A Hierarchical Semantics-Aware Distributional Similarity Scheme

International Joint Conference on Natural Language Processing, pages 596–604, Nagoya, Japan, 14-18 October 2013. A Hierarchical Semantics-Aware Distributional Similarity Scheme∗ Shuqi Sun1, Ke Sun2, Shiqi Zhao2, Haifeng Wang2, Muyun Yang1, and Sheng Li1 1Harbin Institute of Technology, Harbin, China {sqsun,ymy}@mtlab.hit.edu.cn, lisheng@hit.edu.cn 2Baidu, Beijing, China {sunke,zhaoshiqi,wanghaifeng}@baidu.com Abstract The context type and similarity calcula- tion are two essential features of a distri- butional similarity scheme (DSS). In this paper, we propose a hierarchical semantic- aware DSS that exploits semantic rela- tion words as extra context information to guide the similarity calculation. First, we define and extract five types of semantic relations, and then develop relation-based similarities from the distributional similar- ities among the top-ranked relation words. Finally, we integrate various similarities using learning-to-rank technique. Exper- iments show that semantic relations are beneficial to predicting accurate similar- ity. On 6904 pairwise similarity compar- isons, the predictive accuracy of our ap- proach reaches 83.9%, which significantly outperforms the baseline approaches. We also conduct intrinsic analysis by varying the quality of semantic relations and the usage of individual similarities. 1 Introduction Distributional similarity is an essential measure of the semantic relatedness between a pair of linguis- tic objects, including words, phrases, or even sen- tences. Confident distributional similar objects, are useful in various NLP applications, such as word sense disambiguation (Lin, 1997), lexical substitution (McCarthy and Navigli, 2009), para- phrase ranking (Dinu and Lapata, 2010), text clas- sification (Baker and McCallum, 1998), etc. In this paper, we focus on the semantic similarity be- tween words. Well-known implementations of word-level dis- tributional similarity scheme (DSS) mainly fall ∗This work was done when the first author was visiting Baidu. into two categories according to the choice of the context: a) text-window based, and b) depen- dency path based. The former has the advantages of language-independence and computational effi- ciency, while the latter captures finer word-word relationships. However, both approaches focus on the usage aspect of words’ meanings, which is only indirect indicator of the underlying semantic interactions. Meanwhile, a great many successful efforts have been made to extract word-level semantic knowledge from the text, including synonyms / antonyms, sibling terms, hypernyms / hyponyms, holonyms / meronyms, etc. Despite of such delib- erated studies, there are few considerations about how these semantics-oriented outcomes could contribute to the construction of DSS. To realize the full potential of such semantic ev- idences, we propose a semantics-aware DSS by using semantic relation words to guide the sim- ilarity calculation. Our motivation is that words having similar semantic relational words, e.g. sib- ling terms or hypernyms, tend to be more semanti- cally similar. The proposed DSS has a hierarchical layout, in which text-window based distributional similarity is first established from the corpus serv- ing as the basis of the relation-specific similarities for 5 semantic relations. Finally, these similari- ties are linearly combined using learning-to-rank technique into a single measure, capturing both the context distributions and the semantic relations of the target words. Our contribution is three-fold. First, by deriving semantic relation based …

Content truncated to 500 words