Memory-Based Word Sense Disambiguation for Senseval
Walter Daelemans, Antal van den Bosch, Sabine Buchholz, Jorn Veenstra,
Jakub Zavrel
ILK / Computational Linguistics, Tilburg University
PO-box 90153, 5000 LE Tilburg, the Netherlands
walter,antalb,buchholz,veenstra,zavrel@kub.nl
In the context of the Senseval project, we developed a Memory-Based
Learning (MBL) architecture for word sense disambiguation (WSD).
Senseval is set up as a competitive evaluation for WSD systems, and
provides a fixed set of sense distinctions, tagged corpus data, and a
gold standard for measuring system performance. In this year's
Senseval workshop, the WSD task was limited to a relatively small set
of ambiguous words.
Our system consists of separate "word-experts" for each ambiguous
word. In line with our general approach to NLP, the experts are
trained on annotated corpus examples, i.e. sense-tagged words in
context. As far as definitions and usage patterns from a dictionary
are available, we add these to the training material as a further type
of examples. The examples are tagged with parts-of-speech and
processed by a module which extracts a large amount of features from
each example. The three main information sources are neighboring
words, neighboring parts-of-speech, and keywords. The keywords are
words within a wider context which are informative for a particular
sense. The extrapolation from the training examples in MBL is based on
a weighted similarity metric. Because of this, we can conveniently
integrate diverse information sources and large numbers of features
without harmful independence assumptions and without the need to
disregard infrequent or atypical patterns.
The training data for each target word has different characteristics
in terms of numbers of senses and the amount of training examples. To
obtain the best possible performance on the test set, the parameters
for keyword extraction and for the learning algorithm were tuned
separately for each word-expert using tenfold crossvalidation on the
training data. The performance of these word experts varied between
35.4% and 100%, depending e.g. on the data sparseness of that word
and the evaluation criteria. The overall generalisation accuracy was
77%. In the Senseval competition, the MBL system described in this
paper was third best from around twenty participants.