Results

As was described in [Pinto, Jiménez-Salazar, and Rosso2006], the best results are obtained by the transition point technique together with the use of the KStar clustering method (for both, CICLing and hep-ex corpora). We show in Tables 5 and 6, the maximum values obtained for each feature selection technique by using five different clustering methods, for the CICLing and hep-ex corpus, respectively. As may be seen, the transition point technique obtains always the best results, also for the other four clustering methods. This fact suggests that this feature selection technique could be clustering method independent.

Table: Maximum F-Measure obtained using the CICLing corpus

	TP	DF	TS
Kstar	0,72	0,64	0,62
SLC	0,60	0,56	0,52
CLC	0,74	0,68	0,65
NN1	0,73	0,66	0,68
KNN	0,69	0,60	0,60

Table: Maximum F-Measure obtained using the hep-ex corpus

	TP	DF	TS
Kstar	0,6878	0,6844	0,6714
SLC	0,7651	0,5921	0,7367
CLC	0,8676	0,8631	0,8618
NN1	0,6052	0,5439	0,5455
KNN	0,2223	0,2220	0,2206

The behaviour of each feature selection technique (PT, DF, and TS) upon the use of the hep-ex corpus and by using the five clustering methods are shown in Figures 1, 2, and 3.

Figure: Behaviour of PT with each clustering method using hep-ex corpus

Figure: Behaviour of DF with each clustering method using hep-ex corpus

Figure: Behaviour of TS with each clustering method using hep-ex corpus

Finally, in Figure 4, an average of the results with standard deviation is presented.

Figure: Average behaviour of all FSTs with each clustering method using hep-ex corpus