Supersense Tagging of Unknown Nouns Using Semantic Similarity

Proceedings of the 43rd Annual Meeting of the ACL, pages 26–33, Ann Arbor, June 2005. c⃝2005 Association for Computational Linguistics Supersense Tagging of Unknown Nouns using Semantic Similarity James R. Curran School of Information Technologies University of Sydney NSW 2006, Australia james@it.usyd.edu.au Abstract The limited coverage of lexical-semantic re- sources is a significant problem for NLP sys- tems which can be alleviated by automati- cally classifying the unknown words. Su- persense tagging assigns unknown nouns one of 26 broad semantic categories used by lex- icographers to organise their manual inser- tion into WORDNET. Ciaramita and Johnson (2003) present a tagger which uses synonym set glosses as annotated training examples. We describe an unsupervised approach, based on vector-space similarity, which does not require annotated examples but significantly outper- forms their tagger. We also demonstrate the use of an extremely large shallow-parsed corpus for calculating vector-space semantic similarity. 1 Introduction Lexical-semantic resources have been applied successful to a wide range of Natural Language Processing (NLP) problems ranging from collocation extraction (Pearce, 2001) and class-based smoothing (Clark and Weir, 2002), to text classification (Baker and McCallum, 1998) and question answering (Pasca and Harabagiu, 2001). In par- ticular, WORDNET (Fellbaum, 1998) has significantly in- fluenced research in NLP. Unfortunately, these resource are extremely time- consuming and labour-intensive to manually develop and maintain, requiring considerable linguistic and domain expertise. Lexicographers cannot possibly keep pace with language evolution: sense distinctions are contin- ually made and merged, words are coined or become obsolete, and technical terms migrate into the vernacu- lar. Technical domains, such as medicine, require sepa- rate treatment since common words often take on special meanings, and a significant proportion of their vocabu- lary does not overlap with everyday vocabulary. Bur- gun and Bodenreider (2001) compared an alignment of WORDNET with the UMLS medical resource and found only a very small degree of overlap. Also, lexical- semantic resources suffer from: bias towards concepts and senses from particular topics. Some specialist topics are better covered in WORD- NET than others, e.g. dog has finer-grained distinc- tions than cat and worm although this does not re- flect finer distinctions in reality; limited coverage of infrequent words and senses. Cia- ramita and Johnson (2003) found that common nouns missing from WORDNET 1.6 occurred every 8 sentences in the BLLIP corpus. By WORDNET 2.0, coverage has improved but the problem of keeping up with language evolution remains difficult. consistency when classifying similar words into cate- gories. For instance, the WORDNET lexicographer file for ionosphere (location) is different to exo- sphere and stratosphere (object), two other layers of the earth’s atmosphere. These problems demonstrate the need for automatic or semi-automatic methods for the creation and mainte- nance of lexical-semantic resources. Broad semantic classification is currently used by lexicographers to or- ganise the manual insertion of words into WORDNET, and is an experimental precursor to automatically insert- ing words directly into the WORDNET hierarchy. Cia- ramita and Johnson (2003) call this supersense tagging and describe a multi-class perceptron tagger, which uses …

Content truncated to 500 words