CHAPTER 7. CONCLUSION, FUTURE AND RELATED WORK 75 The last related to our work remark from the above observed by Gonzalo and Verdejo (2007) approaches for automatic acquisition of examples for WSD is the fact that the performance of unsupervised strategies is far poorer than the performance of supervised ones. Similar to our unsupervised experiment, the authors mention, that in most of the cases the approaches of this family do not even reach the MFS baseline.
CHAPTER 7. CONCLUSION, FUTURE AND RELATED WORK 76 7.2 Future Work An interesting question in our work, which we assume can lead to very good performance is the choice of words for which examples are extracted. Consider again our experiment based on the addition of the closest instances from the automatically prepared training set to the manually annotated one. As we reported, for a big subset of the words there is accuracy variance of the separate word-experts of about 4%. To figure out the tendency with which the results rise or fall would be a valuable information that can be used in order to add selectively the closest instances. Another issue worth further investigation is what exactly is the effect of the proportional distribution of senses on the final results. It will be interesting to see if the performance of the supervised systems would stay the same if the training and test sets provided by Senseval-3 did not keep such close proportions of the senses. Parameter optimization is also an extremely important part of a WSD system, which can lead to a considerable increase of its performance and respectively of the performance of the separate word-experts. In our work we managed to show only a very limited parameter optimization concentrated only on three different values for a single parameter. The exceptionally good results of the attempt prove that further investigation of the parameter optimization is essential and will most certainly lead to improvement of the performance of the WSD system. 7.3 Conclusion This thesis describes the design and performance of a simplistic memory-based word sense dis- ambiguation system, which makes use of automatic feature selection and minimal parameter optimization. We show that the system performs competitive to other state-of-art systems and use it further for evaluation of automatically acquired data for word sense disambiguation. In our work we conducted several experiments employing the automatically extracted data as training instances. The results of our approach prove that automatic extraction of examples for word sense disambiguation can help to a great extend for the improvement of manually anno- tated training data. We showed that by selecting only the most similar automatically annotated instances and adding them to the manually annotated ones the performance of the separate word-experts can rise up to 4%. Nowadays, the increasing number and size of freely available corpora extremely accomodates the arising need for resources, which are required for approaches similar to ours. Once provided such corpora, however, we demonstrated that the potential of those approaches for improvement of the performance of word sense disambiguation systems by automatically extracted data is immense and determines their further exploration as more than worthwile.