Does Semantic Information Help in the Text Categorization Task?
Edgardo Ferretti, Marcelo Luis Errecalde, P. Rosso · Journal of Intelligent Systems · 2008
In this paper, we investigate how effective the use of semantic infor-mation could be in text categorization tasks. To this end, we consider distinct representations of documents differing in the kind of information incorporated: (a) information about terms only, (b) semantic information, and (c) a combi-nation of both types of information. Moreover, we study how the vocabulary size reduction affects this task. The k Nearest Neighbours method was used to perform the categorization, and the vocabulary size was reduced by means of the Information Gain technique. A number of different document codifications were tested. The experimental results showed that the inclusion of semantic information in syntactically and semantically richer corpora could improve the text categorization task, if vocabularies with a sufficient number of features are considered. In our view, however, it is not yet possible to affirm that the introduction of semantic information ensures an improvement on the text categorization task. Our results suggest that the performance depends heavily on the particular characteristics of the corpus used in each case. KEYWORDS semantic text categorization, word sense disambiguation, k nearest