Research On Aspect Level Sentiment Analysis Method That Merge Position Information And Opinion Span

Online we media and e-commerce platforms have gradually become an indispensable part of everyone’s life today.Users are more and more inclined to leave their own comments and opinions on the network.Because these text messages with emotions and opinions affect the direction of public opinion and serve as a reference for businesses and sellers,sentiment analysis has become a hot direction.In a single comment,there are often multiple emotional expressions.In order to analyze the text more accurately,aspect level sentiment analysis was born.Aspect level sentiment analysis can obtain more targeted and accurate emotion expression,which has become a hot topic in current research.Aspect level sentiment analysis models often use long-term and short-term memory network(LSTM)or gated loop unit(GRU)to extract the features of context information.This process often encodes the context alone,ignoring the importance of location information for sentiment analysis,and can not effectively reflect the key information in the text.Aiming at the above problems,this thesis proposes a Bi-LSTM-CRF Based on Position Information and Opinion Span(BLCRF-POS).The main research contents and contributions are as follows:1.Aiming at the problem that the attention mechanism is to calculate the weight of a single word,this thesis improves the popular multi attention network for aspect level sentiment analysis tasks,and introduces conditional random fields to capture the structural dependence of sentences,because the views that affect the target sentiment are usually short and coherent spans rather than incoherent words.In this thesis,conditional random field is used to capture the structural information corresponding to aspect words,so as to identify the corresponding opinions and classify sentiment.In the global aspect,the global features of sentences are extracted through attention network,and the weight of each word in sentence representation is calculated.Through this global and local way,emotional information can be fully extracted from sentences.2.Aiming at the problem that neither LSTM nor attention network encodes the position,this thesis proposes two position attenuation methods to highlight the key information near aspect words.The first method is the position attenuation method based on function.This method takes the aspect word as the center and makes the local context of the aspect word gradually attenuate the weight to highlight the importance of the words around the aspect word.The second method is the location attenuation method that completely ignores the information of nonlocal context.In this experiment,the combination of the two is adopted.The attenuation function is used for weighted attenuation in the local context,and the information of non local context is shielded.3.In order to verify that this research has practical application scenarios,this paper designs and implements a text emotion classification prototype system.