U4RASD at KnowledgeGraphEval-2026: Ontology-Constrained Encoder and LLM Systems for Arabic Relation Extraction

Overview of the U4RASD Submission

The U4RASD team submitted a comprehensive system to KnowledgeGraphEval-2026 Subtask 2, focusing on directed Arabic relation extraction. The task involves predicting one of 40 relations or no_relation from an Arabic sentence and a subject-object pair. The team explored both encoder-based and large language model (LLM) approaches, highlighting the importance of constrained prediction and complementary modeling strategies in Arabic relation extraction.

Encoder-Based Systems

The encoder systems developed by the U4RASD team utilized typed entity markers and separate relation-existence and relation-classification heads. These components are designed to improve the model's ability to distinguish between relevant and irrelevant relations in the context of Arabic sentences. The use of typed entity markers helps in identifying the specific entities involved in the relation, while the separate heads allow for more precise classification of the relation type.

Large Language Model Approaches

The LLM systems were based on fine-tuning Yehia and Qwen3.5 models using QLoRA and ontology-constrained prediction. This approach ensures that the models adhere to the predefined set of relations, reducing the likelihood of generating incorrect or irrelevant relations. The team's best system was an ensemble of Yehia-RE, Qwen-RE, and TypedPair, which achieved a micro-F1 score of 94.06 and ranked first in Subtask 2.

Performance and Key Insights

The results from the U4RASD submission highlight the effectiveness of constrained prediction and complementary modeling approaches in Arabic relation extraction. The team's best system outperformed other models, including the best encoder and Qwen-RE, which achieved scores of 92.69 and 91.94, respectively. The use of an ensemble approach, combining different models, appears to be a key factor in achieving high performance on this task.

Conclusion

The U4RASD team's submission to KnowledgeGraphEval-2026 Subtask 2 demonstrates the potential of encoder-based and large language model approaches in Arabic relation extraction. The use of constrained prediction and complementary modeling strategies has proven to be effective in improving the accuracy of relation extraction tasks. The results from this submission provide valuable insights into the development of future systems for Arabic relation extraction.

Source: arXiv NLP. Analysis based on publicly available information; no hands-on testing is claimed.