Cross-domain Fine-grained Sentiment Analysis Based On Deep Learning

Sentiment analysis is an important research direction in the field of natural language processing.Its main task is to dig out the emotional tendencies of text creators from text resources such as blogs and comments.Existing sentiment analysis methods usually perform analysis based on coarse-grained information such as document level and sentence level,and the result is a generalized evaluation of the whole.This makes the text resources not fully exploited and utilized,resulting in the waste of some resources,and also faces certain limitations in practical applications.In contrast,finegrained sentiment analysis is more in line with the needs of the current era,it can fully mine and utilize text resources,and can obtain more detailed and comprehensive analysis results.With the complexity and diversification of application scenarios,people often need to analyze emerging fields.However,emerging means that there is insufficient annotation data in this field,and manual annotation methods are both inflexible and time-consuming.Therefore,it is particularly important to utilize the labeled data in the existing domain to achieve cross-domain tasks.Based on the above background,this thesis combines cross-domain sentiment analysis with fine-grained sentiment analysis and conducts the following research work:(1)Considering that the static word embedding method cannot distinguish the polysemy of a word,this thesis uses the dynamic word embedding vector to represent the text.In order to avoid the negative impact of non-emotional keywords that are too close to the aspect word on the judgment of sentiment polarity,this thesis calculates the semantic relative distance from the aspect word for each word above and below,and sets a dynamic weighting algorithm for sentiment keywords.Giving larger weights improves the accuracy of sentiment classifications.(2)In order to obtain more fine-grained interactive information and learn more context-related information,this thesis divides the original input sequence into three parts: the above,the aspect words and the following three parts.Using the dual interactive attention mechanism to detect the aspect.The interaction information between the word and the above,below and between the three internal words is fully extracted.The effectiveness of it is verified by experiments.(3)The above two works are combined to form the main part of the cross-domain fine-grained sentiment analysis model which is utilized to extract domain-specific features.Based on the idea of domain adversarial training,a domain classifier including a gradient reversal layer is introduced.Therefore,the model can acquire domain shared features at the same time.Using pseudo-labels for semi-supervised learning avoids the waste of unlabeled data in the target domain and enhances the ability of the model to express across domains,which completes the cross-domain fine-grained sentiment analysis task consisting of four different domains.The good transfer learning ability of the model is proved by the performance on experimental datasets.