{"id":"a44b8522-f136-44af-b08d-446cf076309a","arxiv_id":"2412.13668","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A BERT-GRU model fusing personal, neighbor, and historical emotion signals predicts users' next emotion on rumor conversations and beats DeGroot, Voter, and the authors' earlier E-USIM on one Twitter dataset.","lead":"This paper proposes E-USIM+, a machine learning model that predicts a social media user's next emotion during rumor and non-rumor conversations by combining the user's own emotional history, neighbors' emotional reactions, and text meaning. If the approach holds up, it could help platforms identify when false information is driving emotional contagion.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported gain over E-USIM may come from BERT content leakage rather than emotional influence: Fu(i) appears to encode the same tweet whose emotion is the prediction target, and no content-only ablation is reported.","rationale":"The reader's weakest assumption concerns the causal validity of temporal co-occurrence as influence. I agree that causal validation is missing, but the more pressing technical risk is that the reported accuracy advantage is uninterpretable because the BERT feature may encode the target tweet. If Fu(i) indeed uses the target tweet's content, the comparisons to DeGroot, Voter, and E-USIM do not hold fixed content features, so the Table 1 gains could be a text-classification artifact rather than evidence that historical emotional influence helps. This is a concrete, checkable leakage/ablation issue, not merely a philosophical critique of causal inference. Still, the paper could be correct: E-USIM+ may genuinely outperform the baselines, and the temporal influence may contribute. Because the paper provides no code, no hyperparameters, no confidence intervals, and no target-tweet alignment statement, the appropriate verdict remains CONDITIONAL. My concern and the reader's concern are distinct but complementary, so I partially agree with the reader's weakest-assumption analysis.","tokens_in":4889,"tokens_out":6702,"duration_ms":60516,"concrete_test":"Obtain the code or a written specification and check whether Fu(i) and the label Su(i) come from the same tweet ID. Then rerun Table 1 on the same filtered Kwon data with three controls: (i) a BERT-only classifier on the same content features used in Fu(i), with no GRU temporal input; (ii) E-USIM+ with Fu(i) removed; (iii) E-USIM+ with the temporal order of the GRU input shuffled. If BERT-only matches or exceeds E-USIM+ accuracy, or if removing or shuffling the content or temporal features does not degrade performance, the central emotional-influence claim would be undermined; if both controls are clearly worse, the claim would survive.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that E-USIM+ predicts a user's future emotion by fusing GRU-captured emotional influence (Eqs. 1-4) with BERT semantic features, and Table 1 shows clear gains over E-USIM on rumour/non-rumour data. The load-bearing issue is that the object of the BERT encoding is never defined as past or neighbor content. In Section 2.3, Fu(i) is 'the learned tweet semantic representation vector' obtained from 'the tweet content within the length n' and social structural vectors. In Section 2.4, the target is P(Su(i)|Hu,i(t)) with Hu,i(t) = Cu,i(t) ⊕ Fu(i). If Fu(i) is the BERT embedding of the very tweet i whose emotion Su(i) is the label, then the label can be inferred directly from content, and the GRU influence chain is bypassed. The paper never rules this out, and no baseline that uses BERT content features without the temporal GRU is reported. Even if Fu(i) instead encodes neighboring incoming tweets, the comparison is still confounded: DeGroot and Voter have no content features, and E-USIM differs from E-USIM+ in the semantic component, so the Table 1 advantage cannot be attributed to historical temporal influence. The conclusion that 'integrating historical information... is better' is therefore not supported without ablations or an explicit statement of the input-to-label alignment.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes E-USIM+, an extension of the authors' earlier E-USIM model, to predict a user's future emotional state on rumour and non-rumour tweets in online social networks. The model combines a GRU over temporal sequences of emotional influence signals (prior neighbouring, prior personal, current neighbouring, historical) with BERT-derived semantic and social structural features, fused as H_{u,i}(t) = C_{u,i}(t) ⊕ F_u(i), followed by a softmax classifier over four emotional regions. Experiments on the Kwon et al. (2017) dataset report accuracy improvements over DeGroot, Voter, and E-USIM baselines for both rumour and non-rumour conversations (e.g., rumour accuracy 0.6728 vs. 0.6234 for E-USIM).","tokens_in":5116,"tokens_out":1858,"duration_ms":17655,"significance":"If the reported gains are real and the model is specified without leakage, the paper would make a useful contribution to emotion dynamics and misinformation research by showing that combining temporal emotional influence with semantic content features improves future emotion prediction. The paper uses a publicly available dataset, compares against three baselines, and extends the authors' prior model, which provides continuity. However, the central empirical claim rests on a single dataset, a single run, and a model specification that is ambiguous about the alignment between the BERT-encoded content and the prediction target. The