Abstract: Interactive argument pair extraction aims to identify multiple argument pairs from two related argumentative texts. The task is decomposed into two subtasks by current methods: argument mining and argument pair extraction, which are then jointly modeled for multi-task learning. However, the same sentence semantic representations are shared by these methods, which may lead to information bias within each subtask. Therefore, a two-stage interactive argument pair extraction method based on multi-semantic representation is proposed. First, a pre-trained language model and a bidirectional encoder are utilized in the argument mining stage to capture the semantic information of a single text for argument identification. Then, the two arguments are directly modeled as a text pair for semantic representation in the argument pair extraction stage. A large language model is used to extract key information from lengthy argument content, new samples are generated, key semantic information is introduced, and ultimately the identification of interactive argument pairs is completed. Experiments are conducted on two benchmark datasets, the results show that the proposed method outperformed the baseline method, demonstrating its effectiveness.

Key words: multi-semantic representation, large language model, interactive argument pair extraction