Combining ConceptNet and WordNet for Word Sense Disambiguation

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 686–694, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Combining ConceptNet and WordNet for Word Sense Disambiguation Junpeng Chen Computer School, Wuhan University, Wuhan, P. R. China chenjp@whu.edu.cn Juan Liu* Computer School, Wuhan University, Wuhan, P. R. China liujuan@whu.edu.cn Abstract Knowledge-based Word sense Disambiguation (WSD) methods heavily depend on knowledge. Therefore enriching knowledge is one of the most important issues in WSD. This paper proposes a novel idea of combining WordNet and ConceptNet for WSD. First, we present a novel method to automatically disambiguate the concepts in ConceptNet; and then we en- rich WordNet with large amounts of semantic relations from the disambiguated ConceptNet for WSD. The evaluation experiments on the Semeval-2007 coarse-grained all-words dis- ambiguation task show that the enriched WordNet can significantly improve the per- formance of knowledge-based WSD methods. 1 Introduction & Motivation 1.1 Introduction Word sense Disambiguation (WSD) is a task of selecting the proper sense of ambiguous terms in the text. WSD is an intermediary step within many Natural Language Processing (NLP) appli- cations, such as text summarization, machine translation, text processing, and so on. Finding a solution to the WSD problem is essential or even compulsory for such NLP applications. There are two main classes of WSD ap- proaches: supervised WSD, and knowledge- based WSD. The former employs statistical learning to learn a classifier from training data. The preparation for these training data may be laborious and erroneous. Moreover, building a manually annotated corpus, as required by su- pervised WSD methods, to cover all word senses in lexicon is expensive, and even infeasible. In contrast, knowledge-based WSD methods rely on the use of wide-coverage knowledge resources. These methods do not need the human labeled training data. The widely used knowledge re- source in such methods is WordNet (Fellbaum, 1998). However, WordNet-based WSD methods usually achieved lower performance compared to supervised methods, mainly due the fact that the lexical and semantic knowledge contained in WordNet is not sufficient for WSD. Therefore, many methods (see Section2) have emerged to enrich WordNet with the lexical and semantic knowledge for WSD purpose, such as by using Wikipedia, on-line lexicons, domain, and so on. The literature research results show that the use of enriched knowledge does improve the performance of WordNet based WSD sys- tems, in terms of both accuracy and coverage. However, most of the present methods mainly focus on enriching WordNet with limited lexical and taxonomic knowledge. Though Wikipedia contains many semantic relations, the knowledge extracted from Wikipedia may contain noises, for some of them are derived from weak seman- tic links, and lack of confidence. In this paper, we propose to use ConceptNet (Havasi&Alonso, 2007) with the large amounts of semantic relations between concepts, to enrich WordNet. For there are many ambiguous con- cepts existed, ConceptNet cannot be directly used to enrich WordNet. Thus, we develop a novel methodology to automatically disambigu- ate the ambiguous ConceptNet concepts. By us- ing the disambiguated …

Content truncated to 500 words