A Model Ensemble Approach for Conversational Quadruple Extraction

Fine-grained sentiment analysis of dialogue text is crucial for the model to understand the conversational participants’ viewpoints and provide accurate responses in generating replies. Unfortunately, in the field of conversational opinion mining, coarse-grained dialogue emotion analysis remains the mainstream approach, despite being unable to meet the actual needs in some specific scenarios such as customer service question and answer system. This work focuses on conversational aspect-based sentiment quadruple analysis, which aims to detect the sentiment quadruple of target-aspect-opinion-sentiment in a dialogue. In this study, we mainly extract triplets and judge the unique sentiment, which is determined by the target and opinion terms together. For this purpose, we fine-tune the pre-trained language models using the DiaASQ dataset. We optimize the rotation positional information embedding by combining the actual length of the dialogue text and use adversarial training to enhance the model’s performance and robustness. Finally, We use beam search ensemble algorithm to improve the entire triplet extraction system’s performance. Our system achieved an average F1 score 40.50 that ranked second in the Chinese dataset and fifth in the general dataset for the Conversational Aspect-based Sentiment Quadruple Analysis shared task at NLPCC-2023.