Fine-grained Sentiment Analysis For Product Reviews

Nowadays,more and more users shop online as the scale of electronic commerce becomes larger and larger.As a result,many reviews for various products are springing up.Both users and manufactures can be benefit from these reviews,as users can get more detailed information about the products and manufactures are able to collect feedbacks without any additional efforts.However,the reviews cannot be examined manually due to its scale,therefore the overall evaluation from customers cannot be easily retrieved.Benefit from the growth of the computing capacity and the arrival of the big data era,Natural Language Processing,as an important research and application field of Artificial Intelligence,has played an irreplaceable role on many tasks.With the capacity of manipulating large scales of data and the advantages of lower cost and higher effectiveness,computers will help to reduce manpower and resources if they can analyze the users' attitudes in product reviews.Sentiment Analysis is one of the most important research areas of Natural Language Processing.Sentiment Analysis algorithms can not only automatically retrieve users' attitudes from documents and sentences,such as positive,negative or neutrally,and also analyze the user's emotions,such as joy,sadness,surprise,etc.However,sentiment analysis at document level or sentence level usually fails to capture the object of the attitude expressed by the user.In the analysis of product reviews,we are interested in the aspects to which users expressed the affirmation or dissatisfaction.Fine-grained Sentiment Analysis can solve this problem very well,and finding the aspects and the opinions from users' reviews is one of the most steps in Finegrained Sentiment Analysis.In this paper,we use methods that based on Recurrent Neural Networks to extract aspects and opinions.Aspects and opinions are identified by annotating each word in the sentence.For sentiment collocation extraction,we use two relation classification algorithms in this paper,respectively as the method based on Support Vector Machines with Tree Kernels,and the method based on Neural Networks that embeds the sentence structure information.These two methods are combined with syntactic relations by fully using the syntactic relationship between words,to determine whether the two specific words are the correct collocation of aspects and opinions.The experiment results show that these three methods are all effective in their tasks.In details,the LSTM or GRU-based methods is out performed than the rulebased words extraction algorithm.For the task of collocation extraction,the performance of the methods that combined with syntactic structures is significant improved,which shows the effectiveness of syntactic structures in the task of relation classification.Finally,the hybrid model based on Convolutional Neural Networks and Recursive Neural Networks achieves a better performance because they can better model the semantic information expressed in sentences.