How to measure whether a cited source supports the answer

A system that answers from manuals or doctrine can attach a citation to each claim, and the citation helps only if the cited page supports the claim. Research tests this with one rule: a cited source supports a statement when a reader who sees only that source would accept the statement. Evaluations turn the rule into two scores, one for how many statements are backed by their citations and one for how many citations back the statement they are attached to.

The attribution test

Rashkin and colleagues proposed Attributable to Identified Sources (AIS) as a way to judge generated text against its sources. The test asks whether a generic hearer would affirm that, according to the identified source passage, the statement holds.

Annotation runs in two stages. In the first, annotators see the output without the source and decide whether it can be understood on its own. In the second, only interpretable outputs continue: annotators see the source and decide whether it supports the output. They judge the entire output at once rather than individual sentences or spans.

AIS measures support, not truth. The authors state that they avoid absolute judgments about factuality, so a statement can be fully attributable to a source that is itself wrong. The score says whether the cited page backs the answer and says nothing about whether the page is right.

Two scores for two failures

Liu, Zhang, and Liang defined the two measures in an audit of generative search engines. Citation recall is the proportion of verification-worthy statements that are fully supported by their citations. Citation precision is the proportion of generated citations that support their statements.

Two measures are needed because they catch opposite failures. A response can make claims with no citation, which lowers recall, or attach citations that do not back the claim, which lowers precision. In the audit, human evaluators rated four popular engines, and on average 51.5 percent of generated sentences were fully supported by their citations while 74.5 percent of citations supported their sentence.

The audit's precision measure includes a partial category. A citation that supports only part of a statement counts toward precision only when the full statement is supported by the union of its citations.

Automating the check

Gao, Yen, Yu, and Chen released ALCE, which they describe as the first benchmark for automatic citation evaluation of large language models (LLMs). It pairs questions with document collections, so a system must retrieve evidence and cite it. One of its datasets, ELI5, is built on the Reddit forum Explain Like I'm Five.

ALCE scores citations with a natural language inference (NLI) model, which judges whether a set of passages entails a statement. The model is TRUE, a T5-11B model fine-tuned on a collection of NLI datasets.

• Citation recall is scored per statement. A statement scores 1 only if it has at least one citation and the NLI model judges that the concatenated cited passages entail it.
• Citation precision is scored per citation. A citation is irrelevant when it does not support the statement alone and the statement's other citations still support it without this one. A citation scores 1 only if its statement has recall 1 and the citation is not irrelevant.

The ALCE abstract reports that on ELI5 even the best models lack complete citation support 50 percent of the time. Bohnet and colleagues, working on attributed question answering, use human annotations as the gold standard and report that an automatic metric correlates with human judgments, which is the premise behind this kind of scoring.

Where automatic scoring falls short

ALCE's authors compared the automatic scores with human labels on 100 randomly sampled examples from two of its datasets, ASQA and ELI5. Cohen's kappa between the human labels and the automatic ones was 0.698 for citation recall, which they call substantial agreement. For citation precision it was 0.525, which they call moderate.

The authors also list a limitation: the NLI model cannot detect partial support. The audit paper counts partial support under a stated condition and the automatic benchmark cannot detect it, so precision figures from the two designs should not be treated as interchangeable.

Two more limits recur across the papers. Liu and colleagues score at the sentence level for simplicity and note that the scope of a single in-line citation is often ambiguous, while AIS judges whole outputs. And every automatic score above is validated against human annotators, so the automatic number inherits their definition of support.

A reported citation score is defined by four choices: the unit of a statement, the judge (human or NLI model), the source collection, and the treatment of partial support. A score reported without those four choices cannot be compared with another score.

What to ask

• What counts as a statement: a sentence, a whole answer, or something else?
• Who judged support, human annotators or an NLI model, and how closely did the two agree on this material?
• Were the cited sources drawn from the same collection the system will answer from in service?
• How is a citation that supports only part of a statement counted?
• Are citation recall and citation precision both reported?

Azcuna's BABEL reads, translates, and searches foreign-language documents entirely on-device, with no cloud and no connection, and keeps every claim tied to the passage it came from. The long read on air-gapped RAG goes further.

Sources

1. 1Measuring Attribution in Natural Language Generation Models · arXiv (Rashkin et al.)
2. 2Evaluating Verifiability in Generative Search Engines · arXiv (Liu, Zhang, and Liang)
3. 3Enabling Large Language Models to Generate Text with Citations · arXiv (Gao, Yen, Yu, and Chen; EMNLP 2023)
4. 4Attributed Question Answering: Evaluation and Modeling for Attributed Large Language Models · arXiv (Bohnet et al.)