Insight Rumors: Token-Level Detection & Analysis

• Insight Rumors are a novel framework that redefines rumor detection as a token-level sequence labeling task, enabling precise localization of rumor spans.
• The approach integrates Chinese BERT embeddings, Att_BiMamba2 feature extraction, and CRF decoding to achieve high performance on annotated datasets.
• Insight Rumors further extends analysis to propagation dynamics and user behavior, providing actionable insights for early warning and multiscale monitoring.

Searching arXiv for recent and related papers on “Insight Rumors” and rumor analysis/detection. “Insight Rumors” denotes a line of work that treats rumors not only as items to be classified, but as objects to be detected, localized, interpreted, and analyzed across content, users, propagation, and time. In the narrowest sense, the term names the model “Insight Rumors: A Novel Textual Rumor Locating and Marking Model Leveraging Att_BiMamba2 Network,” which reframes rumor detection as a sequence labeling problem over text spans rather than a post-level binary decision (Ma et al., 18 Aug 2025). In a broader research sense, the term aligns with systems and analyses that ask what is being said, who is saying it, how it spreads, where and when it emerges, and whether it is false or misleading, using combinations of text matching, sequence models, graph methods, visual analytics, and fact-checked corpora (Jin et al., 2017, Wang et al., 2022, Zhang et al., 2024, Schafer et al., 2024).

1. Conceptual scope and problem formulation

At the most general level, rumor research distinguishes rumor detection from rumor verification. Rumor detection decides whether a post should be labeled as rumor or non-rumor, given the text and associated metadata, whereas rumor verification decides whether a post already identified as a rumor is true, false, or unverified (Sheikhqoraei et al., 28 Jan 2026). One formulation defines a rumor as a “controversial and fact-checkable statement” in election discourse (Jin et al., 2017), while another defines a rumor as content that remains unverified at the time of dissemination (Sheikhqoraei et al., 28 Jan 2026). A further operationalization treats rumors as true/false labeled claims in binary text classification settings, especially when source tweets or headlines are used without propagation context (Ni et al., 2021).

Within this landscape, “Insight Rumors” in the model-specific sense shifts the target from “Is this text a rumor?” to “Where is the rumor in the text?” (Ma et al., 18 Aug 2025). The task is posed over a token sequence

$X = \{x_1, x_2, \dots, x_n\},$

with output label sequence

$Y = \{y_1, y_2, \dots, y_n\},$

using the label set

$$\text{Label} = \{\text{B-Rumor}, \text{I-Rumor}, O\}.$$

Here B-Rumor denotes the beginning of a rumor span, I-Rumor denotes continuation inside a rumor span, and $O$ denotes non-rumor context (Ma et al., 18 Aug 2025). This is explicitly a sequence labeling problem similar to named entity recognition, but the “entities” are rumor spans (Ma et al., 18 Aug 2025).

This problem formulation stands alongside several other established formulations. Some work treats rumor detection as text matching between posts and verified rumor articles, rather than as supervised classification, so that each matched post can be linked to a specific fact-checked claim (Jin et al., 2017). Other work formulates early rumor detection as a temporal decision problem over repost sequences, introducing a microblog-specific credible detection point $\beta$ at which a stable high-confidence prediction first becomes possible (Song et al., 2018). Visual analytics systems instead assume a set of suspected rumors has already been identified and focus on interactive analysis and validation through linked views of space, time, features, and propagation (Wang et al., 2022).

This suggests that “Insight Rumors” is best understood as a multiscale program rather than a single method: post-level discrimination, token-level locating and marking, event-level temporal detection, propagation-level surveillance, and analyst-facing validation each correspond to distinct but complementary subproblems (Ma et al., 18 Aug 2025, Song et al., 2018, Wang et al., 2022).

2. Textual rumor locating and marking

The core contribution of the model named “Insight Rumors” is a pipeline with three stages: word encoding with bert_base_chinese, Att_BiMamba2 rumor feature extraction, and a Rumor Locating and Marking module followed by CRF decoding (Ma et al., 18 Aug 2025). The design aim is not only to detect rumors accurately but also to locate and mark them in context precisely (Ma et al., 18 Aug 2025).

