The authors adapt the open-source IndicTrans2-1B translation system to handle conversational register across 21 Indic languages using only public datasets. By combining experience replay with model souping, they achieve significant improvements in automatic metrics without degrading performance on general domain tasks.
- The adaptation uses OpenSubtitles, BPCC-H-Daily, and Tatoeba data to fine-tune the model for conversational input.
- Experience replay mixes general data back into training to prevent catastrophic forgetting of the general domain.
- Model souping averages the fine-tuned weights with the base model weights to balance performance across domains.
- The resulting model beats IndicTrans2-1B on conversational chrF in all 21 languages with a mean gain of +6.2.
- Performance on the FLORES general domain benchmark remains stable, with a mean change of -0.17 chrF.
The study demonstrates that these techniques allow for effective register matching to references, though human evaluation did not confirm perceived quality improvements.