Improving Conversational Aspect-Based Sentiment Quadruple Analysis with Overall Modeling

In this paper, we describe the experimental schemes of Team HLT-base for NLPCC-2023-Shared-Task-4 Conversational Aspect-based Sentiment Quadruple Analysis (ConASQ). Different from the aspect-based sentiment quadruple analysis task, the ConASQ task requires modeling the relationship between different utterances in context. Previous works commonly apply the attention mechanism (e.g., self-attention, transformer) to model the interaction of different utterances after extracting the feature of each utterance. However, this approach may not capture the interaction of different utterances effectively with a single self-attention layer or a transformer layer. To address this issue, we propose a simple and efficient method in this paper. Specially, we concatenate all utterances as a single sentence and feed this sentence into the pre-trained model, which can better construct the representation of utterances from scratch. Then, we utilize different mask matrices to model the features of dialogue threads, speakers, and replies. Finally, we apply the gird-tagging method to quadruple extraction. Extensive experimental results show that our proposed framework outperforms other competitive methods and achieves 2nd performance in the ConASQ competition.