The Språkdata-ML System as Used for SENSEVAL-2

The Sprakdata-ML System as Used for SENSEV AL-2 Dimitrios KOKKINAKIS Spd'tkdata, Goteborg University Box 200, SE-405 30 Goteborg, ·Sweden Dimitrios.Kokkinakis @svenska.gu.se Abstract This paper describes the Sprakdata-ML system as used in the SENSEV AL-2 exercise. The main focus of the paper is devoted to the process of feature extraction, preparation and organization of the test and training data. Introduction The methodology followed for sense disambiguation of the Swedish data by the Sprakdata-ML system is supervised, based on Machine Learning (ML) techniques, particularly Memory Based Learning (MBL). The MBL implementation we used originates from the university of Tilburg in a system called TiMBL; details can be found in Daelemans et al. ( 1999). Thus, our main contribution in this task has been the effort to try and isolate a set of features that could maximize the performance of the MBL software. However, it is rather difficult to give the exact number of features and examples required for an adequate description of a word's sense or which algorithm performs best. We think that there is space for improvement of our system's performance by better modeling of the available resources (e.g. context, annotations), choice of parameters and algorithms, a claim that we have not explored to its full potential, further exploration is required. Intelligent example selection for supervised learning is an important issue in ML, an issue that we have not fully explored. In previous experiments for a similar problem for Swedish, the algorithm that performed best in TiMBL was a variant of the k- nearest neighbor (Mitchell, 1997) called IB 1, an algorithm that we also used in the exercise; (Kokkinakis & Johansson Kokkinakis, 1999). 91 1 Data Preparation (Train) To enhance the lexical disambiguation results using the available resources, we perform pre- processing in both the dictionary and the text to be sense-disambiguated. This is motivated by the fact that by making certain normalizations and simplifications in the resources we (hopefully) contribute to the production of qualitatively better results. Initially, a text to be disambiguated is pre- processed by a tokeniser, a sentence boundary identifier, an idiom1 and multiword identifier, a Name-Entity recogniser2, a part-of-speech tagger, a lemmatiser and a semantic tagger3. Then, the input texts are transformed to the specified format that the MBL requires, which is feature-vectors of a specific length and content. The vectors we use consist of 102 features, the last two being the id-number and class or sense assigned to the vector. Since we do not know in advance which features will be useful for each particular word and sense, we chose to include features from a number of different information sources. 2 Vector Creation The vectors consisted of: (i) selected information gathered from the dictionary entries (5 features); (ii) near-context (5 features); (iii) annotations applied on the training corpus (5 1 The idioms originate from the Gothenburg Lexical Data Base/Semantic Database (GLDB/SDB) (http://spraakdata.gu.se/lb/gldb.html) and were used for the recognition and marking of idioms in the test/training corpus (over 4,000 idioms). 2 See http://spraakdata.gu.se/svedk/ne.html for a …

Content truncated to 500 words