Context-Sensitive Inference Rule Discovery: A Graph-Based Method

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 2902–2911, Osaka, Japan, December 11-17 2016. Context-Sensitive Inference Rule Discovery: A Graph-Based Method Xianpei Han Le Sun State Key Laboratory of Computer Science Institute of Software, Chinese Academy of Sciences Beijing, China {xianpei, sunle}@nfs.iscas.ac.cn Abstract Inference rule discovery aims to identify entailment relations between predicates, e.g., ‘X ac- quire Y à X purchase Y’ and ‘X is author of Y à X write Y’. Traditional methods discover inference rules by computing distributional similarities between predicates, with each predicate is represented as one or more feature vectors of its instantiations. These methods, however, have two main drawbacks. Firstly, these methods are mostly context-insensitive, cannot accurately measure the similarity between two predicates in a specific context. Secondly, traditional meth- ods usually model predicates independently, ignore the rich inter-dependencies between predi- cates. To address the above two issues, this paper proposes a graph-based method, which can discover inference rules by effectively modelling and exploiting both the context and the inter- dependencies between predicates. Specifically, we propose a graph-based representation— Predicate Graph, which can capture the semantic relevance between predicates using both the predicate-feature co-occurrence statistics and the inter-dependencies between predicates. Based on the predicate graph, we propose a context-sensitive random walk algorithm, which can learn context-specific predicate representations by distinguishing context-relevant information from context-irrelevant information. Experimental results show that our method significantly outper- forms traditional inference rule discovery methods. 1 Introduction Inference rule discovery aims to identify entailment relations between predicates, such as ‘X acquire Y à X purchase Y’ and ‘X is author of Y à X write Y’, with each predicate is a textual pattern with (two) variable slots (X and Y in above). Inference rules are important in many fields such as Question Answer- ing (Ravichandran and Hovy, 2002), Textual Entailment (Dagan et al., 2006) and Information Extraction (Hearst, 1992). For example, given the problem “Which company purchases WhatsApp?”, a QA system can extract the answer “Facebook” from the sentence “Facebook acquires WhatsApp for $19 billion” based on the inference rule ‘X acquire Y à X purchase Y’. Given a set of predicates and their instantiations in a large corpus, most traditional methods identify inference rules by computing distributional similarities between predicates, where each predicate is rep- resented as one or more feature vectors of its variable instantiations. For example, given the predicates and their instantiations in Figure 1, we can represent ‘X acquire Y’ as {X=‘Google’, Y=‘YouTube’, X=‘children’, Y=‘skill’} and measure the similarity between ‘X acquire Y’ and ‘X purchase Y’ based on their common features {X=‘Google’, Y=‘YouTube’}. To achieve the above goal, many similarity measures have been proposed for inference rule discovery, such as DIRT Similarity (Lin and Pantel, 2001), Balanced-Inclusion similarity (Szpektor and Dagan, 2008) and Soft Set Inclusion similarity (Nakashole et al., 2012), etc. This work is licensed under a Creative Commons Attribution 4.0 International Licence. Licence details: http://creativecommons.org/licenses/by/4.0/ 2902 Figure 1. Some predicates and their variable instantiations These distributional similarity based …

Content truncated to 500 words