O problema da incoerência e a regularização semântica para inferência textual

Abstract

Recognition of textual entailment is a task of natural language processing that aims to measure the ability of algorithms to compare a pair of sentences at the semantic level. It is useful directly or indirectly in various applications such as machine translation, summarization and question answering. This task received attention with the release of the SNLI dataset, making possible the application of complex deep learning techniques that obtained several expressive results. Some works, however, begin to question such results, observing the biases explored by the learning algorithms. This dissertation discusses another possible problem of these methods: the incoherence between the answers. A formal definition, based on propositional logic, is presented for what is a coherent response. It is also given a solution that seeks to reduce the incoherence of the models, applicable to any deep learning algorithm, as well as experiments that evaluate some possible impacts of incoherence and the effectiveness of the proposed solution.