On September 4 and 7, 2026, arXiv successively released LexFlip, a tool for diagnosing the dissociation of legal significance indicators. This tool based on 373 minimal perturbation tests of Quebec French regulations found that these perturbations reversed legal significance while only changing 0.93 elements. The test results showed that seven embedding models and metrics such as BERTScore achieved extremely low scores for such edits (0.022 to 0.039), far below the scores for bidirectional natural language inference (0.670). In the FrJudge benchmark, only length features performed better than all semantic metrics, with the smallest gap from the human measurement upper limit (r=0.597).
Integrated timelineUNIFIED TIMELINE
-
2026-09-04
LexFlip: A Dissociation Diagnostic for …
LexFlip: A Dissociation Diagnostic for Legal Meaning Preservation Metrics
-
2026-09-07
LexFlip: A Dissociation Diagnostic for …
arXiv:2609.05296v1 发布 LexFlip,一种用于法律意义保留指标的解离诊断工具。该工具包含 373 个对魁北克法语法规的最小扰动,这些扰动在保留 0.93 个词元的同时逆转了法律效力。测试显示,七种嵌入和 BERTSc…
SignalsSIGNALS
All reports (2)SOURCES
LexFlip: A Dissociation Diagnostic for Legal Meaning Preservation Metrics
arXiv:2609.05296v1 发布 LexFlip,一种用于法律意义保留指标的解离诊断工具。该工具包含 373 个对魁北克法语法规的最小扰动,这些扰动在保留 0.93 个词元的同时逆转了法律效力。测试显示,七种嵌入和 BERTScore 指标在此类编辑上的得分仅占其相同 - 无关范围值的 0.022 至 0.039,远低于双向自然语言推断(0.670)。在 FrJudge 基准上,与人类测量上限 r=0.597 相比,仅长度特征的表现优于所有语义指标且差距最小。