Research On Aspect-based Sentiment Analysis For User Reviews

With the development and popularization of social media and e-commerce platforms,the internet has accumulated massive amounts of user reviews,which contains rich emotional information.Analyzing and mining these reviews can clarify users’ emotional polarity,provide comprehensive guidance and decision-making references for government,businesses,and consumers.However,it is difficult to process the exponentially growing review texts by hand,how to analyzing and mining on user reviews is a current research hot topic in the field of natural language processing.Aspect-based sentiment analysis,as a fine-grained sentiment analysis,aims to obtain the sentiment polarity of specific aspects.Performing aspect-based sentiment analysis on user reviews can obtain the sentiment polarity of different aspects,which has high research significance and reference value for business,social,and political fields.With the expansion of deep learning technology,extracting semantic features through pre-trained models has gradually become the mainstream approach in the sentiment analysis field.However,most existing models suffer from significant gaps in upstream and downstream tasks,insufficient interaction between aspect words and contextual words,and a relatively simple semantic feature extraction.Therefore,aiming at the characteristics of online reviews,an algorithm that integrates prompt knowledge and a multi-feature model of fusion position-aware are proposed.The main contributions of this thesis are as follows:(1)Aiming at the problems of the inconsistent pre-training objective and downstream tasks and model the relationship between aspect words and context is generally limited,an algorithm that integrates prompt knowledge is proposed.Firstly,in order to capture the semantic relation between aspect words and context effectively and enhance the model’s awareness of sentiment analysis,a prompt text is designed and inserted in original sentence and aspect words as the input of BERT model.Then a sentimental vocabulary is built and integrated into verbalizer,which bridges a projection between vocabulary and label words,so as to limit search space of label words,obtain rich semantic information in the vocabulary and improve the leaning ability of the pre-rained model.The F1 scores achieves77.42%,75.20%,and 94.89% on the Restaurant,Laptop,and Chn Senti Corp datasets,which are higher than other compared models and verifies the effectiveness of the proposed method.(2)Aiming at the problems of the limited semantic feature and insufficient semantic information extraction in long reviews,a multi-feature model of fusion position-aware is proposed.Firstly,in order to guide the pre-trained model to adapt to downstream tasks and enhance the model’s semantic features extraction ability,a prompt text is inserted in model’s input.Then,the Roformer model,which is based on rotary position embedding,is used to model text positional information and word-level semantic information to fully explore the semantic features contained in the position relationship and obtain important information adjacent to aspect words.Finally,a multi-feature extraction layer that integrates positional information is constructed based on Bi LSTM and CNN to capture global and local semantic features fusing position-aware in long review text.Those above multi-level features are fused to enhance context semantic relevance of long reviews and strengthen interaction between aspect words and adjacent key information.The experimental results show that F1-score on the AI Challenger 2018 dataset can achieve77.81%,demonstrating a certain advantage compared to other mainstream models and verifying the effectiveness of the proposed model.