Broad-Coverage Sense Disambiguation and Information Extraction with a Supersense Sequence Tagger

Proceedings of the 2006 Conference on Empirical Methods in Natural Language Processing (EMNLP 2006), pages 594–602, Sydney, July 2006. c⃝2006 Association for Computational Linguistics Broad-Coverage Sense Disambiguation and Information Extraction with a Supersense Sequence Tagger∗ Massimiliano Ciaramita Inst. of Cognitive Science and Technology Italian National Research Council m.ciaramita@istc.cnr.it Yasemin Altun Toyota Technological Institute at Chicago altun@tti-c.org Abstract In this paper we approach word sense disambiguation and information extrac- tion as a unified tagging problem. The task consists of annotating text with the tagset defined by the 41 Wordnet super- sense classes for nouns and verbs. Since the tagset is directly related to Wordnet synsets, the tagger returns partial word sense disambiguation. Furthermore, since the noun tags include the standard named entity detection classes – person, location, organization, time, etc. – the tagger, as a by-product, returns extended named en- tity information. We cast the problem of supersense tagging as a sequential label- ing task and investigate it empirically with a discriminatively-trained Hidden Markov Model. Experimental evaluation on the main sense-annotated datasets available, i.e., Semcor and Senseval, shows consid- erable improvements over the best known “first-sense” baseline. 1 Introduction Named entity recognition (NER) is the most stud- ied information extraction (IE) task. NER typi- cally focuses on detecting instances of “person”, “location”, “organization” names and optionally instances of “miscellaneous” or “time” categories. The scalability of statistical NER allowed re- searchers to apply it successfully on large col- lections of newswire text, in several languages, and biomedical literature. Newswire NER per- formance, in terms of F-score, is in the upper ∗The first author is now at Yahoo! Research. The tag- ger described in this paper is free software and can be down- loaded from http://www.loa-cnr.it/ciaramita.html. 80s (Carreras et al., 2002; Florian et al., 2003), while Bio-NER accuracy ranges between the low 70s and 80s, depending on the data-set used for training/evaluation (Dingare et al., 2005). One shortcoming of NER is its over-simplified onto- logical model, leaving instances of other poten- tially informative categories unidentified. Hence, the utility of named entity information is limited. In addition, instances to be detected are mainly re- stricted to (sequences of) proper nouns. Word sense disambiguation (WSD) is the task of deciding the intended sense for ambiguous words in context. With respect to NER, WSD lies at the other end of the semantic tagging spec- trum, since the dictionary defines tens of thou- sand of very specific word senses, including NER categories. Wordnet (Fellbaum, 1998)1, possibly the most used resource for WSD, defines word senses for verbs, common and proper nouns. Word sense disambiguation, at this level of granularity, is a complex task which resisted all attempts of robust broad-coverage solutions. Many distinc- tions are too subtle to be captured automatically, and the magnitude of the class space – several orders larger than NER’s – makes it hard to ap- proach the problem with sophisticated, but scal- able, machine learning methods. Lastly, even if the methods would scale up, there are not enough manually tagged data, at …

Content truncated to 500 words