absence of code, ablations, and statistical tests limits the strength of the evidence, and the paper currently does not establish that the improvement is attributable to temporal emotional influence rather than content leakage or an uneven comparison.","major_comments":[{"comment":"The most load-bearing issue is that the paper never states whether F_u(i) is the BERT embedding of the user's own tweet i whose emotion S_u(i) is the prediction target, or an embedding of preceding neighbour tweets and the user's own prior tweets. In §2.4 the model predicts P(S_u(i)|H_{u,i}(t)) with H_{u,i}(t) = C_{u,i}(t) ⊕ F_u(i); §2.3 defines F_u(i) as 'the learned tweet semantic representation vector' from 'the tweet content within the length n'. If F_u(i) encodes the very tweet whose emotion is the label, then the label can be inferred directly from content and the GRU influence chain is bypassed, making the comparison against E-USIM and the content-free DeGroot/Voter baselines uninformative. The paper must specify the exact alignment (which tweets and which timestamps are encoded in F_u(i)) and must report a content-only baseline (BERT features without the temporal GRU) and an ablation without content features to disentangle the contribution of temporal influence from content semantics.","section":"§2.3–2.4, Eq. (5)–(6)"},{"comment":"The reported accuracy differences, e.g., rumour accuracy 0.6728 vs. 0.6234 for E-USIM, are presented without any uncertainty quantification. There are no confidence intervals, no significance tests, no standard deviations over repeated runs, and only one dataset. Since the improvements are modest and the dataset is filtered to users with at least two messages and tweets that show emotion, the generalizability of the claim 'integrating historical information into an influential model is better' is not established. The authors should provide error bars, significance tests, and ideally results on additional datasets or at least a discussion of the filtering's impact on scope.","section":"§3, Table 1"},{"comment":"The temporal chain construction assumes that incoming neighbour tweets received between a user's posts, together with the user's own historical emotional sequence, are sufficient to determine the user's next emotion. This is a strong causal assumption. The paper does not control for exogenous events, topic, network structure, or text content beyond the BERT features, and does not rule out the possibility that temporal co-occurrence reflects common response to an external event rather than social influence. The authors should at least discuss this limitation and, ideally, add a control experiment (e.g., permuting the order of incoming tweets or using a shuffled-timestamp baseline) to test whether the temporal ordering, rather than mere co-occurrence, is responsible for the predictive gains.","section":"§2.2, Eq. (4)"}],"minor_comments":[{"comment":"The abstract contains a grammatical error: 'we systematically examining' should be 'we systematically examine'.","section":"Abstract"},{"comment":"Figure 1 is essentially unreadable in the provided manuscript; the labels and the flow are illegible. A larger, clearly annotated figure is needed.","section":"§2, Fig. 1"},{"comment":"The notation for the temporal tweet chain (e.g., arrows for incoming/outgoing tweets, the subscripts and superscripts) is difficult to parse. The authors should provide a clean example with concrete timestamps and users to show how the chain is built.","section":"§2.1"},{"comment":"The definitions of the gates r_{u,i}(t) and g_{u,i}(t) in Eqs. (1)–(2) use nonstandard symbols and the text does not define the reset and update gates explicitly; a standard GRU formulation with clear gate names would improve readability.","section":"§2.2"},{"comment":"No details are provided on hyperparameters, number of GRU layers, hidden dimensions, BERT version, or training configuration, making the experiments hard to reproduce. The authors should include a reproducibility statement or a hyperparameter table.","section":"§3"},{"comment":"Reference [6] is the authors' own prior E-USIM paper; the relation between E-USIM and E-USIM+ should be stated more explicitly in the text, not only in the comparison table.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper's central idea is interesting and the empirical direction is plausible, but the missing specification of what exactly the BERT branch encodes is a fundamental ambiguity that could render the headline result a content-leakage artefact. This is fixable with a clear statement and proper ablations, but it is load-bearing and cannot be resolved by presentation alone. I also note that the paper is quite short for a full workshop/conference submission; the experimental section would benefit from a robustness analysis."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a modest extension of the authors' earlier E-USIM model, and the stress-test concern is real. The paper never says what the BERT semantic vector Fu(i) sees. If it encodes the tweet whose emotion is the prediction target, then Table 1 mostly shows content classification and the GRU influence chain is bypassed.\n\nWhat is new: E-USIM+ adds BERT semantic features to the GRU-based temporal emotion model and reports a fresh evaluation on the Kwon rumor dataset. The four-quadrant Russell emotion space and the temporal chain construction come from E-USIM, so the delta is the fusion of content/social features with the emotion-sequence GRU. The method is described well enough for someone with the existing pipeline to reimplement, though no code or hyperparameters are given.