Humans tend to organize information in documents in a logical and intentional way. This organization, which we call textual structure, is commonly in terms of sections, chapters, paragraphs, or sentences. This structure facilitates the understanding of the content that we want to transmit to the readers. However, such structure, in which we usually encode the semantic content of information, is not usually exploited by the filtering methods for the construction of a user profile. In this work, we propose the use of term relations considering different context levels for enhancing document filtering. We propose methods for obtaining the representation, considering the existence of imbalance between the documents that satisfy the information needs of users, as well as the Cold Start problem (having scarce information) during the initial construction of the user profile. The experiments carried out allowed to assess the impact, in terms of T11SU measure, on the filtering task of the proposed representation.
Seed-Guided Topic Model for Document Filtering and Classification
