Machine Question Answering Technology Based On Deep Learning

Machine question answering technology is one of the fastest growing and most im-portant technologies in natural language processing technology.Machine question an-swering technology can help people quickly and conveniently obtain answers to questions from the huge amount of text information generated by the rapid development of the Inter-net and information technology.In this thesis,we focus on various aspects of long-context multi-hop QA techniques.We propose a two-stage long-context multi-hop QA framework based on support facts inference stage and precise question answering stage,to help users get answers to questions from long text information quickly.The main work of this thesis is as follows:1.This thesis proposes a two-stage long-context multi-hop QA framework based on support facts inference stage and precise question answering stage.For long-context multi-hop QA tasks,this thesis proposes a two-stage QA framework based on the two-step process of information selection and question answering that humans perform when doing long-context reading comprehensions.The two-stage QA framework consists of two mod-ules to complete two different types of work in long-context reading comprehension.By understanding the question,the support facts inference module finds the support sentences that can be used to answer the question from the long context,and filters out the sentences that are irrelevant to the question or cannot be used to get the answer,thus significantly re-ducing the context length of the QA task.The precise question answering module uses the supporting sentences obtained by the support facts inference module as the short context of the QA task,and performs the QA task to obtain the answer.Experiments have proved that compared with the traditional End-to-End QA model,the two-stage QA framework this thesis proposed can greatly improve the performance of long-context multi-hop QA tasks,and solve the problem of low QA system performance caused by the interference of irrelevant sentences in long context.2.This thesis proposes a multi-step support facts inference model based on clue up-dating for the support facts inference module of the two-stage QA framework.The multi-step support facts inference model is designed based on the idea that humans perform multi-hop QA tasks.The model searches for support sentences from long contexts iter-atively,and makes use of all the existing clues as much as possible,which improves the performance of the supporting facts inference module.3.This thesis proposes a support sentence inference model based on interaction at-tention between sentences for the support facts inference module of the two-stage QA framework.The support sentence inference model based on interaction attention between sentences,through the attention mechanism to make the interaction between sentence en-codings,so that there is information sharing between the sentences in the context and the contextual sentences are no longer independent units.The model improves the accuracy when supporting facts inference,thus providing more accurate and cleaner short contexts for the precision question answering module of the two-stage QA framework and improv-ing the performance of long-form multi-hop QA systems.