Combination of Contextual Features for Word Sense Disambiguation: LIU-WSD

Combination of contextual features for word sense disambiguation: LIU-WSD Magnus MERKEL & Mikael ANDERSSON Dept. of Computer and Information Science, l.inkoping University S581 83 Linkoping, Sweden, magme@ ida.liu.se, miand @ida.liu.se Abstract This paper describes a system for word sense disambiguation that participated in the Swedish Lexical Sample task of SENSEVAL-2. The system LIU-WSD is based on letting different contextual features cast votes on preferred senses according to a ranking scheme. Introduction The addition of new languages to the SENSEV AL-2 workshop, among these languages also Swedish, presented an opportunity to learn more about WSD applied to Swedish by participation in the event. Previously, we had had no experience of building word sense disambiguation software, but the Swedish Lexical Sample task seemed like a suitable occasion for trying another field of NLP (in recent years our focus has been on word alignment and parallel corpora). Due to time constraints our initial plans of implementing some kind of version of decision lists (Yarowsky, 2000; Pedersen 2001) were abandoned in the end and we decided to go for a slightly simpler approach based on a general algorithm and voting strategies for contextual features on different levels. The contextual features that were being considered were unigrams and bigrams, both in fixed and variable positions, together with possibilities to include parts-of-speech, lemmas and graph words (inflected words). 1 Data and pre-processing The data and resources used in the LIU-WSD system were apart the following: • Sample and training data, provided by the task organisers - the sample data were 123 a great help in order to understand the format of the provided material. • Part of the lexicon data provided for the Swedish lexical sample task. Here only the information on the number of senses and division of main and sub senses was used. Information contained in examples, definitions and for valency was left out. As the system was a first attempt to word sense disambiguation for us, we set up a small corpus containing for five lexical items with around 60 contexts for each lexeme. This was used to test various approaches and algorithms as well as making sure of conversions to and from the format used in the task. As we wanted to use morphosyntactic information and lemmas in the system, the Swedish Constraint Grammar package, SWECG-2 from Conexor was used (Karlsson et al., 1994). The training corpus was fed into the tagger, SWECG-2, which returned an XML file where the text was POS tagged and lemmatised. Below is a sample of how the sentence Men pastaendet ar litet missvisande. came back in XML format. <instance id="barn.l9"> <answer instance="barn.19"/> <context> <w id="wl" base="rnen" pos="CC">rnen</w> <w id="w2" base="pastaende" pos="N">pastaendet</w> <w id="w3" base="vara" pos="PRES">ar</w> <w id="w4" base="litet & litet" pos="ADV & A">litet</w> <w id="w5" base="rnissvisa" pos="NDE">rnissvisande</w> <w id="w6" base="." pos="interpunction">.</w> ... 2 Training The actual training was a matter of building tables of information from the training corpus. The task was first to decide on a number of contextual features to be observed, then set …

Content truncated to 500 words