The first stage uses Chinese BERT for contextual token embeddings. Token embeddings $TEx_i$, segment embeddings $Sx_i$, and positional embeddings $Px_i$ are combined as

$$E = [TEx_1 + Sx_1 + Px_1, \dots, TEx_n + Sx_n + Px_n],$$

and BERT produces contextual outputs

$T = \{T_1, T_2, \dots, T_N\}$

after 12 Transformer encoder layers (Ma et al., 18 Aug 2025). The self-attention sublayer follows

$Y = \{y_1, y_2, \dots, y_n\},$0

and the feedforward block is written as

$Y = \{y_1, y_2, \dots, y_n\},$1

(Ma et al., 18 Aug 2025).

The second stage is Att_BiMamba2. After a linear adjustment

$Y = \{y_1, y_2, \dots, y_n\},$2

the model applies a forward Mamba2 encoder

$Y = \{y_1, y_2, \dots, y_n\},$3

and a backward Mamba2 encoder

$Y = \{y_1, y_2, \dots, y_n\},$4

Inside a Mamba2 block, the sequence is linearly projected,

$Y = \{y_1, y_2, \dots, y_n\},$5

passed through depthwise 1D convolution and SILU activation,

$Y = \{y_1, y_2, \dots, y_n\},$6

then processed by the state space model

$Y = \{y_1, y_2, \dots, y_n\},$7

followed by residual connection,

$Y = \{y_1, y_2, \dots, y_n\},$8

RMS normalization,

$Y = \{y_1, y_2, \dots, y_n\},$9

and final linear projection,

$$\text{Label} = \{\text{B-Rumor}, \text{I-Rumor}, O\}.$$0

(Ma et al., 18 Aug 2025).

The bidirectional outputs are fused by dot-product attention rather than simple concatenation. Scores are computed as

$$\text{Label} = \{\text{B-Rumor}, \text{I-Rumor}, O\}.$$1

and the final rumor representation is

$$\text{Label} = \{\text{B-Rumor}, \text{I-Rumor}, O\}.$$2

The paper states that this enhances the representation of high-dimensional rumor features (Ma et al., 18 Aug 2025).

The third stage maps these high-dimensional rumor features into label features. A skip-connection network performs staged projection: $$\text{Label} = \{\text{B-Rumor}, \text{I-Rumor}, O\}.$$3

$$\text{Label} = \{\text{B-Rumor}, \text{I-Rumor}, O\}.$$4

$$\text{Label} = \{\text{B-Rumor}, \text{I-Rumor}, O\}.$$5

A CRF layer then imposes strong constraints on the output label features (Ma et al., 18 Aug 2025). The CRF sequence score is

$$\text{Label} = \{\text{B-Rumor}, \text{I-Rumor}, O\}.$$6

with conditional probability

$$\text{Label} = \{\text{B-Rumor}, \text{I-Rumor}, O\}.$$7

negative log-likelihood

$$\text{Label} = \{\text{B-Rumor}, \text{I-Rumor}, O\}.$$8

and decoding objective

$$\text{Label} = \{\text{B-Rumor}, \text{I-Rumor}, O\}.$$9

solved by the Viterbi algorithm (Ma et al., 18 Aug 2025).

The associated dataset, IR-WEIBO, contains 3,200 Sina Weibo text posts from verified rumors, annotated with B-Rumor, I-Rumor, and O labels, and split into 80% train, 10% validation, and 10% test (Ma et al., 18 Aug 2025). On IR-WEIBO, the reported test-set metrics for the proposed model are $O$0 $O$1 for B-Rumor, $O$2 $O$3 for I-Rumor, and $O$4 accuracy for O, with the paper stating that the scheme outperforms state-of-the-art sequence labeling baselines adapted to rumor spans (Ma et al., 18 Aug 2025).

3. Detection, identification, and early warning

A major thread adjacent to “Insight Rumors” treats rumor detection as either retrieval against verified claims or early sequential prediction. In election rumor detection on Twitter, one paper avoids large manually labeled training sets by matching tweets against 1,723 Snopes rumor articles and selecting the article with highest similarity

$O$5

If the best score exceeds a threshold $O$6, the tweet is labeled as a rumor referring to $O$7 (Jin et al., 2017). This design has two stated advantages: minimal manual labeling and interpretable results, because one knows which verified rumor article each tweet refers to (Jin et al., 2017).

That work compares TF-IDF, BM25, Word2Vec, Doc2Vec, and lexicon-based matching. BM25 achieves the best classification $O$8 of $O$9, compared with $\beta$0 for TF-IDF, $\beta$1 for Word2Vec, and $\beta$2 for Doc2Vec; for rumor identification, BM25 reaches $\beta$3 accuracy (Jin et al., 2017). For large-scale deployment, the chosen BM25 threshold is $\beta$4, yielding $\beta$5 precision and $\beta$6 recall on the test set (Jin et al., 2017). This suggests a conservative operating point geared toward high-confidence analysis of relative rumor activity rather than exhaustive recall.

