Using Three Way Data for Word Sense Discrimination

Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), pages 929–936 Manchester, August 2008 Using Three Way Data for Word Sense Discrimination Tim Van de Cruys Humanities Computing University of Groningen t.van.de.cruys@rug.nl Abstract In this paper, an extension of a dimen- sionality reduction algorithm called NON- NEGATIVE MATRIX FACTORIZATION is presented that combines both ‘bag of words’ data and syntactic data, in order to find semantic dimensions according to which both words and syntactic relations can be classified. The use of three way data allows one to determine which dimen- sion(s) are responsible for a certain sense of a word, and adapt the corresponding feature vector accordingly, ‘subtracting’ one sense to discover another one. The intuition in this is that the syntactic fea- tures of the syntax-based approach can be disambiguated by the semantic dimensions found by the bag of words approach. The novel approach is embedded into cluster- ing algorithms, to make it fully automatic. The approach is carried out for Dutch, and evaluated against EuroWordNet. 1 Introduction Automatically acquiring semantics from text is a subject that has gathered a lot of attention for quite some time now. As Manning and Sch¨utze (Man- ning and Sch¨utze, 2000) point out, most work on acquiring semantic properties of words has fo- cused on semantic similarity. ‘Automatically ac- quiring a relative measure of how similar a word is to known words [...] is much easier than deter- mining what the actual meaning is.’ (Manning and Sch¨utze, 2000, §8.5) c⃝ 2008. Licensed under the Creative Commons Attribution-Noncommercial-Share Alike 3.0 Unported li- cense (http://creativecommons.org/licenses/by-nc-sa/3.0/). Some rights reserved. Most work on semantic similarity relies on the distributional hypothesis (Harris, 1985). This hy- pothesis states that words that occur in similar con- texts tend to be similar. With regard to the context used, two basic approaches exist. One approach makes use of ‘bag of words’ co-occurrence data; in this approach, a certain window around a word is used for gathering co-occurrence information. The window may either be a fixed number of words, or the paragraph or document that a word appears in. Thus, words are considered similar if they ap- pear in similar windows (documents). One of the dominant methods using this method is LATENT SEMANTIC ANALYSIS (LSA). The second approach uses a more fine grained distributional model, focusing on the syntactic re- lations that words appear with. Typically, a large text corpus is parsed, and dependency triples are extracted.1 Words are considered similar if they appear with similar syntactic relations. Note that the former approach does not need any kind of linguistic annotation, whereas for the latter, some form of syntactic annotation is needed. The results yielded by both approaches are typ- ically quite different in nature: the former ap- proach typically puts its finger on a broad, the- matic kind of similarity, while the latter approach typically grasps a tighter, synonym-like similarity. Example (1) shows the difference between both approaches; for each approach, the top ten most similar nouns to the Dutch …

Content truncated to 500 words