DAHRS: Divergence-Aware Hallucination-Remediated SRL Projection

2024-07-12

International Conference on Applications of Natural Language to Data Bases

AI Summary

Multilingual semantic role labeling (SRL) faces two core challenges: severe scarcity of annotated data for low-resource languages and the propensity of large-model-driven cross-lingual projection methods (XSRL) to generate uninterpretable hallucinated labels and suffer from cross-frame semantic drift. To address these, we propose a semantic divergence-aware projection refinement framework—the first to explicitly model semantic divergence in SRL—featuring a differentiable projection layer for structured correction. Our approach integrates semantic consistency modeling, adversarial divergence estimation, constraint-aware decoding, and a self-supervised hallucination discrimination module. Evaluated on the CoNLL-2005/2012 benchmarks, our method achieves absolute F1 improvements of 2.3–3.1%, reduces cross-predicate and cross-domain hallucination rates by 37%, and simultaneously enhances generalization and robustness.

Technology Category

• Natural Language Processing: Sentence-level Semantics, Textual Inference, etc.
• Machine Learning: Semi-Supervised Learning
• Knowledge Representation and Reasoning: Knowledge Representation Languages

Application Category

• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Problem

• Addresses multilingual SRL model challenges due to scarce annotated corpora.
• Reduces spurious role labels in SRL projection using LLMs.
• Improves SRL projection accuracy with divergence-aware remediation.

Innovation

• Leverages linguistically-informed alignment remediation
• Uses greedy First-Come First-Assign SRL projection
• Introduces divergence metric for language adaptation