Early rumor detection emphasizes temporal stability rather than only end-of-sequence accuracy. The Credible Early Detection model groups reposts into intervals of $\beta$7, represents each interval by TF-IDF or CNN features, and processes the sequence with a GRU (Song et al., 2018). It defines the credible detection point

$\beta$8

where $\beta$9 is the first time step at which $TEx_i$0 crosses a confidence threshold $TEx_i$1 (Song et al., 2018). Its objective combines prediction accuracy after CDP, earliness via

$TEx_i$2

and temporal stability through $TEx_i$3, which penalizes post-CDP threshold violations (Song et al., 2018).

On Weibo-all, the best baseline GRU-2 has $TEx_i$4 accuracy and $TEx_i$5 $TEx_i$6 with early rate $TEx_i$7, whereas CED-CNN with $TEx_i$8 achieves $TEx_i$9 accuracy, $Sx_i$0 $Sx_i$1, and $Sx_i$2 early rate (Song et al., 2018). On Twitter, CED reaches $Sx_i$3 accuracy and $Sx_i$4 $Sx_i$5 with $Sx_i$6 early rate, and CED-CNN with $Sx_i$7 reaches $Sx_i$8 accuracy and $Sx_i$9 $Px_i$0 with $Px_i$1 early rate (Song et al., 2018). The paper states that the proposed model can reduce the time span for prediction by more than $Px_i$2 across settings (Song et al., 2018).

A different recent direction addresses multi-domain rumor detection under domain shift. The PerFact study introduces a Persian dataset from X with 8,034 annotated posts in rumor and non-rumor categories and proposes a domain-gated Mixture-of-Experts model that aggregates text and publisher information (Sheikhqoraei et al., 28 Jan 2026). Annotator agreement is reported as Fleiss’ Kappa $Px_i$3, and the model attains an $Px_i$4-score of $Px_i$5 and an accuracy of $Px_i$6 in multi-domain settings (Sheikhqoraei et al., 28 Jan 2026). This suggests that domain-aware routing is one response to lexical and topical shift, though the task remains binary rumor detection rather than token-level rumor marking.

4. Propagation, users, and multiscale analytics

Rumor analysis extends beyond text into propagation geometry, user roles, and analyst workflows. RumorLens is an interactive visual analytics system developed with administrators from Sohu, NetEase, and TouTiao over four months, structured around three requirements: explore the overall space–time distribution of suspected rumors, inspect suspected rumor cases through feature comparisons, and explore propagation details of individual suspected rumors (Wang et al., 2022). Its Sina Weibo dataset contains 936 suspected rumors, approximately 80,000 corresponding retweets and comments, and 53,843 user profiles from 2019/12/27–2020/12/14 (Wang et al., 2022).

The system integrates TF-IDF keywords, BiLSTM-based sentiment recognition, BERT-based topic classification, influence measures, t-SNE projection, and a circular propagation view (Wang et al., 2022). In the propagation view, concentric rings encode retweet depth, sectors encode temporal progression across days, and cells encode retweets or comments with sentiment as color and text length as size (Wang et al., 2022). The case study reports that a domain expert identified a highly suspicious rumor through a large glyph with low integrity and few fans, then validated it via six retweet rings, dominance of red and green cells, and comments indicating that the original tweet was misleading (Wang et al., 2022). The paper states that RumorLens supports a traceable reasoning process combining spatial-temporal patterns, feature-based anomaly detection, propagation analysis, and rich content and sentiment evidence (Wang et al., 2022).

Cross-platform cascade analysis further differentiates rumor and non-rumor dynamics. A study on Twitter and Weibo with nearly one million crawled user profiles reports that rumors tend to spread more deeply, while non-rumors distribute more broadly (Hou et al., 2024). It also reports that rumors are slower, persist longer, and, in most cases, involve fewer participants than non-rumors, while a small minority of sensational rumors produce very large cascades (Hou et al., 2024). Source users of non-rumor cascades are much more likely to be verified, whereas rumor participants are described as active, long-standing, decently credible users termed “onlookers,” who inadvertently or unwittingly spread rumors due to extensive online interactions and the allure of sensational fake news (Hou et al., 2024).

