Tagging polysemous word with its most frequent sense (MFS) is a popular back-off heuristic in word sense disambiguation (WSD) systems when the training data is inadequate. In past evalua­tions, MFS from WordNet performed even bet­ter than most of the unsupervised systems (Snyder and Palmer, 2004; Navigli et al., 2007).

MFS is usually obtained from a large scale sense tagged corpus, such as SemCor (Miller et al., 1994). However, some polysemous words have different MFS in different domains. For example, in the Koeling et al. (2005) corpus, target word coach means "manager" mostly in the Sports domain but means "bus" mostly in the FINANCE domain. So when the MFS is applied to specific domains, it needs to be re-estimated.

McCarthy et al. (2007) proposed an unsuper­vised predominant word sense acquisition method which obtains domain specific MFS without sense tagged corpus. In their method, a thesaurus, in which words are connected with their distribu­tional similarity, is constructed from the domain raw text. Word senses are ranked by their preva­lence score which is calculated using the thesaurus and the sense inventory.

In this paper, we propose another way to con­struct the thesaurus. We use statistical machine

POS tagging

Figure 1 : The architecture of HIT-CIR translation (SMT) techniques to extract paraphrase pairs from bilingual parallel text. In this way, we avoid calculating similarities between every pair of words and could find semantic similar words or compounds which have dissimilar distributions.

Our system is comprised of two parts: the word sense ranking part and the word sense tagging part. Senses are ranked according to their prevalence score in the target domain, and the predominant sense is used to tag the occurrences of the target word in the test data. The architecture of this sys­tem is shown in Figure 1.

The word sense ranking part includes following steps.

1. Tag the POS of the background text, count the word frequency in each POS, and get the polysemous word list of the POS.

In our method, word senses are ranked according to their prevalence score in the specific domain. According to the assumption of McCarthy et al. (2007), the prevalence score is affected by the fol­lowing two factors: (1) The reiatedness score be­tween a given sense of the target word and the target word's neighbor word. (2) The similarity between the target word and its neighbor word. In addition, we add another factor, (3) the impor­tance of the neighbor word in the specific domain.

In this paper, "neighbor words" means the words which are most semantically similar to the target word.

Figure 2 illustrates the word sense ranking pro­cess of noun backbone. The contribution of a neighbor word to a given word sense is measured by the similarity between them and weighted by the importance of the neighbor word in the tar­get domain and the reiatedness between the neigh­bor word and the target word. Sum up the con­tributions of each neighbor words, and we get the prevalence score of the word sense.

Formally, the prevalence score of sense Si of a target word w is assigned as follows:

sx Esenses

rs(w, rij) is the reiatedness score between w and a neighbor word rij. Nw = {rii,ri2, ■ ■ ■ ,rik} is the top k reiatedness score neighbor word set. ns(si, rij) is the normalized form of the sense sim­ilarity score between sense Si and the neighbor word rij (i.e. sss(si,rij)). We define this score with the maximum WordNet similarity score be­tween Si and the senses of rij (i.e. sss'(si,rij)). In our system, lesk algorithm is used to measure the sense similarity score between word senses.

The neighbor words of the target word as well as the reiatedness score are obtained by extracting paraphrases from bilingual parallel texts. When a word is translated from source language to tar­get language and then translated back to the source language, the final translation may have the same meaning to the original word but with different ex­pressions (e.g. different word or compound). The translation in the same language could be viewed as a paraphrase term or, at least, related term of the original word.

For example, in Figure 3, English noun back­bone can be translated to columna, columna verte­bral, pilar and convicciones etc. in Spanish, and these words also have other relevant translations in English, such as vertebral column, column, pil­lar and convictions etc., which are semantically re­lated to the target word backbone.

We use a statistical machine translation sys­tem to calculate the translation probability from English to another language (called as pivot lan­guage) as well as the translation probability from that language to English. By multiplying these two probabilities, we get a paraphrase probabil­ity. This method was defined in (Bannard and Callison-Burch, 2005).

In our system, we choose the top k paraphrases as the neighbor words of the target word, which have the highest paraphrase probability. Note that there are two directions of the paraphrase, from target word to its neighbor word and from the neighbor word to the target word. We choose the paraphrase score of the former direction as the reiatedness score (rs). Because the higher of the score in this direction, the target word is more likely paraphrased to that neighbor word, and hence the prevalence of the relevant target word sense will be higher than other senses. For­mally, the reiatedness score is given by where / is the pivot language word.

We use the English-Spanish parallel text from Europarl (Koehn, 2005). We choose Spanish as the pivot language because in the both directions the BLEU score of the translation between English and Spanish is relatively higher than other English and other languages (Koehn, 2005).

The organizers of the SemEval-2 specific domain WSD task provide no training data but raw back­ground data in the environmental domain. The En­glish background data is obtained from the offi­cial web site of World Wide Fund (WWF), Euro­pean Centre for Nature Conservation (ECNC), Eu­ropean Commission and the United Nations Eco­nomic Commission for Europe (UNECE). The size of the raw text is around 15.5MB after sim­ple text cleaning. The test data is from WWF and ECNC, and contains 1398 occurrence of 436 tar­get words.

For the implementation, we used bpos (Shen et al., 2007) for the POS tagging. The maximum number of the neighbor word of each target word k was set to 50. We employed Giza++ and Mosesto get the phrase table from the bilingual paral­lel corpus. The WordNet: :Similarity package was applied for the implement of the lesk word sense similarity algorithm.

For the target word that is not in the polysemous word list, we use the MFS from WordNet as the back-off method.

The recall of our system is 43.5%, which is lower than that of the MFS baseline, 50.5% (Agirre et al., 2010). The baseline uses the most frequent sense from the SemCor corpus (i.e. the MFS of WordNet). This means that for some target words, the MFS from SemCor is better than the domain MFS we estimated in the environmental domain. In the future, we will analysis errors in detail to find the effects of the domain on the MFS.

For the domain specific task, it is better to use parallel text in the domain of the test data in our method. However, we didn't find any available parallel text in the environmental domain yet. In the future, we will try some parallel corpus acqui­sition techniques to obtain relevant corpus for en­vironmental domain for our method.

This work was supported by National Natural Science Foundation of China (NSFC) via grant 60803093, 60975055, the "863" National High-Tech Research and Development of China via grant 2008AA01Z144, and Natural Scientific Re­search Innovation Foundation in Harbin Institute of Technology (HIT.NSRIF.2009069).