Research Of Sentiment Analysis Based On The Relations Of Words Model

With the explosive growth of information in big data age,the area of text analysis steps on a new stage.The sentiment analysis area mainly focuses on analyzing and processing information,then extracts or classifies the sentiment of the information.Being profit from the development of machine learning,the trend of sentiment analysis is to be finer-grained,from document level to sentence level,and then to entity level.The difference of texts' grain depicts the difference of demands,and also shows the evolving trail of sentiment analysis area.Comparing to early researches,recent the problems that the researches aim at became finer.By classifying the elements in the text,regarding these elements as entities,and focusing on the entities,researchers got better results.The entities being focused by todays researches are part of text that carry important information,such as opinion towards products in reviews,attitudes towards electors in a social network,and keywords in a stock forum.Entities can be useful in many situations,so how to extract entity is significant for the opinion mining results.Among the present methods of entity extraction,many do the extraction in a pipeline manner,that is to process data only once,and they define relations of words in a relatively simple way.In this paper,a hybrid network model for extracting entity is proposed,which utilizes the relationship "opinion-holder" and "opinion-target" between entities.The model can find entity couples through relations.In the process of constructing this model,our main work are as follow:1)After preprocessing the text,adding features such as part of speech.Sentences in texts are regarded as sequences of words,according to the probability of adjacent roughly extracted result of entities is obtained by a Conditional Random Field model doing sequence tagging task.2)Adding relations between words as a feature in the result of 1).By using the objects as the nodes and the relations as the edges,the hybrid network model is constructed.It concludes two kinds of relations mentioned before.A random walk method is used to grade the candidate entities.3)Analysing the sentiment of the extracted entities.The Recursive Neural Tensor Network is obtained to process entities and sentences,in order to gain the conclusion that sentiment of entities can represent the sentiment of sentences.A series of experiments with two baseline methods show that our model based on the relations of words can extract entities as designed,and outperforms other baseline models.By improving the accuracy of the relationship,the model has a further potential of improvement.