Research On Sentiment Analysis Based On Deep Semantic Features

In recent years, deep learning methods, which are able to express deep semantic features of a sentence by feature combination, have drawn more and more attention in natural language processing(NLP) field. This paper studies three key problems in sentiment analysis based on deep neural network model. They are domain adaptation of sentiment classification, opinion target extraction and opinion target sentiment identification. These methods not only can be used for sentiment analysis tasks, but also can be applied to many other NLP tasks. So the paper has great theoretical value and significance.While researching domain adaptation of sentiment classification, considering that the abundant labelled corpora need hardwork to obtain, the paper proposes joint training method using labelled and unlabeled corpora simultaneously to generate word vectors,and also adopts the strategy of active learning with self-training together to migrate to a new domain. The results of experiment show that joint training method can obviously improve generalization ability of word vectors especially when the size of labelled training corpus is small. Because the word vectors store the most knowledge of sentiment analysis, they just need to be changed a little when migrating to a new domain. As a result of this, the classifier of target domain just needs small times to iterate in order to achieve satisfactory performance. What’s more, using active learning method can alleviate the phenomenon of self-misleading which happens when merely depending on self-training.The experiments show that after employing active learning, the performance of classifier doesn’t decrease much in late iterations and the performance is slightly better than supervised method.While studying the problem of opinion target extraction, considering that in statistical learning methods, features need to be built manually and are also constrained by the size of context window, the paper adopts bidirectional recurrent neural network(BRNN)to extract opinion targets. BRNN introduces forward hidden layer and backward hidden layer to store the information of previous text and latter text respectively which is also not constrained by the size of context window. The paper also adds features such as POS, relation in dependent tree and so on, and compares BRNN with conditional random fields(CRF) model. The results of experiment show that adding linguistic features can improve the extraction performance, and BRNN is better than CRF in the aspect of recall. This method got first in limited resources test in the 7th Chinese Opinion Analysis Evaluation(COAE2015).While studying the task of predicting the sentiment of opinion target, the paper proposes a novel model jointly using Long Short-Term Memory(LSTM) and CNN in order to deal with the problem of using context information insufficiently in traditional methods. The model firstly identifies the sentiment of clause where the specific opinion target lies, then infers the sentiment of opinion target by the result in the first step. In this model,LSTM is used to generate the vector of context and CNN is used to extract features from the vector sequence. This model not only can utilize the information of context adequately, but also can deal with many problems that may influence the identification results such as negators and network new words. The results of experiment show that the model is efficient to identify the sentiment of clause, and the results of predicting the sentiment of opinion target according to the result of clause is obviously better than the methods based on sentiment words.