Research On Sentiment Analysis Of Product Reviews Based On Transfer Learning

In the Internet era,product reviews on ecommerce platforms are important information sources for consumers and businesses.Sentiment analysis can automatically extract sentiment tendencies from user reviews by using artificial intelligence and natural language processing technology.The existing sentiment analysis methods require a lot of labeled data,and the cost of human and computing power is high.The emergence of transfer learning methods has improved this.Transfer learning can acquire shared knowledge from the source task and apply it to the target task learning,leading to better model generalization performance and less time and resource consumption caused by the large amount of labeled data required by the target task.Therefore,based on the transfer learning theory,this paper uses deep learning method to conduct research on sentiment analysis of product reviews,and the main research contents are as follows:(1)A cross-domain sentiment analysis model based on fused feature and attention mechanism is proposed for the feature extraction problem in cross-domain sentiment analysis tasks.Firstly,the source domain and the target domain are projected into the same feature space by the pre-trained word vectors and Skip-gram word representation model,and then the importance weights of global and local features are extracted by a Bi-directional Long Short-Term Memory(Bi LSTM)network combined with an attention mechanism.The controlled experimental results on Amazon product public review dataset show that the proposed model efficiently acquires emotional semantic information by learning fused features through the Bi LSTM network layer,effectively solves the problem of sentiment feature drift and importance distinction,largely avoids negative transfer caused by large differences between domains,and significantly improves the accuracy of cross-domain product review sentiment classification.(2)An aspect-level sentiment analysis model based on multi-task transfer learning is proposed for the problem of matching aspect items and sentiment features in aspect-level sentiment analysis tasks.Domain adaptation is achieved by fine-tuning the BERT model in a self-supervised manner on a domain-relevant corpus.Two independent domain adaptive representation learning layers are used to model the local context and the global context respectively,and the syntactic dependency relative distance is used to judge whether a word belongs to the target local context.The context features dynamic mask and the context features dynamic weight are fused as local context features,which are combined with the multi-head self-attention mechanism to interact with the global context features.At the same time,the multi-task joint training of aspect term extraction and aspect polarity classification is carried out.The controlled experimental results on Amazon product public dataset show that the model improves the accuracy of aspect term extraction and aspect polarity classification tasks.The experimental results on book review domain verify that the model can effectively improve the aspect level sentiment analysis performance in book review domain.