\n\nWhat it does well: the paper is honest about the incremental nature, and the GRU equations are standard. The comparison is at least directional on the data: E-USIM+ beats DeGroot, Voter, and the re-implemented E-USIM on both rumor and non-rumor accuracy.\n\nSoft spots: Table 1 has no confidence intervals, significance tests, or repeated runs, and it rests on one filtered Twitter dataset. DeGroot and Voter are weak baselines, and the E-USIM baseline is described as a 'simplified version' of the authors' own model, so the comparison is not strong. The bigger problem is Section 2.3/2.4: Fu(i) shares the index i with the target Su(i), and the paper never states whether it includes the content of the target tweet. Without that definition, and without a content-only ablation, the improvement cannot be attributed to historical temporal influence. This is a missing input-output definition, not just a missing robustness check. The writing also has typesetting errors that make the temporal chain hard to follow. The citation pattern is not what bothers me: leaning on [6] is normal for an incremental extension. The issue is that the baseline is called simplified.\n\nWho this is for: researchers working on emotion forecasting in rumor threads, who might use this as a baseline method. I would not cite it as evidence until the leakage question is resolved.\n\nRecommendation: I would desk reject the current form. The missing definition, no code, no statistics, and one dataset are too weak to justify referee time. If the authors resubmit with an explicit statement that Fu(i) is built only from prior/neighbor content, a content-only ablation, and significance tests, it becomes a reasonable short-paper submission.","headline":"A plausible incremental extension of the authors' own E-USIM, but the BERT term is never tied to a defined input window, so the headline gain may be content leakage rather than emotional influence.","tokens_in":5705,"tokens_out":6098,"would_cite":false,"duration_ms":54147,"reading_group":"no","serious_thinker":"unclear","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"E-USIM+ fuses GRU-encoded emotional history and neighbor emotions with BERT semantics to predict users' next emotions on rumor and non-rumor conversations, beating DeGroot, Voter, and E-USIM.","keywords":["emotion prediction","false information","rumor detection","emotional contagion","GRU","BERT","temporal social network","sequential influence"],"falsifier":"Randomly permute the order of incoming neighbor tweets within each user's inter-post gap and retrain E-USIM+; if accuracy stays roughly the same, the temporal ordering the model exploits is not carrying the prediction, and the claimed sequential influence is an artifact of content features.","tokens_in":4630,"feed_emoji":"📣","tokens_out":9671,"duration_ms":75918,"temperature":0.7,"pith_summary":"The paper proposes E-USIM+ (Emotional-based User Sequential Influence Model plus) and claims that a person's next emotion in a false-information conversation is predictably shaped by four sequential influences: what neighbors said emotionally before, what the person themselves felt before, what neighbors are saying now, and the person's accumulated emotional state. The model feeds these temporal signals through a gated recurrent unit (GRU), fuses the resulting memory vector with BERT-encoded semantic and social-context features, and classifies the next emotion into one of four regions: excitement, contentment, depression, or distress. On the 111-event Twitter rumor dataset (60 rumors, 51 non-rumors), E-USIM+ reaches 0.6728 accuracy on rumor conversations and 0.7187 on non-rumor conversations, outperforming DeGroot, Voter, and the earlier E-USIM. If this is right, emotional contagion in online rumor threads can be tracked at the individual level from message timing and content alone.","feed_headline":"Emotion history plus neighbor moods predicts rumor reactions","feed_subtitle":"E-USIM+ fuses BERT semantics with GRU-tracked personal and neighbor emotions, beating DeGroot, Voter, and E-USIM.","key_machinery":"The load-bearing object is the temporal tweet chain: for each user, outgoing posts are interleaved with incoming posts from neighbors, and each post carries one of four emotion regions from the circumplex model of affect. From these chains the model derives four influence signals—prior neighboring, prior personal, current neighboring, and historical—and feeds them into a GRU whose hidden state $C_{u,i}(t)$ acts as a running emotional memory. The GRU's reset and update gates decide what to keep and what to drop from the emotional past. The semantic branch uses BERT's multi-head self-attention to encode tweet words plus social structural features into $F_u(i)$, and the final vector $H_{u,i}(t) = (C_{u,i}(t) \\oplus F_u(i))$ is passed to a softmax layer for classification.","core_discovery":"The central claim is that historical emotional state is the missing ingredient in sequential influence models of emotion on false information. E-USIM+ replaces the static historical-influence term of E-USIM with the GRU's internal state $C_{u,i}(t)$, updated from prior neighboring influence $PN_{u,i}(t)$, prior personal influence $PP_{u,i}(t)$, and current neighboring influence $CN_{u,i}(t)$. That temporal state is concatenated with a BERT semantic representation $F_u(i)$ of tweet content and social structure, and a softmax layer outputs probabilities over the four emotion regions. The reported result is