Text Comprehension Based On Resolution Of Anaphora

Text understanding refers to the process of extracting information and analyzing the semantics of text to comprehend its true meaning,which can be applied in advanced applications such as knowledge extraction,common-sense reasoning,dialogue systems,and machine reading comprehension.One of the difficulties in text understanding is the existence of a large number of referring relationships,i.e.,the relationships between certain words in the text and other words,such as coreference and bridging relationships,which are crucial for accurately grasping the semantic information of entities,events,and other concepts in the text.Therefore,correctly identifying referring relationships is essential for text understanding.However,existing anaphora resolution algorithms have problems such as insufficient utilization of global information and small data sets,resulting in poor model performance.Based on the above issues,thesis focuses on algorithmic research on two different types of reference relationships,with the following specific research topics:(1)A span-boundary based coreference resolution algorithm is proposed,which models the text as a graph structure and introduces global graph structure information.Existing coreference resolution algorithms based on mention pairs rely on span representations,and frequent span representation construction leads to high model complexity in terms of time and space.In addition,the algorithm also has insufficient utilization of global information.Therefore,thesis proposes a span-boundary based coreference resolution algorithm that calculates using span boundaries instead of entire spans and models the text as a graph structure.The global graph is embedded into the model for iterative optimization,thus effectively utilizing global information.Experimental results show that the performance of the proposed algorithm based on span-boundary representations is comparable to the baseline model,and the performance is improved after incorporating global information.(2)A bridging resolution algorithm based on multi-task learning is proposed,which introduces inter-task consistency constraints and pretrains the shared representation layer.Existing bridging resolution algorithms have difficulty effectively training neural network models due to small dataset size.Therefore,thesis propose a multi-task learning-based bridging resolution algorithm,which jointly learns bridging resolution,coreference resolution,and information status classification tasks.Thesis use these two highly correlated subtasks to improve the performance of the bridging resolution algorithm and introduce inter-task consistency constraints from the information status,supervising the overall training of the model to further improve the performance of each task.Experimental results show that both multi-task learning and inter-task consistency constraints can effectively improve the performance of the bridging resolution algorithm.(3)A multi-turn dialogue system that integrates anaphora resolution is completed.To verify the role of coreference resolution in text understanding,this paper designs and implements the system.Specifically,this paper uses the above-mentioned coreference resolution algorithm to resolve the historical dialogue,and integrates the result of the coreference resolution into the sentence rewriting model to rewrite the user’s current dialogue,eliminating the influence of the current dialogue on the historical dialogue.Rely on,so as to understand the user’s intention more accurately,so as to achieve effective chat interaction.