Abstract: The purpose of this paper is to investigate argument extraction techniques, in order to identify, extract, and analyze argumentative components and structures in textual information. The intelligent analysis of debate fact text is accomplished by extracting arguments related to the topic of debate from multiple sentences and determining whether the position of the argument is supportive or oppositional. Previous research has mainly relied on deep learning models such as convolutional neural networks and recurrent neural networks, which have simple network structures and cannot learn deeper features from arguments. In order to learn richer semantic information from argumentative text for position classification better, this paper proposes an enhanced RoBERTa model (EnhRoBERTa) based on the pre-training language model RoBERTa, which fully utilizes the multi-level multi-head attention mechanism and extracts shallow and deep semantic representations for fusion, enabling a comprehensive understanding of the relationship between arguments and debate topics from multiple feature dimensions, thereby facilitating argument position classification. However, considering the problem of imbalanced distribution of position in argumentative points, this paper adopts data augmentation techniques to enhance the learning ability of scarce samples. The experimental results on the CCAC2022 match data set show that the proposed model can extract more text features than other baseline models, achieving an F1-score of 61.4%, which is approximately 19% higher than that of the baseline models TextCNN and BiLSTM, and 3.8% higher than that of the RoBERTa.

Key words: position classification, data augmentation, pre-training language model, multiple attention, multi-layer feature extraction