Anchor-Preserving Claim Normalization for Time-Sensitive Social Media Posts

When a fact-checking system reads a post like "lol the PM just said schools stay open next week 🙄", it first has to turn it into a checkable sentence. That rewriting step is called claim normalization. Current normalizers produce clean, fluent claims, but they tend to drop exactly the information whose truth depends on time: who "the PM" was on the day of posting, and which calendar week "next week" means. Once those anchors are gone, no downstream verifier can recover them. We propose to borrow the structure of statistical post-editing, an older translation design in which a rule-based system writes a structurally correct but clumsy draft and a learned model then smooths it. Here, a deterministic first stage resolves relative dates against the post's timestamp and links role mentions ("the PM", "Twitter's CEO") to the specific person who held that role at that time. A learned editor then rewrites this draft into a fluent claim, but it may not change the anchors the draft has fixed.

The opportunity comes from two findings that have never been connected. Temporal verifiers such as ChronoFact reach high accuracy only when claims come with explicit events and dates. Normalizers for informal text are judged on overlap with reference claims, and that metric barely penalizes a lost date or a wrongly resolved officeholder. Post-editing research has already measured the characteristic failure of the learned second stage: it breaks things the first stage got right. We treat that failure as the thing to control and measure. The test of the idea is whether informal posts, once normalized this way, can be verified nearly as well as clean temporal claims.

Core Claim

Claim normalization should be framed as constrained post-editing of a time-resolved, entity-linked draft, not as free summarization. Concretely, we claim three things. First, current normalizers lose or corrupt a substantial fraction of time-sensitive anchors. Second, a draft-then-edit pipeline with time-indexed entity linking preserves them without giving up fluency. Third, preserving anchors produces measurable gains in downstream temporal verification that fluency and overlap metrics do not predict.

The closest existing systems each handle a different piece of the problem. From Chaos to Clarity formalizes normalization and adds check-worthiness-aware chain-of-thought prompting (CACN), but it never links entities or resolves time. AKCIT-FN shows that fine-tuned small seq2seq models beat LLM prompting when supervision exists, and it too ignores temporal validity. Neural Machine Translation for Fact-checking Temporal Claims translates claims into formal temporal logic, but only for clean templated claims and with no entity linking. Our proposal differs from all three in one respect: it makes the timestamp-resolved, time-indexed entity state an explicit, protected intermediate representation, and it measures whether the rewriter keeps it intact.

Research Program

The draft stage is deliberately symbolic. It plays the role that SYSTRAN played in Statistical Post-Editing on SYSTRAN's Rule-Based Translation System, where the rule engine fixed structure and a statistical editor trained on (draft, human reference) pairs fixed wording. The draft stage has three parts.

The first part resolves temporal expressions ("yesterday", "last season", "2moro") into absolute dates or intervals relative to the post timestamp. It uses a small learned tagger for spotting these expressions, because Implicit Temporal Reasoning for Evidence-Based Fact-Checking reports that off-the-shelf HeidelTime propagates errors on noisy text. The conversion from a detected phrase to a date stays rule-based.

The second part generates candidate entities for noisy mentions. It combines Wikipedia anchor-text dictionaries, which Dutta et al. showed can bridge web surface strings to clean ontologies, with fuzzy string matching of the kind INSIGHTBUDDY-AI used for clinical text. It also uses the author's own recent posts as disambiguation context, following the "one sense per tweeter" finding of Gella et al..

The third part is what makes the linking time-indexed. For role and office mentions, it resolves the Wikidata entity whose start and end qualifiers contain the post date. This addresses the static-snapshot limitation noted across the entity linking literature, and it matches the observation in MultiFC that its 25,000 Wikipedia links were fixed rather than time-aware.

The output of the draft stage is a skeleton claim containing protected spans, each tagged with an entity identifier or a resolved date. Each span also carries a confidence score.

The editor stage is a fine-tuned T5/BART-class model, the model family that won in AKCIT-FN. It takes the raw post and the skeleton as input and is trained on pairs of skeletons and gold fact-checker claims, directly mirroring the post-editing training recipe. High-confidence protected spans are enforced by constrained decoding. Low-confidence spans are passed as soft hints, so the editor can override a draft that is likely wrong. This is our response to the error-cascade problem that post-editing work identified.

After generation, a gate re-runs the linker and temporal resolver on the edited claim. If a high-confidence anchor has changed, the gate falls back to a minimally edited draft. The gate operationalizes the lesson from Dugast et al. that unconstrained editors degrade correct drafts. It also lets us report an explicit ratio of improvements to degradations, in the spirit of their 59-versus-7 manual count.

The analyses have two parts. The first is an error taxonomy of what current normalizers do to anchors: drop them, keep them unresolved ("next week"), resolve them to the wrong interval, or resolve a role to the wrong holder. The second is a study of which errors actually change downstream verification labels.

