Using automatically acquired predominant senses for Word Sense Disambiguation

Using Automatically Acquired Predominant Senses for Word Sense Disambiguation Diana McCarthy & Rob Koeling & Julie Weeds & John Carroll Department of Informatics, University of Sussex Brighton BN1 9QH, UK � dianam,robk,juliewe,johnca � @sussex.ac.uk Abstract In word sense disambiguation (WSD), the heuristic of choosing the most common sense is extremely powerful because the distribution of the senses of a word is often skewed. The first (or predominant) sense heuristic assumes the availability of hand- tagged data. Whilst there are hand-tagged corpora available for some languages, these are relatively small in size and many word forms either do not occur, or occur infrequently. In this paper we in- vestigate the performance of an unsupervised first sense heuristic where predominant senses are ac- quired automatically from raw text. We evaluate on both the SENSEVAL-2 and SENSEVAL-3 English all- words data. For accurate WSD the first sense heuris- tic should be used only as a back-off, where the evi- dence from the context is not strong enough. In this paper however, we examine the performance of the automatically acquired first sense in isolation since it turned out that the first sense taken from SemCor outperformed many systems in SENSEVAL-2. 1 Introduction The first sense heuristic which is often used as a baseline for supervised WSD systems outperforms many of these systems which take surrounding con- text into account (McCarthy et al., 2004). The high performance of the first sense baseline is due to the skewed frequency distribution of word senses. Even systems which show superior performance to this heuristic often make use of the heuristic where evi- dence from the context is not sufficient (Hoste et al., 2001). The first sense heuristic is a powerful one. Us- ing the first sense listed in SemCor on the SENSE- VAL-2 English all-words data we obtained the re- sults given in table 1, (where the PoS was given by the gold-standard data in the SENSEVAL-2 data it- self). 1 Recall is lower than precision because there are many words which do not occur in SemCor. Use 1We did not include items which were tagged ’U’ (unassignable) by the human annotators. PoS precision recall baseline Noun 70 60 45 Verb 48 44 22 Adjective 71 59 44 Adverb 83 79 59 All PoS 67 59 41 Table 1: The SemCor first sense on the SENSEVAL- 2 English all-words data of the first sense listed in WordNet gives 65% pre- cision and recall for all PoS on these items. The fourth column on table 1 gives the random base- line which reflects the polysemy of the data. Ta- ble 2 shows results obtained when we use the most common sense for an item and PoS using the fre- quency in the SENSEVAL-2 English all-words data itself. Recall is lower than precision since we only use the heuristic on lemmas which have occurred more than once and where there is one sense which has a greater frequency than the others, apart from trivial monosemous cases. 2 Precision is higher …

Content truncated to 500 words