Text Generation Over Emotion-enhanced User Intent Understanding

Intelligent text generation is an important research direction in the field of natural language processing.Building a text generation system that can express high-quality natural language as humans has always been the long-term goal of artificial intelligence.Intelligent text generation can be divided into four categories:text-to-text generation,data-to-text generation,meaning-to-text generation,and image-to-text generation.Text-to-text generation research is the widest direction and the basis of other research categories.Therefore,this thesis focuses on the research of text-to-text generation.In recent years,with the development of deep learning technology and parallel computing capabilities,intelligent text generation has been playing an important role in various applications of natural language processing.For example,it is used in intelligent dialogue systems to meet the companion needs of users and achieve more intelligent human-computer interaction;it also can be applied in text summarization system,by compressing and refining the original text,and finally generating a concise text description,improving work efficiency.The wide applications of text generation technology have prompted much cutting-edge research in academia and influential application products in the industry.The end-to-end deep neural network makes it possible to generate text in an abstractive way.It first converts user input text into deep vector representations and generates target text based on the deep vector representations.Compared with retrieval text generation,neural text generation has been increasingly studied by researchers because it is not restricted by domains.However,the current research on neural text generation still faces many challenges.First of all,understanding the user's feelings and emotions is an important skill of the intelligence text generation system(i.e.,empathy ability).Since user text is often subjective and contains many emotion-based opinions,simple linear operations cannot effectively model the high-quality vectorized representations of the user's input text.Secondly,good semantic representations to represent user input text is the prerequisite for generating high-quality text.Existing natural language generation technologies often generate text directly based on deep semantic representations.Lacking core components analysis and dependency modelling for user's input text further restrictions on the development of neural text generation.Therefore,this thesis puts forward a new exploration angle for the problem of intelligent text generation based on emotion-enhanced user intent understanding.Emotion is an important component of the user's intention.It reflects the user's attitude and sentiment for an item.To build a human-level text generation system,it is necessary to cultivate the emotion perception ability of the machine.Therefore,this thesis further proposes a new text generation paradigm TET:first,perform emotion-enhanced user intent understanding,and then generate text based on emotion enhanced intent representation.This thesis applies this generation paradigm TET to two representative text generation tasks:opendomain dialogue generation(the generation goal is the response generation)and opinion summarization(the generation goal is the opinion tag sequence generation).This thesis also analyzes the importance of emotion-enhanced user intention understanding for text generation performance improvement.In the dialogue scenario,according to the TET paradigm,this thesis designs a dialogue system based on emotion-enhanced intention understanding(TET-DG).TET-DG considers multi-grained emotional information(i.e.,coarse-grained emotional labels and fine-grained emotional words).In order to capture the user's subtle emotions,TET-DG enhances the deep vector representation of the user's intentions through additional modelling of the multi-resolution emotional information.Besides,TET-DG introduces user feedback information to further enrich the deep representations of user intent.It strengthens the dependency modelling between user intent understanding and response generation through the posterior estimation of the generative adversarial learning.In the opinion summarization scenario,according to the TET paradigm,this thesis designs an opinion tagging system(TET-TG)based on emotion-enhanced intention understanding.First,TET-TG combines supervised learning and unsupervised learning algorithms to perform the emotion strength estimation and intent importance analysis for each user's input text.The deep vector representation of the user's intention is the user text weighted by its emotional strength scores.TET-TG particularly incorporates two alignment mechanisms that explicitly learn core information from the deep representation of user intent to guide the generation of opinion tag sequence.For the dialogue task,this thesis uses a public benchmark dataset of empathetic dialogue to conduct a large number of quantitative and qualitative experimental analysis.For the opinion tagging task,due to the lack of emotion-enhanced opinion tagging datasets,this thesis collects a large-scale dataset from several Chinese e-commerce platforms and performs sufficient experiments.Compared with the existing methods,the TET paradigm has a significant effect on the two text generation applications.This thesis additionally designs additional analysis experiments to explain the effectiveness of emotion-enhanced intention understanding and its positive impact on text generation performance.