A Wordnet from the Ground Up

94 Chapter 3. Discovering Semantic Relatednessinitial LU clusters generated by processing the lists of the k most similar LUs, seeII.1 and II.2. Only the groups dissimilar to other selected groups are added to the setof committees, because the committees should ideally describe all senses of the inputLUs, see II.3. The set of committees is also iteratively extended in order to coversenses of all input LUs, see the condition in II.4.Committees only define senses. They are not the final lemma groups we willextract. The final lemma groups – ideally sets of near-synonyms – are extracted inPhase III on the basis of committees. Each lemma can be assigned to one of severalgroups by the similarity to the corresponding committees. It is assumed that eachsense of a polysemous lemma corresponds to some subset of features which describethe given LU assigned to some committee c (the next sense of e has been identified).CBC attempts to identify the features that describe sense c of e and remove them beforethe extraction of the other senses of e. The idea behind this operation is to removesense c from the representation of e, in order to make other senses more prominent.The original implementation of the overlap and remove operations is straightforward:the values of all features in the intersection are simply set to 0 (Pantel, 2003).We found this technique too radical. It would be correct if the association of featuresand senses were strict, but it is very rarely the case. Mostly, one feature derived fromlexico-syntactic dependency corresponds in different degree to several senses.After a manual inspection of data collected in a co-incidence matrix, we concludedthat it is hard to expect any group of features to encode some sense unambiguously.Some features also have low, accidental values, while some are very high. Finally,vector similarity is influenced by the whole vector, especially when we analyse theabsolute values of similarity by comparing it to a threshold such as. σ in step 2b ofthe CBC algorithm.Assuming that a group of features and some part of their strength are associatedwith a sense just recorded, we wanted to look for an estimation of the extent towhich feature values should be reduced. The best option seems to be the extraction ofsome association of features with senses, but for that we need an independent source ofknowledge for grouping features, as it was done in (Tomuro et al., 2007). Unfortunately,it is not possible in the case of a language with limited resources like Polish. Instead,we tested two simple heuristics (w(f i ) is the value of the f i feature, v c (f i ) – the valueof f i in the committee centroid 25 ):• minimal value:w(f i ) = w(f i ) − min(w(f i ), v c (f i ))25 The centroid features are calculated as average from the features of vectors in the committee.