Answers are grounded in this paper and cite their sources.
Cross-lingual legal assistants in India process user queries in both English and Marathi while adhering to deployment speed requirements. A single model for cross-lingual and lightweight matching of legal question pairs is proposed, which consists of three modules: (i) contextualized semantics via a multilingual Transformer encoder, (ii) language-independent legal anchoring signals in the form of statutes/sections/articles in Latin and Devanagari scripts, and (iii) a sentence-embedding branch learned via knowledge distillation from teacher models. The model is trained using a joint classification and ranking loss in combination with a citation format invariance regularizer. The model’s performance is tested using a deployment-faithful protocol in which the decision threshold is tuned on English validation and then kept fixed for cross-lingual evaluation. On the English benchmark set comprising 1,233 pairs, the model achieves an F1 score of 0.964 and an AUC score of 0.996 in 1.24 ms per pair. An LLM classifier achieves a higher F1 score of 0.986, but at a slower rate of 14.6 ms per pair. For the cross-lingual setting where a fixed English-selected threshold is used, the approach achieves an F1 value of 0.935 for English and 0.852 for Marathi (ΔF1 = 0.083), while keeping a consistent rank quality (MRR ≈ 0.82). Given a small amount of labeled samples of Marathi, the F1-score for Marathi rises to 0.867, with ΔF1 equaling 0.055. This result implies that the use of multilingual semantics along with light-weight legal anchoring leads to solid cross-lingual performance in real-world legal information retrieval and question-answering applications.
No takes yet. Share an insight, caveat, or question.
Pathak et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: