A Research Of Knowledge-Enhanced Question Answering Algorithms

Question Answering(QA)is an important research direction in the field of artificial intelligence.In recent years,data-driven deep learning algorithms have been shown to be successful in question answering tasks.However,existing deep learning based-question answering models that rely on large-scale data annotation and training still have some limitations,such as weak interpretability,and the inability to reason well.To address these issues,cognitive intelligence is gaining increasing attention.Knowledge-aware question answering aims to move towards a new generation of cognitive intelligence,where the incorporation of the prior knowledge assists models to perform explainable reasoning and improves the accuracy and generalisation of question answering task.Compared to traditional QA algorithms,the biggest challenge of this combined knowledge solution is how to effectively represent and fuse symbolic knowledge on a neural approach.The new neuralsymbolic algorithms are yet to be investigated.In this thesis,we focus on QA scenarios in everyday life that require the assistance of commonsense knowledge,and proposes novel knowledge-driven QA algorithms to verify the validity and interpretability of the use of knowledge,with the following main work.(1)For activity-assisted QA robots,they need to correctly understand not only visual and verbal input but also reason with commonsense knowledge of everyday activities in order to improve the accuracy and interpretability of responses.For this reason,this thesis extends the activity knowledge base,which covers a rich sequence of everyday activities,and designs the Human Activity QA task to verify the validity of this knowledge base.The task requires the selection of reasonable action sequence answers based on primitive video and textual intent.In order to fully exploit the activity knowledge,a neuralsymbolic planning(NSPlan)method based on knowledge retrieval and reasoning is proposed.The NSPlan approach combines a visual deep model with symbolic knowledge to retrieve relevant sequences of action sequences from a multimodal knowledge base as evidence for probabilistic reasoning about all answers.The experimental results show the performance of NSPlan model beyond deep learning-based schemes.Moreover,this approach can address the weak interpretability of neural methods.(2)In addition to the general knowledge of the activity,question answering machines often need to combine common knowledge of the world with reasoning when faced with more diverse and complex questions.In this thesis,we present KB-GNCAN,a knowledgeenhanced neural model incorporating common knowledge from commonsense knowledge base.The model uses a designed retrieval algorithm to obtain a subgraph of knowledge related to question-answer pairs from knowledge base,from which relationships between question-answer pairs and commonsense knowledge are mined for inference.After considering the influence of question categories on the selection of relational edges,question category identification is introduced and a novel common knowledge graph neural network using question-relational attention is designed for inference.And the QA-knowledge graph interaction are performed through a Co-action feature interaction mechanism.By comparing with some existing pre-trained language modeling methods,the method can effectively use knowledge to improve the accuracy and interpretability of question answering,and it performs superiorly in methods also combined knowledge.