Exploiting Open IE for Deriving Multiple Premises Entailment Corpus

Proceedings of Recent Advances in Natural Language Processing, pages 1257–1264, Varna, Bulgaria, Sep 2–4, 2019. https://doi.org/10.26615/978-954-452-056-4_144 1257 Exploiting Open IE for Deriving Multiple Premises Entailment Corpus Martin Víta NLP Centre, Faculty of Informatics Masaryk University Botanická 68a, 602 00 Brno Czech Republic info@martinvita.eu Jakub Klímek Department of Software Engineering Faculty of Mathematics and Physics Charles University Malostranské nám. 2/25, 118 00 Prague 1 Czech Republic klimek@ksi.mff.cuni.cz Abstract Natural language inference (NLI) is a key part of natural language understanding. The NLI task is defined as a decision problem whether a given sentence – hy- pothesis – can be inferred from a given text. Typically, we deal with a text con- sisting of just a single premise/single sen- tence, which is called a single premise en- tailment (SPE) task. Recently, a derived task of NLI from multiple premises (MPE) was introduced together with the first an- notated corpus and corresponding several strong baselines. Nevertheless, the further development in MPE field requires acces- sibility of huge amounts of annotated data. In this paper we introduce a novel method for rapid deriving of MPE corpora from an existing NLI (SPE) annotated data that does not require any additional annotation work. This proposed approach is based on using an open information extraction sys- tem. We demonstrate the application of the method on a well known SNLI cor- pus. Over the obtained corpus, we provide the first evaluations as well as we state a strong baseline. 1 Introduction Natural language inference (NLI), formerly known as recognizing textual entailment (RTE) task – as a part of natural language understand- ing (NLU) – belongs to one of the most prominent problems in NLP for more than ten years. Gener- ally, the NLI task is to classify the relationship be- tween a given text and a given hypothesis: whether the hypothesis can be inferred from the text. The task is typically formulated in a “sentence-pair set- ting”, i. e., the text is just a single sentence. Ac- cording to (Lai et al., 2017), we refer this setting as a single premise entailment (SPE for short). Current state-of-the-art approaches are based on deep learning and/or ensemble methods. Over the years, solid resources for supervised learning for SPE were developed. In contrast, problems related to NLI and/or problems derived from NLI, like re- lation inference (Levy and Dagan, 2016), ques- tion entailment (Abacha and Demner-Fushman, 2016), partial/facet entailment (Levy et al., 2013) and (Nielsen et al., 2009), and others, are strongly under-resourced. For further development in these fields, this fact may be limiting. Recently, a task of NLI from multiple premises was proposed in (Lai et al., 2017) – the idea is based on relaxing the common assumption that the premise is just a single sentence. Again, ac- cording to this paper, we will call this derived of NLI task multiple premises entailment (MPE for short). Similarly to other mentioned entailment tasks, MPE is also under-resourced: to best of our knowledge, there exists only one annotated corpus (introduced in the …

Content truncated to 500 words