A fact-checker that only says "refuted" leaves the user to work out why. A more useful system points to the exact phrase in the claim that the evidence fails to back up, for example the wrong year, the swapped name, or the inserted "not". This note proposes a simple way to find that phrase using compression. Take the evidence and ask a compressor to encode the claim afterwards. Everything in the claim that the evidence already "explains" becomes cheap to encode. The phrases that stay expensive are what the evidence does not account for. Those leftover expensive phrases, which we call the compression residual, serve as the explanation. They are faithful by construction, because the explanation is the very quantity the system computes, not a story generated after the fact.
The idea comes from normalized compression distance (NCD). NCD measures how similar two texts are by checking how much smaller their concatenation compresses than the two texts compressed separately. It is known as a parameter-free text classifier, and it is also known to fail at exactly what fact-checking needs. A claim and its negation share almost every byte, so NCD calls them nearly identical. Our central observation is that this failure belongs to the global score and disappears at the token level. If we split the conditional compression cost of the claim given the evidence into contributions from each token, the "not" in a negated claim is precisely the token the evidence cannot predict. The same split also exposes the wrong year, the swapped name, and the unsupported qualifier. NCD therefore should not decide whether a claim is true. It should decide where the claim breaks.
Core Claim
Let $c = c_1 \dots c_T$ be the claim tokens, $e$ the evidence text, and $L(c_t \mid \cdot) = -\log_2 p(c_t \mid \cdot)$ the code length a language-model compressor assigns to token $c_t$. This is the cost an arithmetic coder would pay, as in Neural Normalized Compression Distance and the Disconnect Between Compression and Classification. For a candidate phrase $s$ (a set of claim tokens), we define its local compression ratio as
$$\rho(s) = \frac{\sum_{t \in s} L(c_t \mid e, c_{<t})}{\sum_{t \in s} L(c_t \mid c_{<t})},$$
that is, the cost of the phrase when the evidence is visible, divided by its cost without the evidence. This is the per-span analogue of NCD's normalization. A ratio near 0 means the evidence fully explains the phrase. A ratio near 1 means the evidence contributes nothing. A ratio above 1 means the evidence actively makes the phrase less expected than it would be by default, which is the signature of contradiction.
The claim of the proposal is that ranking a claim's phrases by $\rho$ localizes culprit phrases about as well as LOREN's trained reasoning modules, without any training on fact-checking data. The compressor needs only two forward passes per claim, with no pairwise comparison against a corpus. A second claim is that the ratio-above-1 regime separates "refuted" culprits from merely "unsupported" (not-enough-information) content better than chance. We treat this second claim as an empirical question, not a promise.
Relation to Prior Work
Compression-based text methods supply the mechanism and also constrain it. "Low-Resource" Text Classification: A Parameter-Free Classification Method with Compressors showed that gzip plus nearest-neighbour classification is competitive in few-shot and out-of-distribution settings. However, it only labels whole documents, and its authors acknowledge that compression captures surface regularity rather than deep semantics. Neural Normalized Compression Distance and the Disconnect Between Compression and Classification replaced gzip with language models plus arithmetic coding. It found that better compression does not reliably mean better classification: RWKV 169M compressed AGNews to 0.248 against gzip's 0.785, yet underperformed gzip on DBpedia. We reuse its neural compressor but change what is read from it. Instead of a whole-document distance fed to a classifier, we read token-level conditional code lengths. Its negative result also shapes our design: we do not assume that a bigger compressor is better, and we test compressors of several sizes explicitly. Authorship work such as Authorship Verification with Compression Features motivates the bet that compression is sensitive to small local changes in text.
On the fact-checking side, LOREN: Logic-Regularized Reasoning for Interpretable Fact Verification is the closest work and the target to beat. It splits claims into phrases, builds a local premise for each phrase with a reading-comprehension model, and trains phrase-level veracity with soft logic rules. It reports a culprit-identification F1 of 74.3 on 100 hand-labelled refuted claims and a FEVER score of 72.93. We keep LOREN's framing (one culprit phrase per false claim, aggregated by "refuted if any phrase is refuted") and replace its trained phrase verifier with an untrained information measure.
