acceptodds
Under review as a conference paper at ICLR 2027

Joint Neighbor Semantic Modeling for Mitigating Over-squashing in Heterogeneous Social Event Detection

Abstract

Social Event Detection (SED) aims to identify real-world events from social media streams, where event semantics are distributed across heterogeneous interactions among messages, users, entities, and temporal relations. Graph Neural Networks (GNNs) have become a dominant paradigm for SED by aggregating multi-hop neighborhood information. However, as the receptive field expands, exponentially growing neighborhood evidence must be compressed into fixed-dimensional node representations, leading to over-squashing and the loss of critical event semantics. Existing approaches primarily mitigate over-squashing through structural rewiring or topology augmentation, while paying limited attention to how complementary semantic evidence should be organized before message propagation. In this work, we attribute over-squashing to a semantic bottleneck induced by independent neighbor aggregation, in which complementary event cues are fragmented, diluted, and ultimately lost during message passing. To address this problem, we propose **J**oint **N**eighbor **S**emantic **M**odeling (**JNSM**), a semantic bottleneck-aware framework that explicitly organizes complementary neighbors into target-specific semantic units before propagation. JNSM first performs dual neighbor selection to construct candidate pairs from semantically close and distant neighbors. It then evaluates each candidate using a target-centered semantic gain criterion that balances semantic contribution, redundancy suppression, and cancellation correction. High-gain neighbor pairs are selected as joint semantic units and encoded into learnable semantic unit representations that capture complementary interactions and target-aware context. Finally, JNSM performs semantic-driven local rewiring and semantic unit propagation, allowing complementary event evidence to be better preserved and integrated into the target node. Extensive experiments on three heterogeneous SED benchmarks demonstrate that JNSM achieves strong and robust performance across diverse social-media graphs. Further analyses show that JNSM better preserves event evidence in semantically diverse neighborhoods, providing empirical support for semantic organization before propagation as an effective approach to mitigating over-squashing in heterogeneous SED.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Joint Neighbor Semantics, Over-squashing, Semantic Bottleneck, Social Event Detection

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_t9VxOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Joint Neighbor Semantic Modeling (JNSM) to mitigate the semantic bottlenecks associated with oversquashing for heterogeneous social event detection. Prior to messagepassing propagation, the method assembles complementary neighbors into joint semantic units via dual neighbor selection, targetcentered semantic gain evaluation, learnable unit representation, and semanticdriven local rewiring.

Critical strengths

The paper provides an interesting semantic perspective on over-squashing, emphasizing the importance of preserving complementary information across neighboring nodes. The proposed framework is logically structured, and its combination of semantic gain evaluation, joint representation learning, and local rewiring offers a reasonable approach to improving neighborhood aggregation. JNSM achieves competitive results across three benchmarks, outperforming its backbone on most evaluation metrics. The improvements in high-diversity neighborhoods are particularly encouraging. The paper also includes component ablations, parameter sensitivity studies, and runtime comparisons, providing useful empirical support for the proposed design.

Critical weaknesses

Although the integration is reasonable, the paper does not sufficiently clarify how the proposed approach differs fundamentally from existing semantic-aware aggregation and neighbor interaction methods. Although the ablation experiments support its empirical usefulness, the theoretical relationship between this criterion and actual semantic complementarity remains unclear.

Questions for authors

Could the authors please provide more direct evidence demonstrating that JNSM mitigates oversquashing? Beyond the downstream performance observed in ablation experiments, could the authors further validate whether the semantic gain in Eq. 9 is capable of identifying genuinely complementary neighbor pairs?

1 comment

W1-1: limited innovation and distinction from existing semantic-aware aggregation and neighbour interaction methods.

JNSM's core novelty lies in target-specific joint semantic unit construction, integrating three complementary mechanisms: (1) target-centered semantic gain to identify relevant, non-redundant neighbor pairs; (2) adaptive mean-gain thresholding to select informative pairs according to target-specific neighborhood semantics; and (3) residual unit encoding to preserve and integrate complementary interactions before aggregation. Unlike GraphAIR's parallel convolutions for neighbor interactions, JNSM explicitly constructs complementary semantic units for joint propagation. Unlike FoSR and LASER, which rewire existing nodes, JNSM locally connects newly constructed units to their targets. We will clarify these distinctions in the revision.

W1-2: unclear theoretical relationship between semantic gain and actual semantic complementarity.

