Research On Question Answering Based On Unstructured Document Understanding

With the development of deep learning and big data technology,much progress has been made in the research of question answering(QA)system.The QA technology is the key to the success of the QA system.It provides concise and accurate answers through better understanding of the true intentions of users and more effectively meets the needs of users for information.At present,the knowledge contained in the vast amount of unstructured Internet documents can help the question and answer well,and there are still many problems worth researching about QA technology in unstructured documents.This paper discusses the techniques of question-understanding,proposes a method based on unstructured document understanding,and verifies and analyzes through experiments.The main work and research results include:1.Explore the understanding of the questioning technique,including putting forward a model for obtaining implicit understanding of the question based on training question-answer pairs,and a relational extraction model for obtaining the explicit relationship from question.The representations of the question are generated by explicit understanding and implicit understanding.2.For the sentence-level answers from unstructured documents in QA,distance supervised learning has beed used to construct a QA dataset for unstructured documents and a question-answer framework based on understanding representations,information retrieval and ranking learning,has been proposed.First of all,it uses explicit understanding to extend the question to improve the recall,and then add the implicit understanding as a feature to the ranking learning.Experiments show that question understanding can effectively improve the effectiveness of QA for unstructured documents.3.For the fine-grained research of answers in QA,this paper transforms the unstructured document's fragment-based answer into the determination of the starting position and the ending position of the answer fragment,and proposes a model based on deep learning.Improve the network structure for determining the ending location by relying on the initial location information,and the understanding of characters and dependent syntactic information has been creatively introduced on the basis of input word units.Experiments show that the model has a significant improvement over the current excellent model on the SQuAD dataset.This article combines the current popular understanding research in NLP with QA for unstructured documents,which is of substantial significance for the study of open QA.