Research On Sentence Rewriting Based On Relevant Context In Multi-Turn Setting Dialogue

Open-domain dialogue systems have achieved great success due to the availability of one-round corpora and the development of deep learning,but multiple rounds of dialogue make it difficult for machines to understand their real intentions due to frequent references and information omissions.In recent years,the task of rewriting incomplete statements has attracted much attention,that is,rewriting the statements as a preprocessing process to help the subsequent multi-round conversation modeling.The task of this paper is to construct a sentence rewriting model by extracting contextual information related to incomplete sentences,so as to recover all the key references and omitted information,help the machine understand the entire sentence,and make more appropriate responses.Based on the research and analysis of sentence rewriting,this paper constructs two sentence rewriting models with different ideas.Firstly,solve the problem of sentence rewriting from the perspective of text generation,analyze the limitation of general attention mechanism in SEQ2 SEQ model,and introduces multi-headed self-attention mechanism into Seq2 Seq,the relation between incomplete sentences and historical dialogues is captured by self-attention mechanism,and the rewritten sentences are generated by decoder.Secondly,solve the problem of sentence rewriting from the perspective of fragment selection,the problem is transformed into the reading comprehension problem of fragment selection.The data is processed to be suitable for the task of reading comprehension,by using pre-training model Ro BERTa fine-tuning to predict the index of the default information location,then the default information is used to achieve sentence rewriting.Finally,experiments are carried out on the multi-round dialogue rewriting dataset,and the comparison with the classic model demonstrated the effectiveness of the two models in this paper,which are embodied in BLEU,ROUGE and EM indicators.