Following the Link: Finding Which Health Claim a Cross-Platform Cascade Actually Spreads

Health misinformation often hides in plain sight. A post on Facebook or Reddit says something vague like "doctors don't want you to see this" and links to a twenty-minute YouTube video. The false medical claim lives somewhere inside that video, and a classifier that reads only the post sees nothing wrong. Cross-Platform Multimodal Misinformation documented this "launchpad" pattern for COVID-19 and 5G conspiracies. The share of misinformation rose from 37.4% when judging post text alone to 59.5% when judging the post together with its video, and detection F1 rose from 0.49 to above 0.74 once video transcripts were added. Knowing that a post-video pair is suspect is only half the job, though. A fact-checker or public-health team needs to know which specific claim in a long, rambling transcript is the one people are passing around, so that it can be checked against medical evidence.

This note proposes an unsupervised way to find that claim. We build a small graph that spans the platforms involved: the launchpad post, the replies and comments around it, and the segments of the linked video's transcript. We then use graph centrality, a family of methods that score nodes by how much the rest of the graph points to them, to rank transcript segments by how strongly the audience's own words on other platforms point back at them. The selected segment is the claim the cascade propagates. It is not necessarily the video's main topic, and that difference is what separates this idea from simply asking a model to summarize the video. The similarity function that draws edges between colloquial comments and transcript speech comes from a biomedical language model given further pre-training on lay health text. That choice targets the gap between how patients talk and how the medical literature writes.

Core Claim

Our hypothesis is that, in a cross-platform health cascade, the claim worth checking is the transcript segment with the highest audience-seeded centrality. Here "audience-seeded" means a random walk that starts from the posts and comments on the text platforms and flows into the video transcript. Computed over edges from a register-adapted biomedical encoder, this ranking should identify the propagated claim more accurately than three alternatives: centrality over the transcript alone, a large language model asked for the main claim, and a supervised classifier. The ingredient we claim is new is the set of cross-platform edges. Prior graph summarizers rank units inside one document or one collection, and prior cross-platform misinformation work classifies post-video pairs without ever locating the claim.

Research Program

Each cascade becomes a graph. Its nodes are clause-level units from three sources: the launchpad post, its comment thread or quote shares where available, and the transcript of the linked video split into segments. Within the transcript, consecutive segments are linked to preserve the order of the argument. We take seriously the finding in BioBART that scrambling sentence order hurts biomedical reasoning, so we never treat a transcript as an unordered bag of sentences. Edges across platforms are weighted in two ways. The first is cosine similarity under a biomedical encoder. The second is overlap in Population, Intervention and Outcome spans, the "who, what treatment, what effect" structure that Identifying and Aligning Medical Claims Made on Social Media with Medical Evidence showed is the useful bridge between Reddit claims and clinical trials.

Ranking uses personalized PageRank seeded on the audience side:

$$r = (1-\alpha)\, s + \alpha\, W^{\top} r,$$

In this equation, $r$ is the vector of node scores, $s$ is a restart distribution spread uniformly over post and comment nodes and zero elsewhere, $W$ is the row-normalized edge-weight matrix, and $\alpha$ is the probability of continuing the walk rather than restarting. We read off scores only for transcript nodes. Before ranking, we drop nodes outside the graph's dense core using $k$-core decomposition, which keeps the largest subgraph in which every node has at least $k$ neighbors. This removes generic hub phrases such as "your immune system" that would otherwise attract mass, the pathology described in GoWvis and in a study on graph degeneracy. The method needs no training labels, which matters because labeled cross-platform health data is scarce.

The encoder is where domain pre-training enters. HealthFC found that BioBERT beats general BERT on medical fact-checking. Formal biomedical models still struggle with lay phrasing, a limitation the BioBART authors acknowledge. We therefore continue pre-training BioBERT with masked-token prediction on lay health text, using RedHOT Reddit posts and the synthetic colloquial claims (EMCC) from Hughes and Song. Following BioBART, we omit sentence permutation. The encoder is used only to weight edges, so its contribution can be measured cleanly by swapping it out.

The selected claim is then passed to evidence retrieval and verification. This step tests whether finding the right claim actually improves fact-checking against the literature. SciFact-Open showed that open-literature verification is fragile: MultiVerS falls from 72.5 to 52.4 F1, partly because of mismatches in how specific claims are. Our expectation is that a cleanly isolated claim reduces that mismatch compared with feeding in a whole transcript.

Prior work plays three roles here. Micallef et al. and Hughes and Song motivate the problem: the claim is displaced across platforms, and the language is displaced across registers. BioBART and the degeneracy studies constrain the design, telling us to keep order and to drop hubs. Unsupervised document summarization using pre-trained sentence embeddings and graph centrality and Centrality-guided Pre-training for Graph supply the method family. CenPre explicitly does not handle graphs with several node and edge types, so we do not use it as a learner. We borrow only its premise that centrality carries signal about structural role.

