Reading Between the Lines: Overcoming Data Sparsity for Accurate Classification of Lexical Relationships

Proceedings of the Fourth Joint Conference on Lexical and Computational Semantics (*SEM 2015), pages 182–192, Denver, Colorado, June 4–5, 2015. Reading Between the Lines: Overcoming Data Sparsity for Accurate Classification of Lexical Relationships Silvia Necs¸ulescu Universitat Pompeu Fabra Barcelona, Spain silvia.necsulescu@upf.edu N´uria Bel Universitat Pompeu Fabra Barcelona, Spain nuria.bel@upf.edu Sara Mendes Universidade de Lisboa Lisboa, Portugal sara.mendes@clul.ul.pt David Jurgens McGill University Montreal, Canada jurgens@cs.mcgill.ca Roberto Navigli Universit`a “La Sapienza” Rome, Italy navigli@di.uniroma1.it Abstract The lexical semantic relationships between word pairs are key features for many NLP tasks. Most approaches for automatically clas- sifying related word pairs are hindered by data sparsity because of their need to observe two words co-occurring in order to detect the lexi- cal relation holding between them. Even when mining very large corpora, not every related word pair co-occurs. Using novel representa- tions based on graphs and word embeddings, we present two systems that are able to predict relations between words, even when these are never found in the same sentence in a given corpus. In two experiments, we demonstrate superior performance of both approaches over the state of the art, achieving significant gains in recall. 1 Introduction Resources containing lexical-semantic relations such as hypernymy or meronymy have proven use- ful in many NLP tasks. While resources such as WordNet (Miller, 1995) contain many general rela- tions and subsequently have seen widespread adop- tion, developing this type of rich resource for new languages or for new domains is prohibitively costly and time-consuming. Therefore, automated ap- proaches are needed and, in order to create such a lexical-semantic database, a first step is to de- velop accurate techniques for classifying the type of lexical-semantic relationship between two words. Approaches to classifying the relationship be- tween a word pair have typically relied on the as- sumption that contexts where word pairs co-occur will yield information on the semantic relation (if any) between them. Given a set of example word pairs having some relation, relation-specific pat- terns may be automatically acquired from the con- texts in which these example pairs co-occur (Tur- ney, 2008b; Mintz et al., 2009). Comparing these relation-specific patterns with those seen with other word pairs measures relational similarity, i.e., how similar is the relation holding between two word pairs. However, any classification system based on patterns of co-occurrence is limited to only those words co-occurring in the data considered; due to the Zipfian distribution of words, even in a very large corpus there are always semantically related word pairs that do not co-occur. As a result, these pattern- based approaches have a strict upper-bound limit on the number of instances that they can classify. As an alternative to requiring co-occurrence, other works have classified the relation of a word pair us- ing lexical similarity, i.e., the similarity of the con- cepts themselves. Given two word pairs, (w1, w2) and (w3, w4), if w1 is lexically similar to w3 and w2 to w4 (i.e., are pair-wise similar) then the pairs are said to have the same semantic …

Content truncated to 500 words