Two other systems decompose claims but likewise depend on large fine-tuned or instruction-tuned models. QA-NatVer: Question Answering for Natural Logic-based Fact Verification asks boolean questions to Flan-T5 and reports 70.3% FEVER accuracy. Zero-Shot Fact-Checking with Semantic Triples and Knowledge Graphs breaks claims into subject-relation-object triples and checks them with natural language inference (NLI) models, which judge whether one sentence supports or contradicts another. Generation-based explanation work motivates our insistence on extractive, computed explanations. e-FEVER: Explanations and Summaries forAutomated Fact Checking found GPT-3 summaries hallucinate when evidence is missing. Generating Fact Checking Briefs found about 45% of crowdworkers relied solely on provided briefs, so an explanation that invents content does real harm. The survey Explainable Automated Fact-Checking: A Survey notes that attention weights do not make faithful rationales. This rules out the cheapest alternative explanation source.
Research Program
The first step builds the residual extractor. We compute $L(c_t \mid c_{<t})$ and $L(c_t \mid e, c_{<t})$ with off-the-shelf causal language models at three sizes, roughly 125M, 1B and 7B parameters, using a fixed prompt-free format: evidence text, a separator, then the claim. We score candidate phrases taken from LOREN's own phrase extraction, so that both systems choose among identical candidates. A byte-level LZ variant serves as the classical compressor. It marks claim bytes that cannot be matched by back-references into the evidence, which is essentially what gzip's dictionary does.
The second step studies when the residual is meaningful. We stratify FEVER claims by how they were constructed. FEVER: a Large-scale Dataset for Fact Extraction and VERification built claims from Wikipedia sentences by paraphrasing, negation, entity or relation substitution, generalization and specification. We annotate a sample by mutation type and measure localization within each type. The prediction is that neural and LZ residuals tie on entity substitution, where the wrong name is simply absent from the evidence. The neural residual should win on paraphrase and specification, where the culprit is worded differently from anything in the evidence. Negation is the critical case for Gap 4: global NCD is near zero there, yet the local residual should land on the negator.
The third step uses the residual for verdicts in a deliberately limited way. Supported claims should have low maximum $\rho$. For non-supported claims, we pass only the highest-residual phrase and its evidence to a small NLI model, and the logic rule decides between refuted and not-enough-information. This tests whether a compression-chosen focus helps a weak verifier. NCD itself never makes the stance decision.
The fourth step extends to multi-hop claims on EX-FEVER: A Dataset for Multi-hop Explainable Fact Verification. We add evidence hops one at a time and record which claim spans each hop "pays for". This gives an explanation of which document supports which part of the claim, which current BART-generated explanations do not provide, together with a direct test of whether hops are complementary or redundant.
Baselines
The naive reference is lexical non-overlap. It marks as culprit the candidate phrase with the fewest lemmatized tokens present in the evidence, breaking ties at random, with a uniformly random phrase as a floor. This simple heuristic will solve most entity substitutions, so any compression method must beat it on the paraphrase and negation strata, not merely on average.
The key ablation is the LZ/gzip residual. It is the same procedure with a dictionary compressor instead of a language model, and it isolates whether the learned model's predictions, rather than conditional coding as such, carry the signal. We expect it to behave almost identically to lexical non-overlap. If the neural residual does not beat the LZ residual on paraphrase-type claims, then the whole contribution reduces to string matching, and we would report it that way.
The strongest existing method is LOREN, with its reported culprit F1 of 74.3 on 100 manually labelled refuted claims (LOREN). We will request LOREN's labelled set. If it is unavailable, we will rerun the released LOREN model on our own culprit-annotated set of several hundred refuted FEVER claims, labelled by two annotators blind to system outputs. Both systems receive the same gold evidence, then the same retrieved evidence, and the same phrase candidates. LOREN's 74.3 is only quoted for context. The comparison that counts is our rerun on shared data.
The second strong competitor is perturbation saliency with an NLI model. Each candidate phrase is deleted or masked in turn, and the culprit is the phrase whose removal most reduces the contradiction probability of a RoBERTa-large NLI model trained on MultiNLI. This is the kind of model used in e-FEVER and LOREN. It gets an inference budget equal to or larger than ours. A prompted instruction-tuned language model (Flan-T5, as in QA-NatVer) asked to quote the wrong phrase completes the set.