Semantic gain is a geometrically motivated heuristic, not a theoretical guarantee, informed by the information-theoretic motivation in Appendix B.1. It combines midpoint proximity for target relevance, redundancy suppression for complementary evidence, and cancellation correction against spurious alignment. Together, these factors favor target-relevant, non-redundant pairs. We will clarify this rationale in the revised manuscript.

Q1: direct evidence of over-squashing mitigation.

We thank the reviewer for this question. We conducted a semantic recovery experiment using linear probes on frozen representations to recover the mean and standard deviation of PCA-projected SBERT features from each target's same-event 1- and 2-hop neighbors, excluding the target. All methods use identical data splits and evaluation procedures.

RepresentationValidation Recovery MSE
Raw SBERT0.23225
LASER0.21730
JNSM (independent)0.17845
JNSM (joint+keep distant)0.17575

JNSM (joint + keep distant) achieves the lowest recovery error, suggesting a greater recoverability for original neighborhood semantics from the target final representations. We prioritize semantic recovery as a probing-based diagnostic of information preservation [2, 3], rather than Jacobian sensitivity, which primarily measures source-to-target deviation influence along graph connections [1].

References:

[1] Di Giovanni et al. (ICML 2023). On Over-Squashing in Message Passing Neural Networks: The Impact of Width, Depth, and Topology.

[2] Hewitt and Liang (EMNLP-IJCNLP 2019). Designing and Interpreting Probes with Control Tasks.

[3] Hewitt et al. (EMNLP 2021). Conditional Probing: Measuring Usable Information Beyond a Baseline.

Q2: Can semantic gain identify genuinely complementary neighbor pairs in Eq. 9?

To directly validate Eq. (9), we introduce Incremental Evidence Coverage (IEC), an independent metric measuring additional event-relevant keywords jointly covered by a neighbor pair beyond its more informative individual neighbor:

\frac{ |K_v\cap(T_i\cup T_j)| -\max\left(|K_v\cap T_i|,\ |K_v\cap T_j|\right) }{ |K_v| }$$ Here, (K_v) contains TF-IDF-derived event keywords from reference messages, excluding target keywords, while (T_i,T_j) denote neighbor keyword sets. **Pair-level Semantic Complementarity Validation on Twitter2012.** | Gain Group | Selection Criterion | Incremental Evidence Coverage (%) | Same-event Pair Rate (%) | |:---|:---|---:|---:| | Low Gain | $$0 \leq G \leq 0.5\bar{G}_v$$| 0.42 | 36.61 | | Mid Gain | $0.5\bar{G}_v < G \leq \bar{G}_v$| 0.81 | 55.48 | | **High Gain** |\$ G > \bar{G}_v\$ | **3.54** | **83.93** | High-gain pairs achieve 3.54% IEC, compared with 0.81% and 0.42% for mid- and low-gain pairs. The high–low difference is 3.12 percentage points (95% bootstrap CI: [0.45, 6.73]), with a positive mean within-target Spearman correlation (ρ = 0.428). Based on 102 evaluated pairs across eight matched target–relation groups, these findings support semantic gain's ability to identify complementary event evidence, while IEC remains a lexical proxy rather than a theoretical guarantee.
Reviewer_hSUuOpenReview ↗
Rating
2
Confidence
4
Informativeness
1(−1)
Presentation
3
Soundness
1(−1)
Read review
Paper summary

The paper proposes Joint Neighbor Semantic Modeling (JNSM) for heterogeneous social event detection (SED). JNSM partitions neighbors according to semantic distance, selects complementary neighbor pairs, and encodes them into joint semantic units. These units are propagated through local rewiring and GraphSAGE aggregation to reduce semantic information loss. The method is evaluated on three datasets under an offline supervised setting.

Critical strengths
  • The observation that event evidence is distributed across multiple messages is reasonable. Explicitly modeling complementary neighbor interactions before aggregation is intuitively motivated.
  • The framework is clearly organized, covering neighbor selection, pair scoring, joint encoding, and message propagation.
  • The experiments include component ablations and several design analyses. Results on semantically diverse neighborhoods provide some support for the effectiveness of the proposed neighbor organization strategy.
Critical weaknesses
  1. Recent SED research increasingly focuses on open-world detection and unseen-event discovery, while this paper still adopts a dated closed-set setting. To my knowledge, prompt-based LLM question answering can already achieve strong performance in closed-set SED, leaving the need for the proposed method unclear.

