A Two Level Model for Context Sensitive Inference Rules

Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, pages 1331–1340, Sofia, Bulgaria, August 4-9 2013. c⃝2013 Association for Computational Linguistics A Two Level Model for Context Sensitive Inference Rules Oren Melamud§, Jonathan Berant†, Ido Dagan§, Jacob Goldberger♦, Idan Szpektor‡ § Computer Science Department, Bar-Ilan University † Computer Science Department, Stanford University ♦ Faculty of Engineering, Bar-Ilan University ‡ Yahoo! Research Israel {melamuo,dagan,goldbej}@{cs,cs,eng}.biu.ac.il joberant@stanford.edu idan@yahoo-inc.com Abstract Automatic acquisition of inference rules for predicates has been commonly ad- dressed by computing distributional simi- larity between vectors of argument words, operating at the word space level. A re- cent line of work, which addresses context sensitivity of rules, represented contexts in a latent topic space and computed similar- ity over topic vectors. We propose a novel two-level model, which computes simi- larities between word-level vectors that are biased by topic-level context repre- sentations. Evaluations on a naturally- distributed dataset show that our model significantly outperforms prior word-level and topic-level models. We also release a first context-sensitive inference rule set. 1 Introduction Inference rules for predicates have been identi- fied as an important component in semantic ap- plications, such as Question Answering (QA) (Ravichandran and Hovy, 2002) and Information Extraction (IE) (Shinyama and Sekine, 2006). For example, the inference rule ‘X treat Y → X relieve Y’ can be useful to extract pairs of drugs and the illnesses which they relieve, or to answer a ques- tion like “Which drugs relieve headache?”. Along this vein, such inference rules constitute a crucial component in generic modeling of textual infer- ence, under the Textual Entailment paradigm (Da- gan et al., 2006; Dinu and Wang, 2009). Motivated by these needs, substantial research was devoted to automatic learning of inference rules from corpora, mostly in an unsupervised dis- tributional setting. This research line was mainly initiated by the highly-cited DIRT algorithm (Lin and Pantel, 2001), which learns inference for bi- nary predicates with two argument slots (like the rule in the example above). DIRT represents a predicate by two vectors, one for each of the ar- gument slots, where the vector entries correspond to the argument words that occurred with the pred- icate in the corpus. Inference rules between pairs of predicates are then identified by measuring the similarity between their corresponding argument vectors. This general scheme was further en- hanced in several directions, e.g. directional sim- ilarity (Bhagat et al., 2007; Szpektor and Dagan, 2008) and meta-classification over similarity val- ues (Berant et al., 2011). Consequently, several knowledge resources of inference rules were re- leased, containing the top scoring rules for each predicate (Schoenmackers et al., 2010; Berant et al., 2011; Nakashole et al., 2012). The above mentioned methods provide a sin- gle confidence score for each rule, which is based on the obtained degree of argument-vector sim- ilarities. Thus, a system that applies an infer- ence rule to a text may estimate the validity of the rule application based on the pre-specified rule score. However, the validity of an inference rule may …

Content truncated to 500 words