Effective parameters for word sense disambigua­tion (WSD) may vary for different languages and word types. Although, some parameters are com­mon in many languages, some others may be lan­guage specific. Turkish is an interesting language that deserves being examined semantically. Turk­ish is based upon suffixation, which differentiates it sharply from the majority of European languages, and many others. Like all Turkic languages, Turk­ish is agglutinative, that is, grammatical functions are indicated by adding various suffixes to stems. Turkish has a SOV (Subject-Object-Verb) sentence structure but other orders are possible under certain discourse situations. As a SOV language where objects precede the verb, Turkish has postpositions rather than prepositions, and relative clauses that precede the verb. Turkish, as a widely-spoken lan­guage, is appropriate for semantic researches.

TLST utilizes some resources that are explained in Section 2-5. In Section 6 evaluation of the sys­tem is provided. In section 7 some concluding re­marks and future work are discussed.

Lesser studied languages, such as Turkish suffer from the lack of wide coverage electronic re­sources or other language processing tools like on­tologies, dictionaries, morphological analyzers, parsers etc. There are some projects for providing data for NLP applications in Turkish like METU Corpus Project (Oflazer et al., 2003). It has two parts, the main corpus and the treebank that con­sists of parsed, morphologically analyzed and dis-ambiguated sentences selected from the main cor­pus, respectively. The sentences are given in XML format and provide many syntactic features that can be helpful for WSD. This corpus and treebank can be used for academic purposes by contract.

The texts in main corpus have been taken from different types of Turkish written texts published in 1990 and afterwards. It has about two million words. It includes 999 written texts taken from 201 books, 87 papers and news from 3 different Turk­ish daily newspapers. XML and Text Encoding Initiative (TEI) style annotation have been used. The distribution of the texts in the Treebank is similar to the main corpus. There are 6930 sen­tences in this Treebank. These sentences have been parsed, morphologically analyzed and disambigu-ated. In Turkish, a word can have more than one analysis, so having disambiguated texts is very important.

The dictionary is the one that is published by TDK (Turkish Language Foundation) and it is open to public via internet. This dictionary lists the senses along with their definitions and example sentences that are provided for some senses. The dictionary is used only for sense tagging and enumeration of the senses for standardization. No specific information other than the sense numbers is taken from the dictionary; therefore there is no need for linguistic processing of the dictionary.

In Table 1 statistical information about the final training and testing sets of TLST is summarized. The data have been provided for 3 words in the trial set and 26 words in the final training and test­ing sets (10 nouns, 10 verbs and 6 other POS for the rest of POS including adjectives and adverbs). It has been tagged about 100 examples per word, but the number of samples is incremented or dec­remented depending on the number of senses that specific word has. For a few words, however, fewer examples exist due to the sparse distribution of the data. Some ambiguous words had fewer ex­amples in the corpus, therefore they were either eliminated or some other examples drawn from external resources were added in the same format. On the average, the selected words have 6.7 senses, verbs, however, have more. Approximately 70% of the examples for each word were delivered as training data, whereas approximately 30% was reserved as evaluation data. The distribution of the senses in training and evaluation data has been kept proportional. The sets are given as plain text files for each word under each POS. The samples for the words that can belong to more than one POS are listed under the majority class. POS is provided for each sample.

We have extracted example sentences of the tar­get word(s) and some features from the XML files. Then tab delimited text files including structural and sense tag information are obtained. In these files each line has contextual information that are thought to be effective (Orhan and Altan, 2006; Orhan and Altan, 2005) in Turkish WSD about the target words. In the upper level for each of them XML file id, sentence number and the order of the ambiguous word are kept as a unique key for that specific target. In the sentence level, three catego­ries of information, namely the features related to the previous words, target word itself and the sub­sequent words in the context are provided.

A small scale ontology for the target words and their context was constructed. The Turkish Word­Net developed at Sabanci University is somehow insufficient. Only the verbs have some levels of relations similar to English WordNet. The nouns, adjectives, adverbs and other words that are fre­quently used in Turkish and in the context of the ambiguous words were not included. This is not a suitable resource for fulfilling the requirements of TLST and an ontology specific to this task was required. The ontology covers the examples that are selected and has three levels of relations that are supposed to be effective in the disambiguation process. We tried to be consistent with the Word­Net tags; additionally we constructed the ontology not only for nouns and verbs but for all the words that are in the context of the ambiguous words se­lected. Additionally we tried to strengthen the rela­tion among the context words by using the same tags for all POS in the ontology. This is somehow deviating from WordNet methodology, since each word category has its own set of classification in it.

WSD is a new area of research in Turkish. The sense tagged data provided in TLST are the first resources for this specific domain in Turkish. Due to the limited and brand new resources available and the time restrictions the participation was less. We submitted a very simple system that utilizes statistical information. It is similar to the Naïve Bayes approach. The features in the training data was used individually and the probababilities of the senses are calculated. Then in the test phase the probabilities of each sense is calculated with the given features and the three highest-scored senses are selected as the answer. The average precision and recall values for each word category are given in Table 3. The values are not so high, as it can be expected. The size of the training data is limited, but the size is the highest possible under these cir­cumstances, but it should be incremented in the near future. The number of senses is high and pro­viding enough instances is difficult. The data and the methodology for WSD will be improved by the experience obtained in SemEval evaluation exer­cise.

The evaluation is done only for FG and CG senses. For FG senses no partial points are as­signed and 1 point is assigned for a correct match. On the other hand, the CG senses are evaluated partially. If the answer tags are matching with any of the answer tags they are given points.

In TLST we have prepared the first resources for WSD researches in Turkish. Therefore it has sig­nificance in Turkish WSD studies. Although the resources and methodology have some deficien­cies, a valuable effort was invested during the de­velopment of them. The resources and the method­ology for Turkish WSD will be improved by the experience obtained in SemEval and will be open to public in the very near future from http://www.fatih.edu.tr/~zorhan/senseval/senseval.htm.