Finding optimal parameter settings for high performance word sense disambiguation

Finding optimal parameter settings for high performance word sense disambiguation Cristian Grozea Department of Computer Science University of Bucharest Str. Academiei 14, 70109 Bucharest, Romania chrisg@phobos.ro Abstract This article describes the four systems sent by the author to the SENSEVAL-3 contest, the English lexical sample task. The best recognition rate ob- tained by one of these systems was 72.9% (fine grain score) . 1 Introduction. RLSC algorithm, input and output. This paper is not self-contained. The reader should read first the paper of Marius Popescu (Popescu, 2004), paper that contains the full description the base algorithm, Regularized Least Square Classifi- cation (RLSC) applied to WSD. Our systems used the feature extraction described in (Popescu, 2004), with some differences. Let us fix a word that is on the list of words we must be able to disambiguate. Let � be the number of possible senses of this word . Each instance of the WSD problem for this fixed word is represented as an array of binary values (features), divided by its Euclidian norm. The num- ber of input features is different from one word to another. The desired output for that array is another binary array, having the length � . After the feature extraction, the WSD problem is regarded as a linear regression problem. The equa- tion of the regression is ���� where each of the lines of the matrix � is an example and each line of � is an array of length � containing �� zeros and a single . The output � of the trained model � on some particular input is an array of values that ide- ally are just 0 or 1. Actually those values are never exactly 0 and 1, so we are prepared to consider them as an ”activation” of the sense recognizers and con- sider that the most ”activated” (the sense with high- est value) wins and gives the sense we decide on. In other words, we consider the � values an approxi- mation of the true probabilities associated with each sense. The RLSC solution to this linear regression prob- lem is �������  "! � ; The first difference between our system and Mar- ius Popescu’s RLSC-LIN is that two of the systems (HTSA3 and HTSA4) use supplementary features, obtained by multiplying up to three of the exist- ing features, because they improved the accuracy on Senseval-2. Another difference is that the targets � have values 0 and 1, while in the Marius Popescu’s RLSC-LIN the targets have values -1 and 1. We see the output values of the trained model as approximations of the true probabilities of the senses. The main difference is the postprocessing we ap- ply after obtaining � . It is explained below. 2 Adding parameters The obviously single parameter of the RLSC is Some improvement can be obtained using larger . values. After dropping the parser information from features (when it became clear that we won’t have those for Senseval-3) the improvements proved to be too …

Content truncated to 500 words