Looking for Hyponyms in Vector Space

Proceedings of the Eighteenth Conference on Computational Language Learning, pages 68–77, Baltimore, Maryland USA, June 26-27 2014. c⃝2014 Association for Computational Linguistics Looking for Hyponyms in Vector Space Marek Rei SwiftKey 95 Southwark Bridge Rd London, UK marek@swiftkey.net Ted Briscoe Computer Laboratory University of Cambridge Cambridge, UK ted.briscoe@cl.cam.ac.uk Abstract The task of detecting and generating hy- ponyms is at the core of semantic under- standing of language, and has numerous practical applications. We investigate how neural network embeddings perform on this task, compared to dependency-based vector space models, and evaluate a range of similarity measures on hyponym gener- ation. A new asymmetric similarity mea- sure and a combination approach are de- scribed, both of which significantly im- prove precision. We release three new datasets of lexical vector representations trained on the BNC and our evaluation dataset for hyponym generation. 1 Introduction Hyponymy is a relation between two word senses, indicating that the meaning of one word is also contained in the other. It can be thought of as a type-of relation; for example car, ship and train are all hyponyms of vehicle. We denote a hy- ponymy relation between words a and b as (a → b), showing that a is a hyponym of b, and b is a hyper- nym of a. Hyponymy relations are closely related to the concept of entailment, and this notation is consistent with indicating the direction of infer- ence – if a is true, b must be true as well. Automatic detection and generation of hy- ponyms has many practical applications in nearly all natural language processing tasks. Information retrieval, information extraction and question an- swering can be improved by performing appropri- ate query expansions. For example, a user search- ing for arthritis treatment is most likely also inter- ested in results containing the hyponyms of treat- ment, such as arthritis therapy, arthritis medica- tion, and arthritis rehabilitation. Summarisation systems can increase coherence and reduce repe- tition by correctly handling hyponymous words in the input text. Entailment and inference systems can improve sentence-level entailment resolution by detecting the presence and direction of word- level hyponymy relations. Distributionally simi- lar words have been used for smoothing language models and word co-occurrence probabilities (Da- gan et al., 1999; Weeds and Weir, 2005), and hy- ponyms can be more suitable for this application. We distinguish between three different tasks related to hyponyms. Given a directional word pair, the goal of hyponym detection is to deter- mine whether one word is a hyponym of the other (Zhitomirsky-Geffet and Dagan, 2009; Kotlerman et al., 2010; Baroni and Lenci, 2011). In con- trast, hyponym acquisition is the task of extract- ing all possible hyponym relations from a given text (Hearst, 1992; Caraballo, 1999; Pantel and Ravichandran, 2004; Snow et al., 2005). Such sys- tems often make use of heuristic rules and patterns for extracting relations from surface text, and pop- ulate a database with hyponymous word pairs. Fi- nally, the task of hyponym generation is to re- turn a list …

Content truncated to 500 words