Personalized Text Generation Based On User Representation Enhancement

Text generation,which converts different forms of input into textual output,gives computers the ability to communicate with people,and is a hot area of research in natural language processing.Due to the rapid development of computing power,text generation technology based on deep learning has achieved success.With the vigorous development of the Internet industry,many directions of text generation,such as image captioning,neural machine translation,and text summarization have also been widely used.A large number of scenarios provide researchers with a solid data foundation.Personalized text generation in these directions has strong practicability and social value because it is close to the needs of users.Although there are better solutions for general text generation problems,the problem of personalized text generation has not been well resolved:1)The dynamics of users' literal preferences are ignored and the changes in users' literal preferences over time are failed to capture and model.2)The transferability of users' literal preferences are not considered,therefore there is no data of the users in the new field,and there still are many problems in how to use the original data to construct users' representation.This paper researches into the problem of user personalized text generation,from the direction of capturing the dynamics and migration of users' writing to improve users' representation.The main work of this paper is as follows:In the task of generating personalized picture descriptions for users,previous work focused on using user high-frequency words and user attribute information or simple user representations,but ignored the dynamics of users' literal preferences.This paper proposes a multi-modal hierarchical generative model(MHTN)to solve this problem.This work starts with the analysis of the users' short-term preference to the target text,and verifies the contribution of the users' short-term preference to the target text.A hierarchical Transformer model based on the self-attention mechanism is used to model the users' short-term literal preferences,users' long-term literal preferences and image features.In the process of capturing users' short-term literal preferences,this paper uses the self-attention mechanism to obtain text representation,and also uses user characteristics to select short-term preferences.At the same time,it incorporates a time difference representation mechanism to improve the characterization effect of users' short-term characteristics.The experiment on the real data set shows that our generation method has achieved good results,demonstrating the effect of constructing dynamic user representations.In the cross-domain personalized review generation task,the problem of the source domain users' lack of data in the target domain is rarely involved.This paper uses the idea of transfer learning to propose a cross-domain personalized review generation model(CDPG).Based on the domain discriminator,this paper extracts domainindependent feature vectors from the data of the users' source domain,and constructs the user representation by paying attention to the item and the users' current target phrase.In addition,in the decoder part,in order to efficiently use the input characters and the current item information,this work also uses a copy mechanism based on input characters.The experimental results on multiple public data sets show that the method has a significant improvement in the generation effect and demonstrates the transferability of user representations.In summary,this work proposes two novel models in terms of user representation's dynamics and transferability for personalized text generation.The experiments performed on multiple public datasets demonstrate capturing the dynamics and transferability of user writing can improve user representation and boost the performance of text generation.