DAHRS: Divergence-Aware Hallucination-Remediated SRL Projection
Published in International Conference on Applications of Natural Language to Information Systems (NLDB), 2024
Recommended citation: Youm, S., Mather, B., Jayawaeera, C., Prada, J., & Dorr, B. (2024). DAHRS: Divergence-Aware Hallucination-Remediated SRL Projection. In International Conference on Applications of Natural Language to Information Systems (NLDB). https://doi.org/10.1007/978-3-031-70239-6_29
Semantic role labeling (SRL) enriches downstream applications such as translation, summarization, and question answering. However, multilingual SRL is challenging due to limited annotated data and projection errors from large language models (LLMs). This paper introduces DAHRS, a hallucination-remediated SRL projection method that corrects spurious role labels by modeling divergence-aware alignment and applying a First-Come First-Assign (FCFA) algorithm. DAHRS improves projection accuracy and robustness across language domains.