Combining Collocations, Lexical and Encyclopedic Knowledge for Metonymy Resolution

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 910–918, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Combining Collocations, Lexical and Encyclopedic Knowledge for Metonymy Resolution Vivi Nastase and Michael Strube EML Research gGmbH Heidelberg, Germany http://www.eml-research.de/nlp Abstract This paper presents a supervised method for resolving metonymies. We enhance a commonly used feature set with fea- tures extracted based on collocation in- formation from corpora, generalized us- ing lexical and encyclopedic knowledge to determine the preferred sense of the potentially metonymic word using meth- ods from unsupervised word sense disam- biguation. The methodology developed addresses one issue related to metonymy resolution – the influence of local context. The method developed is applied to the metonymy resolution task from SemEval 2007. The results obtained, higher for the countries subtask, on a par for the compa- nies subtask – compared to participating systems – confirm that lexical, encyclo- pedic and collocation information can be successfully combined for metonymy res- olution. 1 Introduction Metonymies are a pervasive phenomenon in language. They occur because in communicating, we use words as pointers to a larger body of knowledge, that encompasses various facets of the concept evoked by a given word. A listener need not understand the cello to be moved by its playing, just as it is unnecessary for a rider to understand technical jargon; all that matters is sensation, and here the Kawasaki excels. The cockpit is sensibly designed, with a narrow front seat portion ... Kawasaki is a company, it has an organization, fa- cilities, employees, it makes specific products. In the context above, the company name stands in for its products – motorcycles. Motorcycles have parts, the cockpit and front seat are some of them, and this provides the discourse links between the two sentences. Constraints on the interpretation of a word w in context comes both from the local and global context, and are applied to the infor- mation/knowledge evoked by w. The local con- straints come from the words with which w is (grammatically) related to. The global constraints come from the domain/topic of the text, discourse relations that span across sentences. Metonymic words have a rather small num- ber of possible interpretations (also called read- ings) which occur frequently (Markert and Nissim, 2002). Idiosyncratic interpretations are also pos- sible, but very rare. One can view the possible interpretations of a potentially metonymic word (PMW) as corresponding to the word’s possible senses (Nissim and Markert, 2003), bringing the task close to word sense disambiguation. The approach to metonymy resolution pre- sented here is supervised, with unsupervised fea- ture enrichment. We apply techniques inspired by unsupervised word sense disambiguation, which allow us to go beyond the annotated data provided in training, and quantify the restrictions imposed on the interpretation of a PMW by its grammat- ically related neighbours through collocation in- formation extracted from corpora. The only anno- tation required for the corpora are automatically induced part-of-speech tags from which we ob- tain grammatical relations through …

Content truncated to 500 words