Aspect-level Sentiment Analysis Research based on XLN et-LCF

Abstract

At present, the commonly used aspect-level sentiment analysis method is to combine the neural network model and the attention mechanism, use the neural network to mine the semantic features of the sentence, and the attention mechanism assigns the weight of the emotional words in the sentence. However, when there are multiple aspects in a sentence, and the aspect words are uncertain, methods that rely solely on the attention mechanism cannot effectively distinguish sentiment words from different aspects. Therefore, an aspect-level sentiment analysis model based on XLNet-LCF is proposed. The model obtains contextual semantic features bidirectionally through XLNet pre-training, and introduces a context focus mechanism to capture the local context of the context with the aspect word as the focus, which can effectively reduce the impact of different aspects. To solve the problem of mutual interference between emotional words, we combined the multi-head self-attention mechanism to deeply extract the semantic features in the global context to construct an emotional weight matrix. Finally, the matrix was normalized to improve the training speed and input the emotional analysis layer to judge the emotional polarity. The model is tested on three public datasets, Laptop, Restaurant, and Twitter. The results show that the accuracy rates of sentiment analysis of the XLNet-LFC model reach 80.12%, 85.24%, and 81.2%, and the F1 values reach 76.59%, 76.9%, and 73.37%, the overall performance is better than the comparison model.