Baselines

The naive reference point is verifying the raw post without normalization. MultiFC shows why this is weak: its claim-only baseline scores 25.3% Macro F1, and its best evidence-plus-metadata model scores 49.2%. We will report raw-post verification on every downstream task, using the same retriever and verifier as all other systems.

The strongest normalization baseline on overlap metrics is CACN from From Chaos to Clarity, which reports a BERTScore of 89.00 on CLAN. The strongest supervised recipe is the fine-tuned seq2seq approach of AKCIT-FN, which reports METEOR 0.5290 on Portuguese in CheckThat! 2025 Task 2. We will reimplement both CACN-style prompting and a fine-tuned small model on identical splits, with the same backbone size as our editor and the same hyperparameter search budget. We will also run a timestamp-aware variant of each, in which the post date is added to the prompt or input, so our comparison is not against a strawman that lacks the date.

"Beating" these baselines is defined in two ways. On overlap metrics, the goal is non-inferiority: no more than one BERTScore point below CACN. On anchor fidelity and downstream temporal verification, the goal is a clear gain.

For downstream temporal verification, the strongest existing method is ChronoFact, which reports 85.49 macro F1 on clean ChronoClaims against 61.93 for zero-shot GPT-4o. We use ChronoFact as a fixed verifier. The question is how much of that clean-input score survives when the input is an informal post passed through each normalizer.

Two ablations isolate the key ingredient. The first, "append anchors," runs our draft stage but simply concatenates the resolved anchors to an unconstrained normalizer's output. The second, "static linking," replaces time-indexed resolution with ordinary snapshot linking. If "append anchors" matches the full system, then the constrained editor is unnecessary. If "static linking" matches it, then time indexing is unnecessary.

Why not just…?

Why not just prompt a frontier LLM with the post date and ask it to keep names and dates? This is the most serious threat, and we include it as a baseline rather than argue against it. There is some prior evidence that it will fall short. On timeline reasoning, zero-shot GPT-4o scored 61.93 against ChronoFact's 85.49 on ChronoClaims (ChronoFact), and LLMs do not have reliable access to who held an office on an arbitrary date. The specific evidence will come from our anchor-fidelity evaluation, broken down by anchor type. If a date-prompted LLM matches our pipeline on role-to-holder resolution, we will narrow the contribution to the measurement protocol and the gate.

Why not just fine-tune the normalizer with the timestamp as an extra input? This is cheap, and it is what AKCIT-FN's recipe becomes with one extra feature. However, it leaves the model to learn date arithmetic and officeholder lookup from roughly 6,000 CLAN pairs, most of which are not time-sensitive. We expect it to handle "yesterday" but fail on roles and on intervals. The timestamp-aware baseline above tests exactly this.

Why not just normalize first and link and date-resolve the output afterward? Once "next week" or "the PM" has been paraphrased away, it cannot be recovered afterward. The input-side version of this idea is the "append anchors" ablation. That ablation is a genuine competitor. It separates the value of the draft from the value of constraining the editor, and we will report it honestly even if it wins on some metrics.

Proposed Experiments

We will build a time-sensitive evaluation set in two ways. First, we will annotate 600 CLAN posts (and, where timestamps can be recovered, CheckThat! 2025 Task 2 English posts) with their post date and gold anchors: resolved dates and time-indexed entity IDs. Fact-checker claims often spell these out, which makes annotation tractable. Second, we will create an "informalized" version of ChronoClaims and T-FEVER. Human writers, with LLM assistance, will rewrite clean claims as posts with relative dates and role mentions relative to a sampled post date. A 300-item subset will be checked by hand.

The primary metrics are anchor fidelity, defined as the fraction of gold anchors correctly present in the output after re-linking, and the harm rate, defined as correct draft anchors destroyed by the editor. We also report BERTScore and METEOR against the published CACN and AKCIT-FN numbers, and human ratings on the CACN scales.

Downstream, we will run ChronoFact on each system's normalized output for informalized ChronoClaims, and report T-FEVER label accuracy against the TACV figure of 52.15% from Time Matters. We will also run the time-bucket verifier of Allein et al. on MultiFC claims, whose reported score is 47.43% Macro F1. CLAN retrieval precision at k is also of interest, but From Chaos to Clarity reports only that it was "higher," so no number is available there for comparison.

The decisive check is the recovery gap. This is the share of ChronoFact's clean-input macro F1 that is retained on informalized input, compared across our system, the date-prompted LLM, and the "append anchors" ablation. A second check is whether anchor fidelity predicts that gap better than BERTScore does. If anchor fidelity does not predict the gap, the central premise is wrong. If the constrained system closes most of the gap while matching CACN on overlap metrics, the post-editing framing is justified.