Multivalent Entailment Graphs for Question Answering

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 10758–10768 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 10758 Multivalent Entailment Graphs for Question Answering Nick McKenna† Liane Guillou† Mohammad Javad Hosseini†‡∗ Sander Bijl de Vroe† Mark Johnson§ Mark Steedman† †University of Edinburgh ‡Alan Turing Institute, UK §Oracle Digital Assistant {nick.mckenna, javad.hosseini, sbdv}@ed.ac.uk {lguillou, steedman}@inf.ed.ac.uk mark.mj.johnson@oracle.com Abstract Drawing inferences between open-domain nat- ural language predicates is a necessity for true language understanding. There has been much progress in unsupervised learning of entail- ment graphs for this purpose. We make three contributions: (1) we reinterpret the Distribu- tional Inclusion Hypothesis to model entail- ment between predicates of different valencies, like DEFEAT(Biden, Trump) ⊨ WIN(Biden); (2) we actualize this theory by learning un- supervised Multivalent Entailment Graphs of open-domain predicates; and (3) we demon- strate the capabilities of these graphs on a novel question answering task. We show that directional entailment is more helpful for infer- ence than non-directional similarity on ques- tions of fine-grained semantics. We also show that drawing on evidence across valencies an- swers more questions than by using only the same valency evidence. 1 Introduction We are reading a mystery about a dark and forebod- ing manor and have one question: “is Mr. Boddy dead?”1 Our text might say “Colonel Mustard killed Mr. Boddy,” or “Mr. Boddy was murdered in the kitchen with a candlestick,” either of which answers the question, but only via natural language inference. An Entailment Graph (EG) is a structure of meaning postulates supporting these inferences such as “if A kills B, then B is dead.” Entailment Graphs contain vertices of open- domain natural language predicates and entail- ments between them are represented as directed edges. Previous models learn predicates of a single valency, the number and types of arguments con- trolled by the predicate. Commonly these are bi- nary graphs, which cannot model single-argument predicates like the entity states “is dead” or “is an ∗Now at Google Research. 1The murder mystery board game Clue (also known as Cluedo) lends inspiration to this project. author.” This means they miss a variety of entail- ments in text that could be used to answer questions such as our example. The Distributional Inclusion Hypothesis (DIH) (Dagan et al., 1999; Kartsaklis and Sadrzadeh, 2016) is a theory which has been used effectively in unsupervised learning of these same-valency entailment graphs (Geffet and Dagan, 2005; Berant et al., 2010; Hosseini, 2021). In this work the DIH is reinterpreted in a way which supports learning entailments between pred- icates of different valencies such as KILL(Mustard, Boddy) ⊨ DIE(Boddy). We extend the work of Hos- seini et al. (2018) and develop a new Multivalent Entailment Graph (MGraph) where vertices may be predicates of different valencies. This results in new kinds of entailments that answer a broader range of questions including entity state. We further pose a true-false question answer- ing task generated automatically from news text. Our model draws inferences across propositions of different valencies to answer …

Content truncated to 500 words