eTracer: Towards Traceable Text Generation via Claim-Level Grounding

TL;DR

This work tackles the verification challenge of system-generated biomedical text by proposing eTracer, a plug-and-play framework that grounds each claim in a generated response against contextual evidence to produce traceable, signed evidence links. It formalizes claim-level grounding, introducing a decomposition model and an entailment model to produce claim–evidence scores, and defines metrics for grounding quality and response faithfulness. Empirical results on PubMedQA, BioASQ-QA, and TracSum show that claim-level grounding yields substantial improvements over both sentence-level grounding and end-to-end claim grounding, while enabling faster verification in user studies. The framework is designed to be easily integrated with retrieval-augmented generation pipelines and is complemented by a human-annotated ground-truth dataset to foster further research in verifiable biomedical generation.

Abstract

How can system-generated responses be efficiently verified, especially in the high-stakes biomedical domain? To address this challenge, we introduce eTracer, a plug-and-play framework that enables traceable text generation by grounding claims against contextual evidence. Through post-hoc grounding, each response claim is aligned with contextual evidence that either supports or contradicts it. Building on claim-level grounding results, eTracer not only enables users to precisely trace responses back to their contextual source but also quantifies response faithfulness, thereby enabling the verifiability and trustworthiness of generated responses. Experiments show that our claim-level grounding approach alleviates the limitations of conventional grounding methods in aligning generated statements with contextual sentence-level evidence, resulting in substantial improvements in overall grounding quality and user verification efficiency. The code and data are available at https://github.com/chubohao/eTracer.

Figures (16)

• Figure 1: eTracer transforms hard-to-verify responses into traceable ones by establishing grounding links between claims ($c_i$) and contextual sentences and evaluating responses using the rates of faithful, ambiguous, hallucinated, and unverified claims (§\ref{['sec:response_evaluation']}).
• Figure 2: eTracer Framework: identifying supportive and contradictory evidence for each response claim and aggregating grounding results to establish explicit links between response sentences and contextual sentences.
• Figure 3: Decomposition Evaluation Results.
• Figure 4: eTracer performance under different evidence threshold values, $\tau \in \{0, 0.25, 0.5, 0.75, 1\}$.
• Figure 5: Faithfulness evaluation of generated responses in PubMedQA, BioASQ-QA, and TracSum corpus.