Research On Algorithm Of Answer Selection Task In Question Answering

Question answering system is a very important and challenging task in Natural Language Processing,and answer selection is the key part of question answering system.Given a question and a list of candidate answers,we need to select the correct answer.Most researches on this task regard it as semantic comparison process,in which building the question and answer as vector and calculating the semantic similarity between them.In this work,we focus on semantic comparison on answer selection task,pointing out some problems in the semantic comparison among Attention-based works,and propose some solutions.We validate the effectiveness of our works on three QA dataset by contrast experiments.The contributions of this work are summarized as two parts following:Propose a joint semantic alignment model of words and local semantics units.In the current works on answer selection,using words semantic information only is not able to fully compare sentence semantics.We take the advantage of local semantic units,and propose a joint semantic alignment model.Propose a dynamic semantic thresh model.Aim at the problem that the traditional Attention mechanism will bring in the irrelevant even negative correlation semantic information,due to the all positive result of softmax.We propose a dynamic semantic thresh model based on word and context,to wipe out the noises.