Shun Wang

Knowledge extraction serves as a critical technique for constructing domain-specific knowledge graphs and enabling intelligent text parsing. In the vertical agricultural domain, wheat variety approval texts contain abundant agronomic trait entities and their relational associations. Nevertheless, they are characterized by dense domain-specific terminology, the co-occurrence of multiple relations, long entity spans, and mixed numerical and textual expressions, which pose considerable challenges to the accurate extraction of entity-relation triples. Existing in-context learning paradigms mostly retrieve demonstrations based on general-purpose semantic embeddings. The retrieved demonstrations rarely achieve relation-structure matching with query instances; furthermore, LLMs are susceptible to format hallucinations, which limit entity and relation extraction performance. Accordingly, we propose Relation-aware Demonstration Retrieval and Progressive Knowledge Calibration for In-Context Learning (ReCal-ICL), a relation-aware in-context learning framework for joint entity and relation extraction. ReCal-ICL employs RADS, a relation-aware demonstration selector based on hierarchical contrastive learning. Through hierarchical quadruples, in-batch dynamic negative sampling, and a hybrid loss, RADS reshapes the representation space to enhance relational-structure matching between retrieved demonstrations and query instances. In addition, a two-layer output constraint strategy is developed, which integrates prompt-template prior constraints and post hoc result verification mechanisms to jointly guarantee the format compliance of extraction outputs. We conduct comparative and ablation experiments on WVAD as well as public NYT and CoNLL04 datasets. Experimental results show that ReCal-ICL achieves competitive joint extraction performance across multiple datasets. Further experiments show that incorporating demonstrations benefits downstream extraction, while the retrieval relevance evaluation indicates that RADS improves the relational relevance of retrieved demonstrations.

The certification information of wheat varieties contains valuable breeding knowledge and plays an important role in germplasm resource management and breeding research. However, most certification information is stored in unstructured text form, making it difficult to support efficient knowledge acquisition and utilization. To address this issue, this paper investigates joint entity-relation extraction from wheat variety certification texts and its application to knowledge graph construction. Specifically, building on existing character-word fusion methods, we propose the Joint Entity-Relation Extraction for Wheat Variety Certification Texts (JERE-WVCT) to address indiscriminate incorporation of candidate features in character-word fusion that obscures key features, a lack of differentiated weight assignment for features, and severe imbalance across relation types. Within JERE-WVCT, a deep character-word fusion mechanism based on hierarchical filtering and ranking is designed to enhance the representation of domain-specific entities. In addition, relation labels are incorporated as prior knowledge to alleviate the impact of relation type imbalance and improve the model’s triple extraction capability. Experimental results show that JERE-WVCT achieves an F1 score of 96.78% on the wheat variety certification corpus, outperforming all baseline models. Based on the extracted triples, a wheat variety knowledge graph is constructed, and exploratory variety relationship analysis is conducted as a downstream application. The results demonstrate the effectiveness of the proposed model for structured knowledge acquisition and support graph-based exploration of wheat variety information.

Chinese social media cyberbullying detection faces three major challenges: implicit aggressive semantics, diverse adversarial expression patterns, and severe long-tailed class imbalance. To address these issues, this paper proposes a Hierarchical Multi-Committee Ensemble (HMCE) framework for Chinese cyberbullying detection. At the representation level, a hybrid architecture integrating pre-trained language models, multi-scale TextCNN, and attention pooling is designed to capture both global contextual dependencies and fine-grained local N-gram offensive cues. At the decision level, a three-tier progressive ensemble architecture is constructed. Specifically, the Tier-0 Committee Ensemble Layer introduces heterogeneous committees with differentiated training strategies to generate diverse Out-of-Fold (OOF) probability features. The Tier-1 Aggregation Layer employs an XGBoost-based meta-learner to model nonlinear complementary relationships among committees and improve decision fusion capability. Furthermore, the Tier-2 Calibration Layer incorporates probability calibration and adaptive threshold optimization to alleviate prediction bias under long-tailed distributions and improve minority-class recognition performance. Experiments conducted on a real-world dataset containing 23,080 Chinese social media posts demonstrate that the proposed HMCE framework achieves a Macro-F1 score of 90.26% and an accuracy of 92.50%, outperforming conventional pre-trained language models and existing ensemble approaches. The results validate the effectiveness, robustness, and generalization performance of HMCE for cyberbullying detection in complex Chinese linguistic environments.