Formulaic language identification model based on GCN fusing associated information.

Area of Science:

• Computational Linguistics
• Natural Language Processing
• Artificial Intelligence

Background:

• Formulaic language, comprising ready-made linguistic structures, is crucial for coherent writing and communication.
• Accurate identification of novel formulaic language is an ongoing research challenge.
• Existing methods may not fully capture the complex relationships within formulaic language.

Purpose of the Study:

• To propose and evaluate a new model for identifying formulaic language.
• To leverage graph convolutional networks (GCN) for enhanced feature extraction.
• To improve the accuracy, recall, and F1-score of formulaic language identification.

Main Methods:

• Constructing sentences as graphs where nodes represent word features (part-of-speech, semantic) and edges represent mutual information and dependency relations.
• Applying a graph convolutional neural network (GCN) to extract deep grammatical features from these graph representations.
• Fusing associated information within the GCN framework.

Main Results:

• The proposed GCN-based model demonstrates superior performance compared to classical formulaic language identification models.
• The model achieved higher accuracy, recall, and F1-score in experiments.
• The approach effectively mines deeper grammatical features by analyzing word associations.

Conclusions:

• The developed GCN model offers a more accurate method for identifying formulaic language.
• This research provides a foundation for future advancements in formulaic language identification tasks.
• The fusion of associated information in a graph structure is key to the model's success.