Research on Behavior Extraction Algorithm Based on RoBERTa-BiLSTM-GCN

Behavior extraction aims to extract structured behavior information from a large amount of unstructured texts. However, existing research work has problems in accurately extracting behaviors and insufficient ability to identify predicates. To address the issues of inadequate semantic representation and incomplete behavior extraction in behavior extraction, a behavior extraction algorithm based on RoBERTa-BiLSTM-GCN is proposed. Firstly, the word vectors are constructed through the RoBERTa pre-trained model. Secondly, the self-attention mechanism is used to capture the associations between different word features. The bidirectional long short-term memory (BiLSTM) network and convolutional neural network are then employed to extract global syntactic features and local phrase features. Finally, BIO sequence tagging is achieved through the conditional random field to complete the behavior extraction. This algorithm is trained and tested on the CoNLL-2012 dataset. Compared with other methods, it has an optimal precision rate that is 3.27 percentage points higher, a recall rate that is 1.92 percentage points higher, and an F1 score that is 2.60 percentage points higher. This demonstrates that the algorithm can effectively extract the dependencies between text sequence data and syntactic graph structure information, and has considerable behavior extraction capabilities.