Research On Fine-grained Sentiment Analysis For Product Defect Discovery From E-commerce Reviews

Product defect discovery refers to the analysis of user-generated text content to mining product defect information.Product defect information can help consumers learn product defects,and manufacturers learn the deficiencies of product to improve products and improve user experience,and quality inspection personnel can supervise and check product quality to improve work efficiency and reduce labor costs.Two sub-tasks were studied in the field of fine-grained sentiment analysis,namely,sentiment analysis based on target extraction and triplet(target,opinion entity,sentiment polarity)extraction.Target and its corresponding sentiment polarity can be extracted from the input text through sentiment analysis based on target extraction.The current mainstream processing method for this task is serialized annotation.The mainstream methods of triple extraction include the pipeline method and the union method based on(serialization annotation,relationship classification).But the existing research methods have some problems when applied to product defect dicovery task : the syntactic parser can not handle existing product reviews which grammar is nonstandard;In addition,the serialized annotation method leads to the independent prediction results of each word in the sentence.Such sentiment inconsistency will lead to inadequate product defect information mining.There are multiple target and multiple opinion entities in the review text,Although the existing methods consider their relationship,they cannot effectively extract the semantic relationship between targets and opinion entities.Aiming at the above three problems,we improves the existing method make the fine-grained sentiment analysis method more suitable for product defect discovery task:1)For defect discovery scenario some product reviews' syntax is nonstandard,syntactic parser of the existing work is difficult to mining the interaction relationship between term in reviews.An opnion guied model is proposed by using kernel method to build higher dimensional interaction information to focus on the relationship between the target and the context,and it does not need to pay attention to reviews' grammatical information.According to the kernel method,the model calculates the feature representation of words to generate the high-dimensional feature information vector,which is used to judge the sentiment assigned to each word.The feature vector containing sentiment is used to guide the model to predict the sentiment label information.Finally,the validity of the proposed method is verified by experiments on the benchmark data set.2)To solve the problem of imperfect product defect information mining caused by sentiment inconsistency,bidiredctional gate-based attention mechanism is proposed to improves the attention mechanism.The bidirectional gating unit enables the attention mechanism generate consistent weight vectors,so that the sentiment information of the current word is consistent with the sentiment information of the same phrase.Finally,the validity of the proposed method is verified by experiments on the benchmark data set.3)For defect discovery scenario product reviews have multiple targets in the text,the existing work extract semantic relations between targets and opnion entities inefficiently.An attention mechanism based on orthogonal regularization is proposed to extract relationship between targets and opinion entities,using the orthogonal regularization restriction targets focusing on different words.we also consider the location information,The position embedding method is used to judge the relationship between words in relation extraction stage to improve the performance of model prediction.Finally,the validity of the proposed method is verified by experiments on the benchmark data set.