that E-USIM+ beats E-USIM by about five accuracy points on both rumor and non-rumor threads, with the largest F-measure gains on distress and excitement.","pith_inferences":["A testable extension the paper does not run is a shuffled-timestamps control: randomizing the order of incoming neighbor tweets within each gap while keeping content identical would reveal whether the temporal influence signal is causal or just correlated with content.","The same GRU memory state could be used as a feature for predicting retweet or sharing decisions, since emotional contagion is believed to drive sharing behavior; this would connect emotion prediction directly to spread prediction.","Because the model relies only on message timing and emotion labels, it could be ported to other languages or platforms by re-labeling emotions with the same four quadrants, though the BERT branch would need retraining on new text.","The paper's implicit claim that influence is mostly local—neighbors and self—implies that platform feed design, which controls who appears in a user's stream, could directly change emotional dynamics in rumor threads."],"forward_implications":["Models that ignore a user's own emotional history, such as DeGroot and Voter, are leaving usable signal on the table for rumor emotion prediction.","The four-region emotion mapping is sufficient to capture differences in how rumors and non-rumors propagate emotionally, since the same architecture yields different accuracy and F-measure profiles on the two conversation types.","Conversation tempo matters: the construction of incoming-neighbor windows between a user's outgoing posts is what creates the sequential signal, so the model is sensitive to the timing of messages.","Combining lexical emotion coverage with BERT semantic features lets the model assign emotions to tweets that contain no explicit emotion word."],"supporting_citations":[{"why":"Supplies the 111-event Twitter rumor dataset (60 rumors, 51 non-rumors) used for all experiments and baseline comparisons.","marker":"[3]"},{"why":"Provides the EmoLex lexicon used to expand emotional word coverage when assigning emotion regions to tweets.","marker":"[5]"},{"why":"Defines the circumplex model of affect from which the four emotion regions (R1–R4) are derived.","marker":"[8]"},{"why":"Presents the earlier E-USIM model that E-USIM+ extends; its simplified version is the strongest baseline reported.","marker":"[6]"},{"why":"Shows that emotionally framed content triggers emotional response posts, the contagion premise the sequential influence model operationalizes.","marker":"[2]"}],"fun_headline_variants":["Emotion history is the missing piece in rumor response models","Temporal emotion state improves false-information prediction","E-USIM+ tracks past feelings to sharpen rumor forecasts","Historical mood key to anticipating reactions to fake news"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The model assumes that emotions in tweets a user receives from neighbors between their own posts, together with the user's own emotional history, are what determine the user's next emotion; it does not test whether outside events, topic, or network structure are the real drivers.","fun_headline_variants_meta":{"raw":{"variants":["Emotion history is the missing piece in rumor response models","Temporal emotion state improves false-information prediction","E-USIM+ tracks past feelings to sharpen rumor forecasts","Historical mood key to anticipating reactions to fake news"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000405,"raw_usage":{"total_tokens":2049,"prompt_tokens":829,"completion_tokens":1220,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":445,"completion_tokens_details":{"reasoning_tokens":1157}},"tokens_in":445,"tokens_out":1220,"duration_ms":12903,"temperature":1.0,"reasoning_tokens":1157,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T12:54:13.802397+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Randomly permute the order of incoming neighbor tweets within each user's inter-post gap and retrain E-USIM+; if accuracy stays roughly the same, the temporal ordering the model exploits is not carrying the prediction, and the claimed sequential influence is an artifact of content features.","supporting_citations":[{"cited_title":"PloS one 12(1), e0168344 (2017)","cited_arxiv_id":null,"evidence_quote":"Supplies the 111-event Twitter rumor dataset (60 rumors, 51 non-rumors) used for all experiments and baseline comparisons."},{"cited_title":"In: Proceedings of the NAACL HLT 2010 Workshop on Computational Approaches to Analysis and Generation of Emotion in Text","cited_arxiv_id":null,"evidence_quote":"Provides the EmoLex lexicon used to expand emotional word coverage when assigning emotion regions to tweets."},{"cited_title":"Journal of personality and social psy - chology 39(6), 1161 (1980)","cited_arxiv_id":null,"evidence_quote":"Defines the circumplex model of affect from which the four emotion regions (R1–R4) are derived."},{"cited_title":"In: Proceedings of the 18th Inter - national Conference on Advances in Mobile Computing & Multimedia","cited_arxiv_id":null,"evidence_quote":"Presents the earlier E-USIM model that E-USIM+ extends; its simplified version is the strongest baseline reported."},{"cited_title":"In: Proceedings of the 15th ACM Web Science Conference 2023","cited_arxiv_id":null,"evidence_quote":"Shows that emotionally framed content triggers emotional response posts, the contagion premise the sequential influence model operationalizes."}],"review_version":1}