Sensiting Inflectionality: Estonian Task for SENSEVAL-2

Sensiting inflectionality: Estonian task for SENSEVAL-2 Neeme Kahusk and Heili Orav and Baldur Oim University of Tartu Research Group of Computational Linguistics Tiigi 78, 50410 Tartu, Estonia { nkahusk,horav,hoim }@psych.ut.ee Abstract This paper describes the all-word sense disam- biguation task provided by Estonian team at SENSEVAL-2. About 10,000 words are manually disambiguated according to Estonian WordN et word senses. Language-specific problems and lexicon features are discussed. 1 Introduction We got interested in word sense disambiguation (WSD) for two reasons. First, already a couple of years ago it was evident that WSD is beco- ming one of the new "hot" topics in computatio- nal linguistics and language engineering as our knowledge of how to handle semantic parame- ters of texts and semantic features of words in texts increased. The second reason was purely practical. Since 1996 we have been involved in a large project of building a semantic database of Estonian; participating in the Euro WordNet project has been a part of it (but a very impor- tant part, of course). The main source of buil- ding this database have been different corpora of Estonian, and in working with corpora the question of whether we are dealing with diffe- rent meanings of a word in case of its concrete occurrences or not arises constantly. So we got interested in the possibility to use some objec- tive methods here. Our task was all-words task. This choice is explained with our "practical" interests explai- ned above. A large amount of work was done to provide training data where disambiguation was done manually. The same kind of work had to be done with test data, of course. The description of this work is given below. Let us note already here that this work appeared to be very useful and informative for us as builders of Estonian WordNet (EstWN). And let us stress that this was our first at- tempt of WSD at all. 2 Corpora and lexicon The test and training texts come from Corpus of the Estonian Literary Language (CELL), the 1980-s. We used this part of the corpus, that was morphologically disambiguated, initially for the syntactic analysis. The morphological analysis was made with ESTMORF (Kaalep, 1997). Lemma and word class in the output of the program are relevant to our task, but it is impossible to get them without morphological disambiguation, because of frequent homonymy among word forms. All training texts and most of test texts (5 of 6 total) are fiction. One of the test texts is from newspaper. Six training and six test files provided for the task contain about 2000 tokens each. More information about the texts used in the task is in Table 1. Table 1: Statistics on training and test corpora Corpus Training Test Total words 12162 11440 Words to disambiguate 5854 5650 of them being verbs 2431 2191 nouns 3423 3459 2.1 Lexicon The Estonian part of EuroWordNet1 served as the lexicon. Like other wordnets, Est WN is a lexical-semantic database, the basic unit of which …

Content truncated to 500 words