CK-16 Entailment-based consistency check (textual entailment; NLI (natural language inference) consistency; SummaC; FactCC; faithfulness checking)
Score whether each sentence of a summary or answer is supported by the source sentences, and produce a consistency score per sentence and per summary.
What it does
Scores whether each sentence of a summary or answer is supported by the source sentences. Artefact: a consistency score per sentence and per summary. Use it as triage, then read the unsupported sentences.
When to use it
Use when a model has summarised or answered from a document and you want to find the sentences the source does not support.
When it does not fit
Fits summarisation of a single document. Misfit: claims needing numerical reasoning or multi-document synthesis (not confirmed). Whether current LLM judges agree with human reviewers was not verified.
Result format
The result is a table with one row per sentence of the summary or answer, giving the consistency score and the supporting source sentence where one exists, plus a consistency score for the whole summary. It lists the unsupported sentences for a person to read.
Requirements on the material
| Property | Requirement type | Requirement |
|---|---|---|
| Content type | Always applies | requires=text |
| Unit of analysis | Always applies | requires=claim |
| Appraises source credibility | Applies in some cases | no |
Not restricted on: Structure, Variable types, Time structure, How the sample was chosen, Level of detail available, Number of sources, Whether sources can be joined, Amount of material, Known data problems, Producer interest, Source dependence, Origin, Document genre, Language, Personal information, Security classification, Licence.
| Minimum material | the source text and the generated text to be checked |
|---|---|
| Additional sources needed | none |
| Not for (rules) | none |
| Not for (text) | Wrong choice for claims that need numerical reasoning or multi-document synthesis, which was not confirmed to work. Use the score as triage and read the unsupported sentences. |
Choosing and ordering
| Kind of technique | Verification technique, runs in step 6 |
|---|---|
| Phase tag | validate |
| Other phase tags | report |
| Kind of finding | contradiction |
| Question answered | what |
| Strongest statement it supports | describe |
| Needs as input |
|
| Makes as output |
|
| Must exist first | None |
| Loop target | proceed with limit stated |
Library checks by stakes
Each stakes level includes the checks of the level below it. Checks that need a person apply at high stakes only, so the process leaves them out at lower stakes (decision 38). Repeatability: repeatable with fixed settings.
| Stakes | Check | Person involved | Counts as a different method |
|---|---|---|---|
| exploratory | spot check | optional | no |
| decision-informing | spot check | optional | no |
| decision-informing | human review sample | needed; left out below high stakes | no |
| decision-informing | source-span citation | none | no |
| high-stakes | spot check | optional | no |
| high-stakes | human review sample | needed | no |
| high-stakes | source-span citation | none | no |
| high-stakes | gold standard agreement | conditional | no |
| high-stakes | independent review | needed | no |
Failure modes and limits
| Model suitability | model plus code |
|---|---|
| Typical model failure | Accepts a plausible paraphrase that changes a number or a hedge. |
| Error sources | input data; language model |
| Certainty limits | Certainty drops for numeric or multi-document claims (indirectness) and when the checker model shares the generator's errors (risk of bias). |
Related techniques
- should follow PE-12 Model summarisation of documents (abstractive summarisation; multi-document summarisation): summaries are the usual output checked
- should follow PE-13 Retrieval-augmented analysis (RAG (retrieval-augmented generation); grounded question answering over a corpus): generated answers from retrieval are checked against passages
- pairs with CK-17 Source-span and quotation check by code (quote-and-verify; extractive grounding; citation verification): span matching confirms the passage exists and entailment confirms it supports the claim
- pairs with CK-18 Numeric and attribution recomputation by code (numeric fidelity check; deterministic recomputation): entailment models miss changed numbers
Evidence and tag basis
Tested. Entailment measures correlated better with human faithfulness ratings than standard summary metrics (2020). One entailment-based checker reached 74.4 per cent balanced accuracy on its benchmark, about 5 points above earlier work (2021), so errors remain. (Maynez et al. 2020, ACL; Laban et al. 2021, SummaC; Kryscinski et al., arXiv 1910.12840)
Tag basis: input profile inference; selection inference; sequencing inference; robustness inference. All tag values are inference unless a source is named.