Introduction to the Special Issue on Word Sense Disambiguation: The State of the Art

ANIMATE, HUMAN, etc. and encode type restrictions on nouns and adjectives and on the arguments of verbs. Subject codes use another set of primitives to classify senses of words by subject (ECONOMICS, ENGINEERING, etc.). Guthrie et al. (1991) demonstrate a typical use of this information: in addition to using the Lesk-based method of counting overlaps between definitions and contexts, they impose a correspondence of subject codes in an iterative process. No quantitative evaluation of this method is available, but Cowie et al. (1992) improve the method using simulated annealing and report results of 47% for sense distinctions and 72% for homographs. The use of LDOCE box codes, however, is problematic: the codes are not systematic (see, for example, Fontenelle, 1990); in later work, Braden-Harder (1993) showed that simply matching box or subject codes is not sufficient for disambiguation. For example, in I tipped the driver, the codes for several senses of the words in the sentence satisfy the necessary constraints (e.g. tip-money + human object or tip-tilt + movable solid object). In many ways, the supplemen7 Note that the assumptions underlying this method are very similar to Quillian’s: Thus one may think of a full concept analogically as consisting of all the information one would have if he looked up what will be called the “patriarch” word in a dictionary, then looked up every word in each of its definitions, then looked up every word found in each of these, and so on, continually branching outward[...] (Quillian, 1968, p. 238). However, Quillian’s network also keeps track of semantic relationships among the words encountered along the path between two words, which are encoded in his semantic network; the neural network avoids the overhead of creating the semantic network but loses this relational information. 13 tary information in the LDOCE, and in particular the subject codes, are similar to those in a thesaurus, which, however, are more systematically structured. Inconsistencies in dictionaries, noted earlier, are not the only and perhaps not the major source of their limitations for WSD. While dictionaries provide detailed information at the lexical level, they lack pragmatic information that enters into sense determination (see, e.g., Hobbs, 1987). For example, the link between ash and tobacco, cigarette or tray in a network such as Quillian’s is very indirect, whereas in the Brown Corpus, the word ash co-occurs frequently with one of these words. It is therefore not surprising that corpora have become a primary source of information for WSD; this development is outlined below in section 2.3. 2.3.2 Thesauri. Thesauri provide information about relationships among words, most notably synonymy. Roget's International Thesaurus, which was put into machine-tractable form in the 1950's8 and has been used in a variety of applications including machine translation (Masterman, 1957), information retrieval (Sparck Jones, 1964, 1986), and content analysis (Sedelow and Sedelow, 1969; see also Sedelow and Sedelow, 1986, 1992), also supplies an explicit concept hierarchy consisting of up to eight increasingly refined levels. Typically, each occurrence of the same word under different categories of the thesaurus represent different senses of that word; i.e., the categories correspond roughly to word senses (Yarowsky, 1992). A set of words in the same category are semantically related. The earliest known use of Roget’s for WSD is the work of Masterman (1957), described above in section 2.1. Several years later, Patrick (1985) used Roget’s to discriminate among verb senses, by examining semantic clusters formed by “e-chains” derived from the thesaurus (Bryan, 1973, 1974; see also Sedelow and Sedelow, 1986). He uses “word-strong neighborhoods,” comprising word groups in low-level semicolon groups, which are the most closely related semantically in the thesaurus, and words connected to the group via chains. He is able to discriminate the correct sense of verbs such as inspire (to raise the spirits vs. to inhale, breathe in, sniff, etc.), question (to doubt vs. to ask a question) with “high reliability.” Bryan's earlier work had already demonstrated that homographs can be distinguished by applying a metric based on relationships defined by his chains (Bryan, 1973, 1974). Similar work is described in Sedelow and Mooney (1988). Yarowsky (1992) derives classes of words by starting with words in common categories in Roget's (4th ed.). A 100-word context of each word in the category is extracted from a corpus (the 1991 electronic text of Grolier's Encyclopedia), and a mutual-information-like statistic is used to identify words most likely to co-occur with the category 8 The work of Masterman (1957) and Sparck Jones (1964) relied on a version of Roget’s that was hand-punched onto cards in the 1950’s; the Sedelow’s (1969) work relied on a machine readable version of the 3rd Edition. Roget’s is now widely available via anonymous ftp from various sites.