  2. The experiments omit direct BERT/SBERT clustering baselines and recent methods such as RagSEDE, HyperSED, and LSED. Without these comparisons, the benefit of the proposed graph modeling remains unclear.

  3. Improvements over GraphSAGE are small, with lower ARI on Twitter2012. JNSM also trails GraphSAGE on the larger Twitter2012 graph and GIN on French across all three metrics.

Questions for authors

See Weaknesses

1 comment

W1-1: Limited motivation given advances in LLM-based SED.

We appreciate the reviewer's perspective on recent advances in LLM-based SED. However, JNSM focuses on an aggregation-level semantic bottleneck mitigation that is distinct from, yet complementary to, LLM-based text representation learning. While LLM embeddings capture rich individual message semantics, modeling complementary interactions across interconnected messages may still require additional retrieval and cross-message modeling. JNSM therefore constructs joint semantic units before aggregation to preserve distributed semantic evidence. As shown in Table 2 above, comparisons with and without unified LLM augmentation demonstrate JNSM's competitiveness without LLM enhancement, while revealing dataset-dependent effects when augmentation is applied. These findings support the complementary role of joint semantic modeling alongside advanced text representations.

W1-2: Limited offline research significance and applicability to unseen-event discovery.

JNSM's significance lies in mitigating aggregation-level semantic bottlenecks beyond overall detection accuracy. The offline supervised setting just enables controlled evaluation of this mechanism. To assess applicability to evolving graphs, we further evaluate JNSM on incremental Twitter2012 streams (Table 3 above), achieving average improvements over GraphSAGE of 1.07, 1.16, and 1.65 percentage points in NMI, AMI, and ARI, respectively. These results support JNSM's applicability to incremental SED, while fully online and open-set unseen-event discovery remain future work.

W2: Missing direct BERT/SBERT clustering baselines and recent SED methods.

We thank the reviewer for suggesting broader comparisons. We now include SBERT clustering, HISEvent, HyperSED, RPLM-SED, and RagSEDE under a consistent evaluation protocol (Table 1 above). To compare fairly with LSED, we additionally evaluate SBERT, LSED, and JNSM with and without unified LLM enhancement (Table 2 above). JNSM achieves the highest NMI/AMI on Twitter2012 and all three metrics on CrisisLexT in Table 1, while RPLM-SED leads on French. Under unified text enhancement, JNSM outperforms LSED without LLM augmentation across all datasets but shows dataset-dependent performance with augmentation. These results demonstrate JNSM's competitiveness with LLM-based text enhancement. We will incorporate these comparisons into the revision.

W3: Marginal and inconsistent gains over GraphSAGE and other GNNs.

JNSM's effectiveness should be assessed in terms of both detection performance and semantic-bottleneck mitigation, as its primary objective is to preserve complementary neighborhood information during aggregation. In high-semantic-diversity neighborhoods, JNSM improves ARI over GraphSAGE by 8.42 and 3.97 percentage points on Twitter2012 and CrisisLexT, respectively (Section 4.3). While JNSM trails GraphSAGE on larger Twitter2012 and GIN on French, it improves all three metrics on larger CrisisLexT (Section 4.5). These results highlight JNSM's benefits in addressing semantic bottlenecks alongside its dataset-dependent performance.

W4: insufficient disentanglement of joint semantic modelling and neighbour filtering.

We conducted controlled ablations to disentangle joint semantic modeling from neighbor selection and noise filtering. Compared with independent propagation of the same selected neighbors, joint modeling improves NMI, AMI, and ARI by 0.70, 1.04, and 3.28 percentage points, respectively. It also outperforms independent propagation when unpaired distant neighbors are retained, demonstrating benefits beyond neighbor selection and noise filtering alone.

Ablation study disentangling joint semantic modelling and neighbor filtering on Twitter2012 (%).

VariantNMIAMIARI
Independent propagation (same selected neighbors)93.8890.9983.52
Joint modelling + unpaired distant neighbors94.5492.0085.95
Independent propagation (all neighbors)93.6290.6283.37
JNSM (Full)94.5892.0386.80
Reviewer_vPq6OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper attempts to address the issue that complementary event semantics are easily lost during neighborhood aggregation due to a semantic bottleneck associated with over-squashing in heterogeneous social event detection. The authors propose the JNSM method, which filters and combines semantically complementary neighboring nodes for information propagation, thereby reducing information loss and improving event detection performance.

