Summary
Financial rules change often and are very detailed, making it hard to answer compliance questions correctly. The authors tested if combining smart searching of relevant text with smaller language models can help get better answers. They improved how well the models find important rules and how accurately they answer questions about a Dubai financial rulebook. However, these improvements didn’t work well for Australian legal questions, and sometimes the models gave confident but unsupported answers. This means finding relevant text helps, but ensuring answers are truly grounded in rules still needs better evaluation.
What this means in practice
- •For compliance teams at financial institutions: Produce more accurate answers to complex financial regulation questions by combining domain-tuned retrieval with efficient language models.
- •For legal tech developers: Build financial rulebook question-answering tools that improve retrieval of relevant passages using LegalBERT adaptations and fusion approaches.
Tested on one dataset.
Authors
Tobias Deußer, Abhishek Pillai, Aurelio F. Bariviera, Dhananjay Bhardwaj, Lorenz Sparrenberg, David Berghaus, Christian Bauckhage, Rafet Sifa
Abstract
Financial institutions operate under dense, frequently amended rulebooks, and answering a compliance question correctly requires not only fluency but verifiable grounding in the authoritative text. Large language models are attractive for this task, yet the models that firms can realistically deploy on-premise are compact ones, and compact models hallucinate obligations. We study whether a carefully domain-adapted retrieval-augmented generation pipeline closes that gap. Our retriever is built in three stages on top of LegalBERT: entailment tuning that recasts question--passage matching as premise--hypothesis reconstruction, contrastive tuning with in-batch negatives, and score-level fusion with BM25. Our generator is a compact model (2B--12B parameters) served under 4-bit quantization, either prompted or adapted with retrieval-aware fine-tuning (RAFT) through LoRA. On ObliQA, a question-answering benchmark built from the Abu Dhabi Global Market rulebooks, the staged retriever raises Recall@10 from 0.256 to 0.774 and outperforms BM25 (0.678) and E5-large-v2 (0.758), the strongest general-purpose dense encoder we tested. RAFT-LoRA then improves the composite RePASs answer-quality score for every model we could adapt, with the largest gain on the weakest one. However, the adapted models do not transfer to Australian case-law questions, and a closed-book model that receives no passages at all scores within 0.011 RePASs of the full pipeline while producing answers that cite nothing and misstate obligations. The retrieval gain is therefore measured directly, the generation gain is a gain in RePASs rather than demonstrated grounding, and grounding itself requires an evaluation protocol that RePASs does not provide.