Baselines

The naive reference points come in two forms. The first judges the post text alone, which Micallef et al. report at 0.49 F1 for detection. The second, for claim identification, selects the first transcript segment or a random one. Any useful method must clear both easily.

The strongest supervised competitor is the multi-task BERT-plus-Bi-LSTM model from Cross-Platform Multimodal Misinformation, which reports detection F1 above 0.74 with transcripts. It is the strongest because it was built for exactly these post-video pairs. We will run it on the same splits, and also in a variant that swaps in BioBERT, giving it the same encoder budget as our method. Our method does not have to beat it at pair-level detection. It has to match it within a few F1 points when the classifier is shown only the selected segment rather than the full transcript. That result would show the selected segment carries the misinformation signal, and it would add the ability to point to the claim, which the classifier lacks.

The strongest unsupervised competitor is centrality over the transcript alone, using the sentence-embedding approach of the graph-centrality summarization paper cited above. It runs on the identical encoder and transcript segments and differs only in having no audience nodes. Beating it on claim identification is the core test of our hypothesis.

A large language model, prompted with the post and the full transcript to name the claim being spread, is the most practical competitor. We will give it the same inputs and a fixed prompt-tuning budget on a held-out development set. We found no reported result on Lacuna for language-model claim extraction on this data, so we will measure it ourselves.

The event-graph network of Cross-document Misinformation Detection based on Event Graph Reasoning is the closest graph-based method. It was built for news clusters and depends on a supervised information-extraction system whose errors cascade downstream. We will report it only if its pipeline can be adapted at reasonable cost, and otherwise treat it as context.

Two ablations isolate our ingredients. The first removes the cross-platform edges while keeping the post and comment nodes as isolated restart points. The second swaps the encoder among general BERT, BioBERT and our register-adapted model.

Why not just…?

Why not just ask a large language model for the main claim? This is the real threat. A capable model given the transcript and post may well find the propagated claim. Our bet is narrower: the propagated claim and the main claim often differ, especially in long videos whose most-shared moment is a side remark. In those cases the audience graph carries information that the transcript alone does not. We will report results separately for cascades where annotators judge the propagated claim to differ from the video's main thesis. If the language model matches us there too, the graph's value shrinks to cost and transparency, and we will say so. A hybrid that gives the model the top centrality-ranked segments is a natural fallback, in the spirit of A Systematic Exploration of Knowledge Graph Alignment with Large Language Models in Retrieval Augmented Generation. That study reports central nodes acting as anchors and an average gain of 7.3% across 15 models on knowledge-graph question answering.

Why not just fine-tune Micallef's classifier with a biomedical encoder? That would improve detection, but a pair-level label does not say which sentence to check, and fact-checking needs that sentence. Hughes and Song show how weak retrieval from clinical trials is even with a good claim in hand (DPR-PIO Precision@1 of 8.17%). Retrieval from an unlocated transcript can only be worse. The retrieval experiment below measures this directly.

Why not just run TextRank on the transcript? Transcript-only centrality finds what the speaker emphasizes, not what the audience picked up. This alternative is identical to our method minus the audience nodes, so the comparison is a direct, controlled test of the core claim.

Proposed Experiments

The primary dataset is the Micallef et al. collection of 3,000 labeled posts linked to 991 YouTube videos, each tied to one of 32 claims verified as false by the International Fact-Checking Network. That link provides gold labels for claim identification: a selected segment counts as correct if two annotators agree it asserts the labeled claim. If the original posts or comments can no longer be retrieved, we will re-collect cascades for the same videos and report how much the data shrank.

The first metric is claim identification: accuracy of the top-1 segment and recall within the top 3. The second is detection F1 when a classifier sees only the selected segment, compared against the reported full-transcript figure of above 0.74.

For evidence grounding, annotators will pair a sample of around 200 selected claims with supporting or refuting literature, following HealthFC's labels (supported, refuted, not enough information) and its evidence-strength scale. We will measure retrieval precision and verification macro-F1 with a HealthFC-trained verifier. As a calibration check, HealthFC itself reports DeBERTa pipeline veracity at 65.1% macro-F1 and evidence selection at 55.5%, so a cross-platform score near those values would indicate that claim isolation has largely closed the gap with clean editorial claims.

The encoder ablation runs on both claim identification and edge-alignment quality. Alignment quality is measured on pairs of comments and transcript segments that annotators judge to be paraphrases.

The decisive checks are two. First, audience-seeded centrality must beat transcript-only centrality on claim identification, most clearly in cascades where the propagated claim differs from the main thesis. Second, the register-adapted encoder must improve on BioBERT in edge alignment. If the first fails, the cross-platform graph adds nothing. If only the second fails, the domain pre-training step can be dropped without harming the core claim.