In the traditional triple extraction task,the extraction of aspect terms and opinion terms is interrelated,when employing pipeline or joint model architectures,issues such as error propagation and misalignment often arise.To tackle these challenges,this paper presents the Position-prompt Dual-channel Recurrent Neural Network (PDRN)model for solving the triple extraction task.The PDRN model utilizes pre-trained BERT models to generate word embeddings as model inputs,establishing a synchronized mechanism through dual-channel explicit interactions in multiple loops for the extraction and pairing of aspect and opinion term pairs,and employs a position-prompt-based BERT-BiLSTM model for sentiment polarity classification.Experimental results on three triple extraction datasets show an improvement in F1 scores by 1 and 2 percentage points compared to the best-performing pipeline model and similar j oint models.Moreover,on the ASOTE task,the F1 score is enhanced by 2.9％ compared to the baseline.
关键词
情感分析/情感三元组抽取/深度学习/BERT模型
Key words
sentiment analysis/sentiment triplet extraction/deep learning/BERT model
维普
万方数据