Word Sense Disambiguation (WSD) is essential for language understanding systems such as in­formation retrieval, summarization, and machine translation systems (Dagan and Itai, 1994; Schutze and Pedersen, 1995; Ng and Zelle, 1997). In par­ticular due to the rapid development of other is­sues in computational linguistics, WSD has been considered the next important task to be solved. Among various WSD tasks, the lexical sample task can achieve a precision rate more than 70% in Chinese, so can the all-words task in English, but currently no Chinese all-words WSD system is available. This study proposes an all-words WSD system conducted on a specific domain which can achieve a 55.9% precision rate.

This system makes use of certain characteristics of WordNet. First, the sense inventory in Chinese WordNet is ordered by the "prototypicality" of the words. In other words, the first sense of a word with multiple senses will be the prototype mean­ing of that word. In addition to semantic relations and sense definitions, Chinese WordNet also in­cludes sense axes which indicate the relations be­tween Chinese senses and corresponding English senses.

Two heuristic rules are devised to characterize domain texts: In a domain text, domain senses are more likely to occur in words if they have one (Heuristic Rule 1); on the other hand, for words with no domain senses, the most generic usages (prototype senses) are more likely to be adopted (Heuristic Rule 2). Therefore, as proposed by Li et al.(1995) for the WordNet-based domain-independent texts WSD task, two heuristic rules (HR) are taken into consideration in the domain WSD test:

for all senses Sk of w do if w has domain sense

choose domain sense Sk else choose prototype sense s\

Figure 1 : Heuristic Rules based WSD

Besides, sense definitions from WordNet were also tested with simplified Lesk algorithm (Lesk, 1986; Kilgarriff and Rosenzweig, 2000) in another experiment to examine the effect of considering sense definitions in domain WSD:

choose domain sense Sk elseif Dk overlaps with C :

choose sense Sk with Dk

that overlaps the most else:

Figure 2: HR with simplified Lesk Algorithm. Dk is the set of content words occurring in the dic­tionary definition of sense Sk- C is the set of con­tent words in the context.

Before the test only preprocessing including seg­mentation and parts of speech tagging will be applied to the target texts, in order to eliminate those senses of the same word form in other parts of speech; the background documents provided by SemEval-2010 are not used for training since this is not a supervised system. According to Wang (2002), with preprocessing of PoS tagging alone, 20% of word sense ambiguity can be distin­guished.

Since the current number of semantic relations in Chinese WordNet is still less than that in En­glish WordNet (PWN), to detect domain senses, the sense axes in Chinese WordNet are exploited. By seeding with English words such as "environ­ment" and "ecology," all English words related to these seed words can be captured with the help of the semantic relations in Princeton WordNet. By mapping these environment-related English words to Chinese words with any kind of semantic rela­tions in the sense axes, the corresponding Chinese domain senses can be identified.

Therefore, the HR-WSD system will first con­sider any domain senses for the words to be dis­ambiguated; if there is no such sense, the proto­type sense will be adopted. Another test where sense definitions from WordNet are considered to facilitate HR-based disambiguation was also con­ducted.

The results were evaluated according to three man­ually tagged documents in SemEval-2010 Task 17: All-words WSD on a Specific domain (Agirre et al., 2010). The most frequent sense baseline (MFS) refers to the first sense in WordNet lexical markup framework (In Chinese WordNet senses are ordered according to annotations in hand-labelled corpora). In these tagged domain texts, only nouns and verbs (two major types of content words) as a single word are disambiguated. There­fore, in this system only these two kinds of words will be tagged with senses. Adjectives, adverbs, or words in multiple forms (e.g., idioms and phrases) are not considered, in order to simplify the test and observe the results more clearly.

By observing that the HR-WSD system* (Rank 1) outperformed other systems and was closest to the MFS performance we can infer that Heuristic Rule 2 works. However, since this system perfor­mance is still worse than MFS, it may indicate that Heuristic Rule 1 does not work well, or even de­creases the system performance, so the mechanism to detect domain senses needs to be refined. Be­sides, the inclusion of simplified Lesk algorithm** did not perform better than the original HR-WSD system, further investigation such as more fine­grained definition can be expected.

Although PoS tagging may help filter out senses from other parts of speech of the same word form, incorrect PoS tagging will lead to incorrect sense tagging, which did happen in the HR-WSD sys­tem, in particular when there is more than one pos­sible PoS tag for the word. For instance, 'nuan-hua' in 'quan-qiu nuan-hua' (global warming) is manually tagged with a verbal sense in the answer key from SemEval-2010, but tagged as a noun in the pre-processing stage of the HR-WSD sys­tem. The difference between manual tagged texts and automatic tagged texts should be examined, or consider allowing more than one PoS tag for a word, or even no PoS pre-processing at all.

To disambiguate with the help of gloss defini­tion, gloss words of the polysemous word must have direct overlapping with that of its context word, which does not always occur. To solve this problem, we may expand gloss words to related words such as hyponyms, hypernyms, meronyms, or the gloss definition of the current gloss words.

Apart from nouns and verbs, if function words and other kinds of content words such as adjec­tives and adverbs are to be disambiguated, the per­formance of the current WSD system needs to be re-examined.

As mentioned in the beginning, WSD is an es­sential part in language understanding systems.