Aspect-Level Sentiment Analysis For End-to-End Solutions And Implicit Aspect Extraction

At present,our country’s Internet users account for more than 70%of the country’s population,generating petabytes of data every day.These data contain rich user emotions,and contain huge economic and social value.The use of the data generated by the Internet to completely construct the user’s emotional world is of great significance for user choice,enterprise product improvement,and government policy implementation.This paper focuses on the end-to-end solution and implicit aspect extraction problem in aspect-level sentiment analysis.Based on the multi-level information extraction model,an end-to-end solution is designed to obtain the emotional information contained in the text in a more fine-grained and systematic manner.The aspect library is constructed using explicit information,and the cooccurrence rules and neural networks are fused to obtain the implicit aspects of the text more completely.The main work of this paper is as follows:1)To effectively utilize the information contained in the text,a two-stage multilevel information fusion mechanism is proposed.Different information extraction methods are designed for corpus,sentence,and vocabulary levels,and specific fusion mechanisms are provided according to the characteristics of tasks.A two-stage training process is adopted to fully integrate information at all levels into natural language processing tasks.This mechanism is used for subsequent end-to-end aspect-level sentiment analysis.2)An end-to-end aspect-level sentiment analysis algorithm based on a multi-level information extraction model is proposed.The algorithm clearly defines the end-to-end task and its core demands,and transforms it into a sequence labeling problem using unified labeling.Based on the two-stage multi-level information fusion mechanism proposed above,the BERT pre-training model is used to obtain word vector semantics;the sentiment word extraction task is used to obtain key sentiment information;the whole sentence domain sentiment extraction task is used to obtain global sentiment and domain,so as to effectively use corpus,Sentence,lexical level information.Through a two-stage training process,multi-level information is incorporated into aspect-level sentiment analysis tasks.Experiments show that the above model has higher effectiveness and accuracy than other mainstream models.3)An implicit aspect extraction algorithm that integrates co-occurrence rules and neural networks is proposed.The algorithm clarifies the concepts of explicit aspect unit,implicit aspect unit and continuous aspect unit,and explicitly defines the implicit aspect extraction task through formulas.The aspect feature library is efficiently constructed using explicit aspects,and the implicit aspect extraction problem is transformed into an implicit aspect unit classification problem.By integrating co-occurrence rules and neural network for aspect prediction,all complete aspect units of the whole sentence are recursively obtained to realize implicit aspect extraction.The experimental results show that the algorithm proposed in this paper can efficiently and accurately extract the implicit aspects appearing in sentences,and has advantages in efficiency,cost and accuracy.