Duality of Link Prediction and Entailment Graph Induction

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4736–4746 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 4736 Duality of Link Prediction and Entailment Graph Induction Mohammad Javad Hosseini⋆§ Shay B. Cohen⋆ Mark Johnson‡ and Mark Steedman⋆ ⋆University of Edinburgh §The Alan Turing Institute, UK ‡Macquarie University javad.hosseini@ed.ac.uk, scohen@inf.ed.ac.uk mark.johnson@mq.edu.au, steedman@inf.ed.ac.uk Abstract Link prediction and entailment graph induc- tion are often treated as different problems. In this paper, we show that these two prob- lems are actually complementary. We train a link prediction model on a knowledge graph of assertions extracted from raw text. We propose an entailment score that exploits the new facts discovered by the link prediction model, and then form entailment graphs be- tween relations. We further use the learned en- tailments to predict improved link prediction scores. Our results show that the two tasks can benefit from each other. The new entail- ment score outperforms prior state-of-the-art results on a standard entialment dataset and the new link prediction scores show improvements over the raw link prediction scores. 1 Introduction Link prediction and entailment graph induction are often treated as different problems. The for- mer (Figure 1A) is used to infer missing relations between entities in existing knowledge graphs (Socher et al., 2013; Bordes et al., 2013; Riedel et al., 2013). The latter (Figure 1B) constructs en- tailment graphs with relations as nodes and entail- ment rules as edges between them (Berant et al., 2011, 2015; Hosseini et al., 2018) for the task of answering questions from text. In this paper, we show that these two problems are complementary by demonstrating how link prediction can help identify entailments and how discovered entail- ments can help predict missing links. Methods to learn entailment graphs (Berant et al., 2011, 2015; Hosseini et al., 2018) process large text corpora to find local entailment scores between relations based on the Distributional In- clusion Hypothesis which states that a word (rela- tion) r entails another word (relation) q if and only if in any context that r can be used, q can be used in its place (Dagan et al., 1999; Geffet and Da- gan, 2005; Kartsaklis and Sadrzadeh, 2016). They E. Macron be nominated for presidency of be elected president of France run for presidency of run for presidency of be nominated for presidency of be elected president of M. Le Pen (A) (B) Figure 1: A link prediction knowledge graph (A) and an entailment graph (B) for entities of types politician,country. The solid lines are discovered correctly, but the dashed ones are missing. However, evidence from the link prediction model can be used to add the missing entailment rule in the entailment graph (B). Similarly, the entailment graph can be used to add the missing link in the knowledge graph (A). use types such as person, location and time, to disambiguate polysemous relations (e.g., person born in location and person born in time). Entail- ment graphs are then formed by …

Content truncated to 500 words