The same study identifies exponential patterns in cascade features and introduces the Credibility Erosion Effect, under which the credibility of a person who repeatedly broadcasts and shares the same information gradually declines over time (Hou et al., 2024). This points toward a propagation-level notion of “insight rumors” in which depth, breadth, source attributes, participant reputation, and temporal persistence matter alongside textual content.

Graph-based infodemic surveillance takes yet another perspective. A unified GNN model for rumor detection, virality prediction, and user vulnerability scoring builds a propagation graph $Px_i$7, a user interaction graph $Px_i$8, community assignments via DiffPool, and time-aware post embeddings with BERT and a fully connected time encoder (Zhang et al., 2024). It defines virality prediction as

$Px_i$9

and user vulnerability as the fraction of rumor events among all events a user engaged in (Zhang et al., 2024). On WEIBO, the MT-META variant reaches $$E = [TEx_1 + Sx_1 + Px_1, \dots, TEx_n + Sx_n + Px_n],$$0 accuracy and $$E = [TEx_1 + Sx_1 + Px_1, \dots, TEx_n + Sx_n + Px_n],$$1 MacF1 for rumor detection, with virality MSE $$E = [TEx_1 + Sx_1 + Px_1, \dots, TEx_n + Sx_n + Px_n],$$2 and vulnerability MSE $$E = [TEx_1 + Sx_1 + Px_1, \dots, TEx_n + Sx_n + Px_n],$$3; on TWITTER it reaches $$E = [TEx_1 + Sx_1 + Px_1, \dots, TEx_n + Sx_n + Px_n],$$4 accuracy and $$E = [TEx_1 + Sx_1 + Px_1, \dots, TEx_n + Sx_n + Px_n],$$5 MacF1 (Zhang et al., 2024). The paper argues that user vulnerability acts as a bridge between rumor status and virality (Zhang et al., 2024).

5. Empirical rumor insights in elections and crises

Election rumor studies provide concrete demonstrations of what “insight rumors” can mean empirically. In the 2016 U.S. presidential election, analysis of 8,731,137 tweets from 7,283 Clinton followers and 7,339 Trump followers finds that rumor tweeting is highly concentrated: the top 10% of users account for about 50% of rumor tweets, and the top 20% account for about 70% (Jin et al., 2017). During April–September 2016, the ratio of rumor tweets is $$E = [TEx_1 + Sx_1 + Px_1, \dots, TEx_n + Sx_n + Px_n],$$6 for Clinton followers and $$E = [TEx_1 + Sx_1 + Px_1, \dots, TEx_n + Sx_n + Px_n],$$7 for Trump followers, while over the entire time span it is $$E = [TEx_1 + Sx_1 + Px_1, \dots, TEx_n + Sx_n + Px_n],$$8 and $$E = [TEx_1 + Sx_1 + Px_1, \dots, TEx_n + Sx_n + Px_n],$$9, respectively (Jin et al., 2017). The paper also states that both camps post rumors about both candidates, with opponent-targeted rumors more frequent for both groups (Jin et al., 2017).

The same study reports event-driven temporal peaks around presidential debates, official nomination events, controversial emergency events such as the Orlando shooting, and developments that directly trigger specific rumors such as Clinton’s public reappearance after pneumonia (Jin et al., 2017). It explicitly notes that stance is not modeled: the same detected rumor can be amplified by one side and questioned by the other (Jin et al., 2017). This suggests that rumor matching alone can conflate rumor spread with rumor discussion.

A later election dataset extends this perspective to the 2022 U.S. midterms. ElectionRumors2022 contains approximately 1.81 million Twitter posts linked to 135 distinct rumors from September 5 to December 1, 2022, with 88.0% retweets and approximately 427,600 unique accounts (Schafer et al., 2024). The collection process starts from about 446 million election-related tweets captured via the Twitter v1.1 Streaming API, then uses rumor leads, query development, inclusion coding, and tweet-level quality assurance (Schafer et al., 2024). Across all incidents and final queries, 99% of sampled top retweets and 96% of randomly sampled non-retweets are correctly associated with the intended rumor (Schafer et al., 2024).

