This research aims to extract structured knowledge from bibliographic summaries of ancient Chinese agricultural books while addressing data imbalance issues. Through the recognition of structural functions, this approach facilitates accurate comprehension of summary content.
The study develops a framework for categorizing structural functions of summaries into six components: title information, responsible person information, book formation process, edition information, document content and evaluative information. To address the few-shot learning challenge in Chinese agricultural bibliographic summaries, the study implements data augmentation for underrepresented categories. Four methods are employed to resolve class imbalance: synonym replacement, random swapping, random deletion and back translation. Large language models, such as ChatGPT, are utilized to augment samples in minority classes, with t-SNE visualization employed to validate the augmentation’s effectiveness. Then BERT-BiLSTM model is implemented for structural function recognition in agricultural bibliographic summaries.
Data enhancement based on large language models can effectively extend the original samples and retain the semantic relationship of the original samples. On this basis, the structural function recognition experiment using BERT-BiLSTM model can identify the structural function of the Chinese agricultural bibliographic summaries more accurately.
This research presents an innovative method for structuring Chinese agricultural bibliographic summaries and pioneers the application of large language models in this domain. It addresses the challenge of making historical Chinese agricultural bibliographic summaries more accessible and transmittable.
