RoBERTa-BiLSTM-GAT Framework for Behavior Extraction and Case Matching from Legal and Multimodal Data

Abstract

As big data and information technology continue to develop, improving the effectiveness of behavior extraction and case matching in intelligent decision-making systems has become an urgent need. To this end, this study proposes a behavior extraction and case matching model combining multi-level feature learning and graph neural networks. Methodologically, the behavior feature extraction module is constructed using a robustly optimized Transformer encoder representation model and a bidirectional long short-term memory network. A graph attention network is introduced to optimize the topological matching mechanism between cases. The model was validated on the CaseLaw and Twitter Sentiment140 datasets. The experimental results showed that the model achieved F1 scores of 90.89% and 93.89% for the behavior extraction and case matching tasks, respectively. The average processing time for matching was as short as 0.63 seconds. Compared with advanced methods such as T5, DGI, and MANN, this model demonstrated significant advantages in terms of accuracy, recall rate, and matching efficiency. Additionally, in testing with text-image multimodal data, the proposed model achieved an average matching adjustment count of approximately 3.5, a matching throughput of up to 210 times per second, and a matching confidence score of up to 0.92. These results fully validate the superiority and practicality of this method in complex behavioral pattern analysis.