Um modelo temporal-relacional para classificação de documentos

Abstract

Automatic Document Classification (ADC) is one of the most relevant and challenging research problems in Information Retrieval. Despite the large number of ADC techniques already proposed, few of them take into consideration characteristics of the human language. As discussed in recent studies [Montejo-Raez et al., 2008; Chen, 1995], understanding and considering such characteristics may benefit ADC. Therefore, in this work we propose a new network-based representation for textual documents that is based on fundamental concepts of Linguistic, in particular those associated with relationships between terms. Using the proposed model, we also introduce a relational algorithm for ADC which exploits such relationships. Experimental evaluation of this algorithm shows that it achieves results that are comparable to SVM in four real datasets. In addition, its simplicity, execution efficiency and a simple parameter tuning are characteristics that make our algorithm an interesting alternative to SVM. A deeper analysis also shows that there are several dimensions in which relational algorithms may be enhanced. Due to its relevance, particular attention is given to the temporal dimension. In fact, changes occur spontaneously at every moment affecting settings and observations made previously on the term network. Considering this evolving behavior may be very useful in the area of Information Retrieval [Alonso et al., 2007]. In order to incorporate the temporal dimension to our algorithm, we attach to every relationship of our network information about the moment of its construction. The evaluation of simple temporal versions of the proposed algorithm showed that considering the temporal evolution has improved the performance of our relational classifier, by providing more accurate information about the behavior of each term. A preliminary assessment of other dimensions of analysis, such as information scarcity and the use of attributes of relationships, also showed that more elaborated techniques to address such dimensions may benefit the proposed algorithm. Further, considering the generality of the linguistic concepts incorporated in this work, we believe that our proposal may be equally successful in various ADC application domains.