Research On Cross-domain Chinese Explanatory Opinion Mining Method Based On Transfer Learning

With the improvement of computing power and the accumulation of data,deep learning algorithms have developed rapidly,and explanatory opinion mining methods based on Chinese have also received a lot of attention.Traditional methods of explanatory opinion mining often rely on a large amount of labeled data in the domain.However,the time and labor costs required to produce high-quality large-scale data are too high.Therefore,this paper constructs a high-quality transfer learning data on the online review data of two vertical domain,mobile phone(target domain)and hotel(source domain),and explores cross-domain explanatory comments based on transfer learning based on this data research on mining methods,including the following three aspects:(1)Research on the recognition method of cross-domain Chinese explanatory opinion elements based on transfer learning.This article first explores the problem of opinion element recognition.Its main purpose is to identify the attribute entities in the online comment sequence,the evaluation of the attribute,and the interpretation of the evaluation.The baseline model to solve this task is the Bi-LSTM CRF supervised learning method and uses only small-scale target domain labeled data.On this basis,this article uses large-scale source domain labeled data to learn invariant features and attempts to use the fine-tuned deep migration method to learn the structural features of the target domain.The specific implementation method is to select different neural network hierarchies for fine-tuning.And achieved better results.(2)Research on cross-domain Chinese opinion explanatory classification method based on transfer learning.In the online review data of the e-commerce platform,the user's sentiment polarity and evaluation of attributes are often explanatory.This article divides the explanation of opinions into three categories: reason,suggestion,and condition according to the expression characteristics and logical relationship of the explanation.The baseline method of explanation opinion interpretation classification tasks also uses supervised learning methods based on the attention mechanism,and only uses small-scale target domain labeled data,and then attempts to use DANN-based unsupervised learning methods and corpus-based enhancements based on keyword features Tri-training semi-supervised learning method.Experiments show that both methods have achieved certain effects.(3)Research on cross-domain Chinese explanation fine-grained sentiment analysis method based on transfer learning.Analyzing the emotional orientation of product attributes often best reflects the audience of a product.The emotional tendency of attribute entities is generally determined by the relevant evaluation in the context,so this paper uses the attention mechanism to combine the relationship between context and attribute entities,and serves as a baseline model.Similarly,existing transfer learning methods often focus on the invariant features of the learning domain and ignore the domain-specific features.On this basis,this paper proposes a semi-supervised learning method based on the domain-specific emotional word attention model.Experiments show that this method is superior to other comparison methods.