Critical strengths
  1. The concept of "semantic bottleneck" is proposed to characterize information loss from the perspective of semantic complementarity, providing a new angle for addressing the over-squashing problem in graph neural networks.
  2. The JNSM model performs well in semantically diverse neighborhoods, and achieves competitive results on all three datasets.
  3. The experiments are relatively comprehensive, including ablation studies, parameter sensitivity analysis, and efficiency testing.
Critical weaknesses
  1. The paper lacks direct evidence to prove that JNSM can alleviate the over-squashing problem in graph neural networks.
  2. Compared with GraphSAGE, JNSM achieves performance improvements mostly below 1 percentage point in the main experiments, with even worse results in some cases.
  3. The method mostly combines existing neighbor selection, pairing and graph rewiring techniques with limited innovation; similar ideas have been explored in GraphAIR, LASER and other prior works.
  4. The SED-Aug baseline was evaluated without its LLM-based augmentation module, and there is a lack of direct comparison with other recent related methods.
  5. Experiments only use static data and random partitioning, lacking validation on new events and dynamic social media data.
Questions for authors
  1. Can you provide more direct experimental evidence to prove that JNSM indeed alleviates over-squashing, rather than just improving detection performance?
  2. Compared with GraphSAGE, JNSM achieves relatively small performance improvements, with some indicators even decreasing. Can the authors verify the reliability of these improvements through significance testing?
  3. What is the core innovation of JNSM compared to existing neighbor interaction modeling methods such as GraphAIR? Can you provide a direct experimental comparison?
  4. Why is the LLM-based augmentation component not used in the SED-Aug baseline? Does JNSM still have advantages compared with its full version?
  5. Can JNSM effectively identify unseen new events and adapt to dynamic social media data under time-based split or cross-event testing?
1 comment

W1&Q1: lack of direct evidence that JNSM alleviates over-squashing in GNNs.

We thank the reviewer for this concern. We conducted a semantic recovery experiment comparing raw SBERT, LASER, JNSM with independent propagation, and JNSM retaining unpaired distant neighbors. Linear probes were trained on frozen representations to recover the mean and standard deviation of PCA-projected SBERT features from each target’s same-event 1- and 2-hop neighbors, excluding the target itself, under identical data splits and validation settings.

RepresentationValidation Recovery MSE
Raw SBERT0.23225
LASER0.21730
JNSM (independent)0.17845
JNSM (joint+keep distant)0.17575

JNSM (joint+keep distant) achieves the lowest recovery error, suggesting a greater recoverability for original neighborhood semantics from the target final representations. We prioritize semantic recovery as a probing-based diagnostic of information preservation [2, 3], rather than Jacobian sensitivity, which primarily measures source-to-target deviation influence along graph connections [1].

References:

[1] Di Giovanni et al. (ICML 2023). On Over-Squashing in Message Passing Neural Networks: The Impact of Width, Depth, and Topology.

[2] Hewitt and Liang (EMNLP-IJCNLP 2019). Designing and Interpreting Probes with Control Tasks.

[3] Hewitt et al. (EMNLP 2021). Conditional Probing: Measuring Usable Information Beyond a Baseline.

W2: JNSM’s marginal and inconsistent improvement over GraphSAGE.

JNSM's effectiveness should be evaluated from two perspectives: overall detection performance and semantic-bottleneck mitigation, as its primary objective is to preserve complementary neighborhood information during aggregation rather than solely optimize clustering accuracy. As shown in Section 4.3, JNSM improves ARI over GraphSAGE by 8.42 and 3.97 percentage points in high-semantic-diversity neighborhoods on Twitter2012 and CrisisLexT, respectively. These findings highlight the benefits of joint semantic modeling in addressing aggregation-level semantic bottlenecks and offer complementary evidence of over-squashing mitigation.

W3&Q3: limited innovation for combining existing neighbour selection, pairing, and graph rewiring.

JNSM’s core novelty lies in target-specific joint semantic unit construction, integrating target-centered semantic gain, adaptive mean-gain threshold, and residual unit encoding. Unlike GraphAIR, which models neighbor interactions through parallel graph convolutions, JNSM explicitly constructs and propagates complementary semantic units. Unlike FoSR and LASER rewirings introduce edges among existing nodes, JNSM locally connects newly constructed semantic units to their targets.

W4-1&Q4: SED-Aug without LLM-based augmentation.