Beating the field means matching or exceeding rerun-LOREN culprit F1 within its confidence interval while using no fact-checking training. It also means strictly beating the LZ and lexical baselines on the paraphrase and negation strata.
Why not just…?
Why not just mark claim words that do not appear in the evidence? For substituted entities this probably works as well as anything, and we expect a tie there. It fails whenever the culprit is expressed in words that do appear in the evidence: negation ("did not win" against "won"), changed quantifiers, or relations reworded across a paraphrase. The stratified evaluation is designed to show where lexical overlap and the neural residual diverge. If they never diverge, the proposal's claim is false.
Why not just mask each phrase and watch an NLI model's verdict change? This is the most serious threat, because it uses a trained entailment model that does understand direction. It has two weaknesses. First, it costs one NLI pass per phrase. Second, its saliency inherits the dataset-artifact reliance that Zero-Shot Fact-Checking with Semantic Triples and Knowledge Graphs documents on FEVER-Symmetric and FEVER 2.0, where models key on claim-only cues. We will compare both methods on FEVER-Symmetric, where claim-only shortcuts are removed. The prediction is that the compression residual degrades less, because it is computed only from evidence-conditioned predictability. If NLI saliency wins there too, we would still report the residual as a cheap, training-free approximation, not as the better method.
Why not just ask a large language model which phrase is wrong? Prompted models produce fluent answers, but the evidence in the file says they fabricate when support is absent (e-FEVER), and that users over-trust such output (Generating Fact Checking Briefs). The residual can only point at spans that actually occur in the claim, and its score is reproducible. We will run a prompted Flan-T5 and a larger instruction-tuned model as baselines and measure how often each names a span that is not one of the candidates or not in the claim.
Proposed Experiments
The primary experiment measures culprit-localization F1 on refuted FEVER claims under gold evidence, reported overall and by mutation type. The decisive comparisons are neural residual against LZ residual on the paraphrase and negation strata, and neural residual against rerun-LOREN overall. A compressor-size sweep (125M, 1B, 7B) tests whether the compression–classification disconnect from Neural NCD reappears: we will report compression rate and localization F1 side by side.
The second experiment repeats localization with retrieved rather than gold evidence. This matters because retrieval is the main failure point: FEVER's baseline evidence recall was 45.89%, and pipeline accuracy fell from 50.91% to 31.87% once correct evidence was required. Here the residual should spread across the whole claim when the evidence is missing, which we will measure as a not-enough-information signal.
The third experiment is the verdict test. It compares FEVER label accuracy for a small NLI model given the full claim against the same model given only the top-residual phrase. QA-NatVer's 70.3% with Flan-T5 serves as a reference point, not as a target. On FEVER-Symmetric, no reported number was found on Lacuna for QA-NatVer or the triples system, so we rerun the baselines.
The fourth experiment is a faithfulness check. We delete the evidence sentence that a gold annotation links to a given claim span and verify that the residual moves to that span; the success rate is our faithfulness score.
The fifth experiment applies hop-wise residual attribution on EX-FEVER, scored against the human-written golden explanations by whether each hop is credited with the claim spans that explanation assigns to it.
A cross-lingual check on DanFEVER, using a multilingual compressor and no translation, closes the study. No reported result was found on Lacuna for DanFEVER culprit localization, so it is purely exploratory.
References
- Neural Normalized Compression Distance and the Disconnect Between Compression and Classification
- "Low-Resource" Text Classification: A Parameter-Free Classification Method with Compressors
- Authorship Verification with Compression Features
- LOREN: Logic-Regularized Reasoning for Interpretable Fact Verification
- QA-NatVer: Question Answering for Natural Logic-based Fact Verification
- Zero-Shot Fact-Checking with Semantic Triples and Knowledge Graphs
- e-FEVER: Explanations and Summaries forAutomated Fact Checking
- Generating Fact Checking Briefs
- Explainable Automated Fact-Checking: A Survey
- FEVER: a Large-scale Dataset for Fact Extraction and VERification
- EX-FEVER: A Dataset for Multi-hop Explainable Fact Verification