Abstract:

Beet is an important sugar crop in China，and its breeding literatures contain rich structured knowledge such as varieties，traits，and genes. However，the scarcity of annotated data and the complexity of terminology limit the performance of existing entity recognition methods. To address this problem，we constructed the first dedicated dataset for named entity recognition（NER）in beet breeding in China，defining eight core entity types. Through a strategy combining small-scale manual annotation with rule-based template augmentation，the strategy effectively expanded the training samples. Using this dataset，the BERT-BiLSTM-CRF model achieved an F1 value of 90.19%，outperforming baseline model. To validate the practicality of the proposed method，we applied the model to recognize entities in 502 literatures and built a knowledge graph that centered on the literature and supports traceability. The graph's coverage and retrieval accuracy were evaluated through structured-retrieval tasks. Experimental results showed that the normalized retrieval accuracy based on the graph reached 90.8%，confirming the usability of the knowledge graph in the domain of beet breeding. This study provides an effective methodological reference for entity recognition and knowledge organization in low-resource scenarios.

Key words: beet breeding, named entity recognition, BERT-BiLSTM-CRF model, knowledge graph, low-resource text mining