The Application Of Heuristic Knowledge In Fine-grained Sentiment Analysis

The thriving of mobile Internet has contributed to the generation of oceans of data every day. A large proportion of these data is customer reviews, which demonstrate the customers' attitudes and opinions. By performing sentiment analysis on these reviews, valuable information such as customers' preference and personality can be mined. In a feature ensemble way, we exploit convolutional neural networks to perform fine-grained sentiment analysis with several word embeddings as features. The major work is as follows:1, the improvement of word embeddings' feature expressing ability.The current word embedding is a general word representation without.optimization for specific task. Targeting at sentiment analysis, we introduce useful heuristic knowledge into word embedding so that its feature expressing ability is enhanced, and thus improving the diversity of features. And we also compare the difference of modified models'expressing ability on several aspects.2, the design and implementation of a multi-branch convolutional neural network. By employing several word embeddings in separate lookup tables, a convolutional neural network is able to exploit differentfeatures at the same time. The experiments on movie reviews and product reviews prove the effectiveness of this feature ensemble method. To understand why, we explore how the features influence the classifier's behavior. Further migrating application to question classification is also effective, and we find the reason behind.