To train Natural Language Generation (NLG) sys- tems, various input-text corpora have been devel- oped which associate (numerical, formal, linguis- tic) 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” Lin- guistic Annotations and (iii) crowdsourced bench- marks.1 In this paper, we focus on how to create data- to-text corpora which can support the learning of micro-planners i.e., data-to-text generation sys- tems that can handle the complex interactions occurring between lexicalisation (mapping data to words), aggregation (exploiting linguistic con- structs such as ellipsis and coordination to avoid repetition), surface realisation (using the appropri- ate syntactic constructs to build sentences), sen- tence segmentation and referring expression gen- eration. We start by reviewing the main existing types of NLG benchmarks and we argue for a crowdsourc- ing approach in which (i) data units are automati- cally built from an existing Knowledge Base (KB) and (ii) text is crowdsourced from the data (Sec- tion 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 …

Content truncated to 500 words