Abstract

We are in the Internet era. Now, more than ever, there is an excessive amount of documents available online, making it harder and harder to find the ones that are relevant to our searches. The new systems, which use belief models to improve the understanding of what the user really wants, need a practical text representation model that can make the extraction of beliefs easier. Profiling is an appropriate method for creating such a text representation. In this thesis we develop the details for a process of building basic profiles from reported speech in newspaper articles . This process recognizes and extracts reported speech, and determines the internal structure of reported speech clauses, fitting it within basic profile frames. In order to accomplish this, three building blocks are used: recognition of reporting verbs, punctuation analysis and the structure of the reported speech . We implemented a fully automated system which extracts basic profiles and analyzes them, differentiating between circumstantial information and primary information. Inside the primary information, a differentiation is made between direct quotes and paraphrases. The source, reporting verb and other circumstantial information are extracted. The reporting verb is analyzed, determining its tense, aspect, modality, mood, voice and polarity. The semantic dimensions of the reporting verb are also analyzed, paving the path for a future belief analysis based on reporting verbs.