Research On Enhancing Question Answering Technologies With Structured Knowledge

Question-Answering(QA)systems can automatically answer users’ questions to provide the specific information users need accurately and quickly.QA systems can improve the user experience and business efficiency,having promising potential for industrial applications.QA systems employ a branch of technology to answer the users’ questions,such as information retrieval and semantic understanding.The research on QA systems can inspire and promote the growth of related natural language processing technologies,making QA a topic with high research value.The ability to capture and understand relative knowledge is essential to answering questions.The methods to effectively retrieve,capture,understand,and leverage knowledge to enhance QA systems are hot research topics that attract many researchers.Therefore,this dissertation focuses on structured-knowledge-enhanced QA techniques that can improve the QA systems’ abilities of question understanding,knowledge retrieving,and answer predicting.Specifically,this dissertation includes the following four research contents:1.The first is structured-knowledge-enhanced supported evidence retrieving for complex problems.The QA system can obtain the relevant knowledge using a retriever.We call this retrieving process the supporting evidence retrieval.When the question becomes complex,multiple pieces of supporting evidence scattered in different documents are required.So the model needs to decide the number of retrieved documents during retrieving adaptively.Moreover,the model needs to locate the desired evidence based on the evidence the model has obtained.For this,we propose a structured-knowledgeenhanced iterative retriever.We first construct the relations between documents based on entity mentions in documents.Then the structured knowledge provided by the relations between documents is encoded and introduced into the retrieval process to enrich the retrieval candidates.The experiments on the open-domain complex question dataset show that the proposed methods can perform better.2.The second is using structured knowledge to analyze the QA ability of pre-trained language models.The patterns that models use to understand and capture the knowledge,whether these patterns are reliable and effective,decide whether these models can give correct answers.Revealing the patterns that models use to process knowledge,and analyzing the reason for underperformance,can pave the way for improving these models.Therefore,we use structured knowledge to construct the probing data and propose a twostage analyzing method,to reveal and analyze how models understand and capture the factual knowledge in the pre-training samples.With the help of feature attribution methods,we first quantify the dependence on patterns that pre-trained language models used to answer factual questions,showing how the model understands and captures factual knowledge.Further,the effectiveness of the different patterns that models depend on is quantified to identify the more beneficial patterns to the model’s performance and provide practical evidence for model improvement.For example,we can guide the model to use more effective patterns for better performance.3.The third is pre-training with structured-knowledge-enhanced masked language modeling.Based on the previous analysis results,we propose to use structured knowledge to improve the masked language modeling samples.We design two pre-training tasks to improve the models’ ability to understand and capture factual knowledge,boosting the model’s performance on QA tasks.Specifically,we use a knowledge base to distinguish and select the pre-trained samples,encouraging models to use reliable relations to understand and capture factual knowledge.Experiments on cloze-style QA,closed-book QA,and extractive QA show the proposed pre-training methods can help the model achieve better results on the QA tasks.4.The fourth is integrating structured knowledge with generative answer prediction.We can refine the model structure,objective function,and training data distribution to improve the answer accuracy.On the other hand,we can control the model output directly.Accordingly,we propose a generative answer prediction that introduces structured knowledge into the encoder-decoder framework to provide more information for answer prediction;and we design a replacement and copy mechanism based on structured knowledge,which injects high-quality structured knowledge directly into the answers to mitigate factual errors produced by the free decoding process.Compared with pre-training,this model introduces structured knowledge more directly.The experiments show that the proposed model can generate more accurate answers.Structured-knowledge-enhanced QA techniques can bring higher accuracy to QA systems and achieve better results.In this dissertation,we propose to use structured knowledge to improve,analyze,enhance,and control the processes of retrieving,understanding,capturing,and outputting knowledge in QA systems.We develop the QA techniques accordingly.Experiments show that using structured knowledge can effectively improve the performance of QA systems,empirically verifying the importance of knowledge for QA.