Personalized Text Generation And Its Applications In Recommendation And Conversation

Recently the developments of big data and artificial intelligence(AI)have demon-strated great influences on users' lifestyles.AI systems capture user's interests and requirements through interactions with the user,and provide corresponding services to gain user's satisfication.Since text is an important medium of user-system interaction,it goes without saying that natural language processing(NLP)is meaningful.Text gen-eration,a classic problem of NLP,is growing to be a popular research topic.Text gen-eration has wide applications;for instance,machine translation,online chat-bot,online marketing and Ads generation.It is worth noticing that in the text generation problem,users have own interests and tastes.Evidence has shown that the users' personal char-acteristics can determine their attitudes toward the given texts.For example,in the ads generation task,the user who is susceptible to social influence may prefer the ads like"10,000 twitter users try it".Thus,personalized text generation leads to improved user satisfaction,trust,experience and persuasion to buy or try an item,and has a significant importance.As an emerging research topic,personalized text generation mainly consists of three problem:1)research scenarios.What kinds of tasks are appropriate and necessary for applying personalized text generation;2)Personalized information selecting.What kinds of personalized information are helpful and can be integrated into text generation task;3)Persaonalized information integrating and user modeling.How to use personal-ized information to effectively modeling users' interests and incorporate them into text generation module.In this thesis,we aim to investigate personalized text generation in the following three traditional/emerging research scenarios including recommendation and conversation.We take fully consideration on the different tasks and different per-sonalized information,and fully integrating them into the text generator to obtain highly personalized texts.We gradually achieve the goal of generating personalized texts from explicit user tags to complex user modeling:(1)In the short text convesation(STC)task,the system is asked to generate a suitable response for a user given post.We consider to incorporate emotion,a major personalized information in the conversation,into text generation.We conduct a data analysis to show the emotion expressed in conversations and design models to mimic human conversations.We propose both stepwise and jointly learning methods that con-sider predicting relevant emotions and generating emotion-aware responses.Our mod-els can determine the relevant emotions based on attention mechanism to reply to a user post and generate responses with appropriate emotions.Our experiments show su-perior performance with emotion-aware responses and generated texts are diverse and personalized.(2)Explainable recommendation is an emerging and popular research topic.Sys-tem recommends an item with text as the explaination,which is helpful for user to understand why the system recommends such an item and can increase user trust and satisfaction.In specific,we investigate how to integrate user historical reviews into both item recommendation and generating personalized explanations.Most existing meth-ods fail to model the cross knowledge of recommendation accuracy and explainability,and cannot jointly optimize the two goals.We propose a co-attentive multi-task learn-ing(CAML)model that tightly couples the recommendation task and the explanation task.We design an encoder-selector-decoder architecture for multi-task learning based on cognitive psychology,propose a hierarchical co-attentive selector to model deep user-item interactions and effectively control the cross knowledge transferred from both tasks by incorporating multi-pointer networks.Extensive experiments demonstrate that our method improves both explainability and accuracy and can generate personalized explanations.(3)In the explainable recommendation,system can provide item recommendations with corresponding text explanations.Evidence show that explanations not only help users understand the working mechanisms of the models,but also serving as a bridge be-tween users and recommender systems and trigger potential user feedbacks.Therefore we introduce explainable conversational recommendation,which provides explanations to help users understand the model,collects user feedbacks to understand,integrate user needs into the model and iteratively refine both recommendation and explanation performances.We design an incremental multi-task learning framework and multiple objectives can be simultaneously achieved through tight collaboration among the rec-ommendation prediction task,the explanation generation task,and the user feedback integration module.Modeling the key concepts that a user likes about an item enables us to derive the cross knowledge between the two tasks,trigger feedbacks about con-cepts,and seamless integrate the feedbacks for incremental model update.We propose a multi-view feedback integration method to achieve effective incremental model up-date.The first view focuses on satisfying user requirements through local propagation of user needs,and the second view better generalizes user feedbacks by updating global model parameters.We evaluate our method with different settings of simulated users and results demnonstrate that the model can effectively integrates user feedbacks and fulfills all the objectives including recommendation accuracy,explanability and user requirement satisfication.In summary,we investigate personalized text generation in diefferent scenarios and introduce how we efficiently model users' personalized information and incorporate it into text generation based on the specific research problem and information.The proposed research works fully take the advantages and effectivenesses of personalized text generation in the different tasks and are also meaningful in practice.