TL;DR
When fact-checking LLM-generated text, breaking sentences down into isolated "atomic facts" often strips away crucial context, leading to verification errors when names or concepts are ambiguous. This paper introduces molecular facts—claims that are rewritten to be completely self-contained (decontextualized) while remaining as concise as possible (minimal). By balancing these two properties, the authors' new two-stage generation method significantly improves fact-verification accuracy in ambiguous settings without bloating the claims.
What Problem Does This Paper Address?
To verify if an LLM is "hallucinating," standard pipelines typically break long-form text down into short, discrete propositions called atomic facts. However, the authors point out a major flaw in this granularity: fully atomic facts often lack the context needed to interpret them correctly.
For instance, if an LLM generates a biography about a Swedish footballer named Ann Jansson, an extracted atomic fact might simply be: "She won a medal at the European Athletics Championships." If a fact-checker looks up "Ann Jansson," it might erroneously validate this claim against a different person—an athlete who was a racewalker, not a footballer.
Conversely, if you add too much context during decontextualization (e.g., including her exact birthdate and every team she played for), you violate minimality. This makes the claim overly specific, meaning a fact-checker might reject the entire claim simply because one minor, ancillary detail couldn't be found in the evidence document. The paper addresses this tension by defining the optimal middle ground: molecular facts.
Methods
The authors propose MOLECULAR-DECONTEXT, a two-stage prompting framework using GPT-4 Turbo to automatically generate molecular facts:
- Stage 1: Identifying Ambiguity: The model analyzes the primary subject of a claim and uses its internal parametric knowledge to determine if multiple entities share that name. If ambiguity exists, it selects a specific disambiguation criterion (such as profession, birth year, or location).
- Stage 2: Molecular Facts Generation: The model rewrites the atomic claim by integrating the original text's context alongside the chosen disambiguation criterion.
This approach is compared against three baselines:
- ATOMIC: Raw, unmodified atomic claims.
- SIMPLE-DECONTEXT: Standard decontextualization using the LLM response as context.
- SAFE-DECONTEXT: A conservative revision method designed to minimize changes.
Datasets
The methodology was evaluated using two primary datasets:
- Factcheck-Bench: A dataset consisting of 812 long-form ChatGPT responses with human-annotated factuality labels, used to analyze how over-contextualization hurts error localization.
- Ambiguous Biographies: A dataset containing LLM-generated biographies of distinct individuals who share identical or highly similar names (e.g., two different people named Dick Hanley). The authors sampled 726 human-annotated claims from this dataset.
Key Insights and Ideas
- The "Entity Switch Point" Challenge: Fact-checking systems struggle most at the exact moment an LLM switches from discussing Entity A to Entity B within the same paragraph (e.g., hallucinating details of a different person with the same name). Molecular decontextualization is vital here to signal to the verifier that the subject has shifted.
- The Pareto Frontier of Length vs. Accuracy: Modifying text generally increases sentence length. While longer, highly detailed decontextualizations can sometimes improve raw accuracy, they frequently break down during error localization because they contain too many sub-facts. Molecular facts successfully find the sweet spot on this frontier.
- Parametric Knowledge as a Filter: Instead of blindly injecting context, using an LLM's internal knowledge to first ask "Is this name ambiguous?" allows the system to apply surgical edits only where necessary.
Key Results
- Minimality vs. Bloat: In controlled experiments, standard decontextualization methods (
SIMPLE-DECONTEXT) produced non-minimal, bloated claims in 23.39% of cases, which directly harms a system's ability to localize errors. - Accuracy in Ambiguous Settings: On the Ambiguous Biographies dataset,
MOLECULAR-DECONTEXTachieved 74.7% overall accuracy, outperforming raw atomic claims (68.7%) and showing a much higher propensity to maintain claim minimality (52.0%) compared to standard simple decontextualization (16.0%). - Robustness at Switch Points: At critical entity switch points, molecular claims demonstrated superior robustness, improving verification accuracy by 3.5% over the next best decontextualization baseline.
Limitations and Future Work
- Scope of Ambiguity: The evaluation focused primarily on English-language biographies and entity name ambiguities. The authors note that other types of semantic or predicate ambiguities remain unexplored.
- Dependence on Parametric Knowledge: Stage 1 relies entirely on GPT-4's internal knowledge to flag ambiguous entities. If the model is unaware of a name twin, the disambiguation pipeline will fail.
- Pipeline Integration: The study evaluates decontextualization as an isolated step; future work is needed to test how molecular facts perform in fully end-to-end, automated retrieval and verification pipelines.
Summary updated