Research And Implementation Of Multi-domain Dialogue Summarization Algorithm In Customer Service Scenario

With the rapid development of the Internet,chat consultation on online platforms have gradually become popular in daily live.In customer service scenarios,summarizing the user’s needs and the customer service’s feedback can back up the problem-solving process,retain the solutions to the problem response,and better improve the service efficiency of customer service.There have been many research works on dialogue summarization tasks that have achieved considerable results,such as introducing context association graphs,triples and other external knowledge to improve the ability of understanding dialogue structure in single-domain dialogue summarization tasks,and achieving domain adaptation by pre-training,introducing auxiliary tasks,etc.in multidomain dialogue summarization tasks.However,there is still little research on customer service dialogue with numerous business domains.The challenges for dialogue summarization research in customer service scenarios are as follows:(1)multiple domains and scarce data;(2)complex dialogue structure and scattered key points;(3)format control in the generation stage.In response to these problems,this paper proposes a multi-domain dialogue summarization framework for customer service scenarios.The main contributions and innovations of this paper are as follows:First,in response to the problem of customer service dialogue involving many domains and scarce data,this paper proposes a low resource domain adaptation algorithm for dialogue summarization based on prompt and adversarial learning,and further explores the ability of model lightweight and domain adaption by combining prompt tuning and adversarial learning.The prompt tuning improves the ability of the summary generation in low resource scenarios by freezing the parameters of the pre-trained language model and introducing a small number of optimizable prefix parameters.The low resource domain adaptation method based on adversarial learning extracts the domain-irrelevant feature through the adversarial learning mechanism,which improves the ability of the model generalization.Second,in response to the problem of complex dialogue structure and scattered key points,this paper proposes an algorithm for focusing dialogue key information based on dialogue state,finds the importance of dialogue states to the generation of customer service dialogue summary,and selects them as the external knowledge。Propose an optimized construction method for dialogue states to improve their compatibility with the algorithm framework of this paper.The method of integrating dialogue state information as domain-specific feature prompt information is input into the model to improve the model’s ability to capture key information.Third,in response to the issue of format control during the generation stage,this paper proposes a style controllable algorithm for generating customer service dialogue summaries.By constructing a reference summary format information that combines hard and soft templates,and designing a prefix-based integration method,this information is integrated into the model decoding stage to guide the model in generating a unified format summary.The algorithm of this paper is experimented on the customer service dialogue summarization dataset with five domains.The experimental results show that the algorithm of this paper has a significant performance improvement,which verifies the effectiveness of the algorithm of this paper.