The study reports a sharp Election Day spike followed by rapid decay: within about 24 hours, rumor tweet volume drops to less than one-fifth of its Election Day peak (Schafer et al., 2024). Arizona dominates the geographic distribution, accounting for 42.1% of incident-level coding and 34.7% of direct textual state mentions among rumor tweets (Schafer et al., 2024). Partisan labeling via coengagement networks assigns right-leaning labels to 262,448 accounts, responsible for 1,430,244 posts and dominating 109 of 135 rumors; left-leaning labels are assigned to 125,915 accounts, responsible for 327,564 posts and dominating 22 rumors (Schafer et al., 2024). The paper also reports that all users exhibit extreme retweet concentration, with a Gini coefficient of approximately 0.96 in 2022 and the top 1% garnering about 84% of all retweets (Schafer et al., 2024).

Crisis settings reveal other dimensions. During COVID-19 in China, rumor influence is modeled through search-engine traces as a proxy for “new insiders,” yielding an exponential form

$T = \{T_1, T_2, \dots, T_N\}$0

where $T = \{T_1, T_2, \dots, T_N\}$1 is the attenuation coefficient and $T = \{T_1, T_2, \dots, T_N\}$2 is the peak coefficient (Lin et al., 2020). A 5-fold cross-validation experiment using MSE shows that a decision tree is suitable for predicting the peak coefficient, while linear regression is ideal for predicting the attenuation coefficient (Lin et al., 2020). The feature analysis states that precursor features are most important for the outbreak coefficient, location information and rumor entity information are most important for the attenuation coefficient, and anxiety is a crucial rumor-causing factor (Lin et al., 2020).

6. Limits, controversies, and future directions

A recurring controversy is whether high-performing models actually learn rumor detection rather than dataset-specific shortcuts. Cross-dataset experiments with BERT-Base uncased on Twitter15, Twitter16, PHEME, GossipCop, and PolitiFact show strong in-domain $T = \{T_1, T_2, \dots, T_N\}$3 scores such as $T = \{T_1, T_2, \dots, T_N\}$4, $T = \{T_1, T_2, \dots, T_N\}$5, $T = \{T_1, T_2, \dots, T_N\}$6, $T = \{T_1, T_2, \dots, T_N\}$7, and $T = \{T_1, T_2, \dots, T_N\}$8, but unsatisfactory out-of-domain performance, with many cross-dataset $T = \{T_1, T_2, \dots, T_N\}$9 values around 40–50 and some below that range (Ni et al., 2021). On a common-sense rumor set of 200 samples, accuracies are about 48–52%, essentially at chance (Ni et al., 2021). The paper further shows that clue words such as “Obama,” “Paul,” and “Sydney” can dominate model attention and that simple semantic reversals cause large accuracy declines, leading to the claim that models take shortcuts and learn absurd knowledge when rumor datasets have serious data pitfalls (Ni et al., 2021).

To address this, the same work proposes PairT, a paired test requiring a model to correctly predict both elements of a semantically linked pair at the same time (Ni et al., 2021). This suggests that evaluation of “insight rumors” systems should emphasize consistency and semantic sensitivity, not only standard accuracy on benchmark splits.

Other limitations are methodological rather than evaluative. The 2016 election rumor-matching approach depends entirely on Snopes.com, misses rumors not covered there, and does not model stance or sentiment (Jin et al., 2017). RumorLens assumes suspected rumors have already been detected and focuses on analysis and validation rather than initial detection (Wang et al., 2022). The Att_BiMamba2 “Insight Rumors” model is built and evaluated on Chinese Weibo text, with bert_base_chinese and IR-WEIBO, so adaptation to other languages and platforms is not addressed in the reported experiments (Ma et al., 18 Aug 2025). Multi-domain work explicitly identifies lexical shift, topical shift, propagation and user behavior differences, and event novelty as major obstacles to generalization (Sheikhqoraei et al., 28 Jan 2026).

Future directions in the provided literature converge on richer integration rather than a single dominant architecture. Retrieval, discrimination, and generation systems seek grounded explanatory debunking rather than only labels, combining an Expert-Citizen Collective Wisdom module, a real-time debunking vector database, and LLM-based generation with retrieval-augmented prompts (Xu et al., 2024). Multi-domain mixture-of-experts models attempt to balance domain-specific and global features (Sheikhqoraei et al., 28 Jan 2026). Token-level locating and marking introduces a more precise textual target for moderation and explainability (Ma et al., 18 Aug 2025). This suggests that the future meaning of “Insight Rumors” may increasingly involve joint reasoning over claim spans, verified evidence, user roles, propagation structure, and human-in-the-loop validation, rather than coarse post-level rumor discrimination alone.