The Estonian part of Euro WordNet served as the lexicon. Like other wordnets, EstWN is a lexical-semantic database, the basic unit of which is concept. Concepts are represented as synonym sets (synsets) that are linked to each other by semantic relations. The description of

EstWN is given in the final document of Eu­roWordNet (Vider et al., 1999).

EstWN is supposed to cover the Estonian base vocabulary in its initial version. The base vocabulary will be determined by statistical analysis of the reference corpus. Even so it is not always easy (nor appropriate) to stop enco­ding words with frequencies below a certain th­reshold. For this reason we expect EstWN to cover more than just the base vocabulary.

Still the EstWN is rather small, there were 9436 synsets, 13277 words and 16961 senses (li­terals) in it when the disambiguation was done. That makes about 1.28 senses per word as ave­rage.

Most of synsets are connected with hypero-nym-hyponym relations building corresponding hierarchies.

Four linguists disambiguated the texts, each text was disambiguated by two persons. Only nouns and verbs were disambiguated, as ente­ring adjectives into EstWN is in the very begin­ning. The sense number was marked according to sense number in EstWN. If the word was mis­sing from the EstWN, "0" was marked as sense number, and if the word was in EstWN, but missed the appropriate sense, "+1" was mar­ked.

If inconsistencies were met, they were discus­sed until agreement was achieved. On about 28% of the cases the disambiguators had diffe­rent opinions.

One of the problems that the disambiguators ran into concerned dividing words into diffe­rent senses in EstWN. It turned out as over-differentiation—word meaning marked as too specific, or over-generalisation—word meaning marked as too general.

Not all senses found in EstWN are represented in texts. Maximum number of senses per word found in texts is 13. This is more than app­ropriate senses in lexicon (see Table 3), but we must remeber about the "+1" that disambigua­tors had, if they found that there are not enough meanings in EstWN. Table 2 describes distri­bution of senses in usage and Table 3 shows the top of lemmas according to number of senses.

It would be the best, if all words to disambi­guate were in the lexicon with all their possibk meanings. Apparently this presumption is not met.

The number of compounds in Estonian is in­definite. It is quite easy for a writer to in­vent new compounds that are not in any dic­tionary, but nevertheless are easily understood by readers. That is one reason, why there are so many sense numbers "0" in the texts. About 46 % of words that are not in EstWN, are com­pounds.

Another remarkable class of words not in lexicon are proper names, as there are no pro­per names in EstWN. There are 17.5 % of words proper names.

If we will postpone phrasal verbs and some strange words that contain hyphens (about 7 %), it leaves us with about half thousand words to check why they are not in EstWN.

But why are there missing senses (tagged with "+1")? The reason is simply historical: such words were included into EstWN as sy­nonyms of some base vocabulary word and the other senses of them are not considered yet.

The initial format of text was as it came from ESTMORF and semantic disambiguation: every word on separate line, followed by an adi-tional line of morphological analysis and sense number, with multi-word phrase marked if word was part of it. The task to convert into Senseval XML format seemed trivial at first, but phrases turned out to be problematic. Unfortunately enough, all the story about phrases is concer­ning the training corpus only, because in test corpus the multi-word phrases were unmarked.

Estonian is a flective language with a free word order and that makes it complicated to figure out all phrases. The elements of a ph­rase can be scattered around the sentence in an unpredictable order.

In the initial texts, the disambiguators mar­ked down the whole phrase on the line where the phrase occured. They were not told to mark it on each line, where the non-disambiguatable parts of the phrase were, and it happened that the phrase was not marked on the line, where the head of the phrase was. The algorithm of calculating head or satellite took into account the part of speech and the form. For verb phra­ses, if both components were verbs, declinable form of verb infinitive was marked as satellite. For noun phrases, substantive makes head and adjective satellite. If both words are substanti­ves, head is the second one... well, mostly.

However, it is known that expressions tend to contain frozen forms, including inflectional en­dings. For example, one may not say "*Human Right" or "*Humans Right". "Human Rights" is the only correct expression and should be ad­ded into thesauri in such form. Phrasal verbs like "ara maksma" (to pay off) and idiomatic verbal expressions like "end tiikkideks naerma" (to laugh oneself into pieces) represent a situa­tion that is different from the occasion desc­ribed above: the verb part may inflect freely, but the other word(s) are frozen forms. He­reby, even if we have determined what is phrase or collocational multi-word unit, we still have a question— are they commonly used and should we add them into the lexicon.

Multiword expressions are included into EstWN if they build up a conceptual unit and are commonly used as lexical units.