Argument Mining Based on Dynamic Perception of Contextual Information

In this task of argument mining, contextual information significantly improves model performance. Existing research usually concatenates propositions with sentences within a fixed context window as the input of context information for the model. This method presents the following issues: first, sentence concatenation leads to excessively long propositions, which reduces encoder efficiency and causes the model to fixate on redundant or unimportant details. Second, the widespread presence of discourse markers in the training data may influence model performance. To resolve these issues, this paper proposes an argument mining approach based on dynamic perception of contextual information (DPAM). This method abandons the traditional sentence concatenation approach and instead encodes each proposition in the document individually, the word vectors with contextual relationships are obtained through the RoBERTa (A Robustly Optimized BERT Pretraining Approach) pretrained language model, and the context information is integrated by combining the dynamic gated sentence-level attention mechanism. In addition, a data augmentation method is employed to reduce the model's reliance on discourse markers. Experimental results on datasets from five different domains demonstrate our model's superiority over the baseline.