Automatic Corpus Extension for Data-driven Natural Language Generation

Automatic Corpus Extension for Data-driven Natural Language Generation Elena Manishina, Bassam Jabaian, St´ephane Huet, Fabrice Lef`evre LIA-CERI, University of Avignon, France, elena.manishina@alumni.univ-avignon.fr, {bassam.jabaian,stephane.huet,fabrice.lefevre}@univ-avignon.fr Abstract As data-driven approaches started to make their way into the Natural Language Generation (NLG) domain, the need for automation of corpus building and extension became apparent. Corpus creation and extension in data-driven NLG domain traditionally involved manual paraphrasing performed by either a group of experts or with resort to crowd-sourcing. Building the training corpora manually is a costly enterprise which requires a lot of time and human resources. We propose to automate the process of corpus extension by integrating automatically obtained synonyms and paraphrases. Our methodology allowed us to significantly increase the size of the training corpus and its level of variability (the number of distinct tokens and specific syntactic structures). Our extension solutions are fully automatic and require only some initial validation. The human evaluation results confirm that in many cases native users favor the outputs of the model built on the extended corpus. Keywords: corpus building, natural language generation, automatic paraphrasing 1. Introduction As data-driven approaches started to make their way into the Natural Language Generation (NLG) domain, the need for automation of corpus building and extension became apparent. Building the training corpora manually is a costly enterprise which requires a lot of time and human resources. Such manually built corpora are often small in size and rather monotone in their content, especially in lim- ited domain tasks. Generation systems built on such cor- pora show little variety in their outputs, thus making the system-user interaction less exciting for the later. The problem of diversifying generation module outputs has recently been gaining interest within the NLG research community, mainly because the users show manifest pref- erence for systems with varied output, which closely re- sembles that of a human interlocutor (Mairesse and Young, 2014). Corpus creation and extension in data-driven NLG domain traditionally involved manual paraphrasing performed by either a group of experts or with resort to crowd-sourcing. Yet, outside NLG, automatic synonym and paraphrase ex- tractions are well established Natural Language Process- ing (NLP) research fields with elaborate methodology and a large number of available tools and resources which are widely used in various NLP applications. Moreover, auto- matic extraction methods have been shown to perform at a level close to that of human annotators (e.g. see (Barzilay and McKeown, 2001)). In our study we propose the fol- lowing extension solutions, which we combine in order to produce a richer and a more diverse training corpus: • Extending the system vocabulary with automati- cally obtained synonyms which consists in replacing open class words (nouns, verbs, adjectives, adverbs) with their synonyms acquired automatically, thus cre- ating new variants of existing sentences; • Diversifying the set of syntactic structures by in- troducing multi-word paraphrases which consists in replacing sub-phrases inside a sentence with auto- matically obtained paraphrases; • Making the system responses more ‘human’ and user-friendly by means of introducing a modal component which is …

Content truncated to 500 words