As clarified in Appendix C.1.2, we initially excluded SED-Aug's LLM augmentation to isolate graph modelling effects under consistent inputs. To address this concern, we now compare SBERT, SED-Aug, and JNSM with and without unified LLM-based text enhancement (Table 2 above). LLM enhancement consistently benefits SBERT but has mixed effects on graph-based methods. Without LLM augmentation, JNSM outperforms SED-Aug on Twitter2012 and CrisisLexT across all metrics. With augmentation, JNSM remains superior on Twitter2012 but underperforms SED-Aug on CrisisLexT and French. These findings highlight the dataset-dependent effects of LLM augmentation while supporting JNSM's distinct contribution to joint semantic neighborhood aggregation.

W4-2: insufficient comparisons with recent SED methods.

We initially prioritized graph message-passing baselines to isolate JNSM's contribution. To broaden the comparison, we additionally evaluate HISEvent, HyperSED, RPLM-SED, SED-Aug, and RagSEDE under a consistent protocol (Table 1 above). JNSM achieves the highest NMI/AMI on Twitter2012 and all three metrics on CrisisLexT, while RPLM-SED leads on French. We will incorporate these comparisons in the revised manuscript.

W5&Q5: lack of evaluation on dynamic social media data, only static experiments.

We extend JNSM to incremental SED on Twitter2012 (Table 3 above). JNSM outperforms GraphSAGE across all three streams and metrics, with average gains of 1.07, 1.16, and 1.65 percentage points in NMI, AMI, and ARI, respectively. These findings support its applicability to evolving social graphs, although fully online detection remains future work.

Q2: Statistical Significance of JNSM vs. GraphSAGE

We use stratified data splits and five fixed random seeds to ensure comparable evaluation conditions and assess variability. Chi-square tests confirm no detectable label-distribution differences across splits on Twitter2012 (p = 0.9978), CrisisLexT (p = 0.9997), and French (p ≈ 1.0). These tests establish split comparability, not statistical significance, between JNSM and GraphSAGE. We will clarify these procedures in the revised manuscript.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Supplement Experiment Results

Table 1: Comparison with recent SED methods (%).

MethodTwitter2012 NMIAMIARICrisisLexT NMIAMIARIFrench NMIAMIARI
SBERT86.0779.1865.1277.6877.6584.3348.7637.2312.04
HISEvent88.9488.4078.6743.8743.8337.6767.4866.9351.72
HyperSED77.2775.5954.856.896.840.8450.5648.6947.47
RPLM-SED93.0789.7281.7385.1885.1691.5777.2172.2874.41
RagSEDE93.6891.1385.785.142.270.9039.0832.4717.70
JNSM94.4291.8184.7790.0790.0594.8474.5268.9139.27

Table 2: Impact of LLM-Based Text Augmentation on SED Performance (%).

MethodTwitter2012 NMIAMIARICrisisLexT NMIAMIARIFrench NMIAMIARI
SBERT w/o LLM86.0779.1865.1277.6877.6584.3348.7637.2312.04
SBERT w/ LLM91.3187.0674.6178.0378.0085.5362.6354.1022.48
SED-Aug w/o LLM92.5288.8979.1187.7887.7793.3274.6869.1842.35
SED-Aug w/ LLM93.6990.6582.4087.7587.7393.3278.6574.0046.08
LSED w/o LLM94.4491.8785.0283.9283.9190.2165.2457.3823.65
LSED w/ LLM92.8689.4180.4782.7382.7189.3573.6767.8337.12
JNSM w/o LLM94.5892.0386.8089.4089.3994.4672.5266.4331.67
JNSM w/ LLM94.5091.8884.3276.9276.8983.6177.3372.3537.60

Table 3: Incremental SED Performance on Twitter2012 (%).

StreamMetricSBERTDoubleGCNGraphFormerGINReDHANSED-AugPIGCNLASERGraphSAGEJNSM (Ours)
S1NMI58.3367.7568.2659.9362.4069.0969.3062.7669.3869.42
AMI42.9065.4365.9957.2059.7266.8667.0860.1667.1767.22
ARI9.5511.8212.0310.829.6012.2911.3714.0111.7912.70
S2NMI44.1663.1272.1068.5676.3171.4170.3869.2667.4269.69
AMI26.1660.7570.3166.5374.7869.5768.4767.3465.2967.74
ARI6.8011.9324.6628.1066.2219.7618.8444.4213.3816.57
S3NMI50.3367.8973.1261.2564.3373.1873.3164.7372.8873.78
AMI28.9065.4571.0158.2261.5271.0571.1962.0270.7471.72
ARI12.6824.8125.2321.7422.9623.8726.3823.4125.9626.80

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.