Generative Knowledge Enhancement For Aspect-level And Dialogue-level Sentiment Analysis

With the rapid proliferation of Internet technology,an increasing number of individuals have joined and utilized the Internet.Consequently,the burgeoning accumulation of Internet data has given rise to the pressing concern of how to efectively harness this data to enhance user experience with utmost quality and efciency.This has emerged as a critical issue in contemporary artifcial intelligence science and technology.Sentiment analysis,a prominent task in the feld of natural language processing,exhibits signifcant scope and utility,providing contextual and technical support for numerous downstream tasks.Unlike traditional sentence-level and document-level sentiment analysis,aspect-level(also referred to as aspect-based sentiment analysis)and dialogue-level(also known as emotion recognition in conversations)sentiment analysis tasks have garnered substantial attention from both the academic and industrial communities due to their distinctive characteristics and inherent challenges.In recent years,signifcant advancements have been made in deep learning techniques,resulting in notable improvements in the performance of aspect-level and dialogue-level sentiment analysis algorithms.Nevertheless,despite these advancements,several challenges persist:(1)In aspect-level sentiment analysis,the model’s comprehension of the interplay between a specifc aspect and its context is constrained due to insufcient information;(2)When addressing dialogue-level sentiment analysis in real-world scenarios where future context remains unknown,accurately discerning the sentiment of a particular utterance solely based on historical information becomes a formidable task.The challenges associated with both aspect-level and dialogue-level sentiment analysis tasks can be succinctly described as the inadequacy of semantic information present in the original text,which poses obstacles to accurate analysis.In order to overcome this challenge,this study proposes the utilization of generative models to generate contextual information as an additional enhancement to the original text,thereby facilitating downstream tasks.The enhanced knowledge obtained through this approach is then integrated into novel models tailored to the characteristics of each task:(1)For aspect-level sentiment analysis,the proposed method combines prompt-based generation to obtain sentence-aware and aspect-aware knowledge for a given aspect in a sentence.Subsequently,an adaptive network incorporates the original sentence embedding and aspect embedding to calculate sentence-enhanced and aspect-enhanced knowledge.The fnal classifcation is performed using an embedding that incorporates external knowledge;(2)In the case of dialogue-level sentiment analysis,the approach involves maintaining local areas of historical and pseudo future information for a specifc utterance.These local areas are then utilized to compute local-aware state,local state,and the evolution of local states,which are subsequently leveraged for the fnal classifcation.Extensive experiments are conducted on three aspect-level sentiment analysis datasets and four dialogue-level sentiment analysis datasets,and the results confrm the efcacy of the proposed methods and models.Furthermore,the experiments and analyses reveal important insights: in aspect-level sentiment analysis,leveraging generative knowledge can aid in detecting texts that exhibit implicit,inconsistent,or multiple sentiments;while in dialogue-level sentiment analysis,utilizing generative knowledge as a proxy for future context can compete with real future context to some extent,particularly in relatively context-independent conversations.Additionally,this work presents the practical application of the proposed approach in the domain of complaints and proposals.