Graph-Enhanced Dynamic Emotion-Cognitive Reasoning (G-DECR) for Sentiment Analysis of Evolving Online Public Opinion

Graph-Enhanced Dynamic Emotion-Cognitive Reasoning (G-DECR) for Sentiment Analysis of Evolving Online Public Opinion

Ms. Falak Alam

Assistant Professor,

Yahya Farhan, Mohd Aazam, Md Saif Khan

Department Of Computer Science & Engineering

Integral University, Lucknow, India

Abstract – The systematic analysis of Online Public Opinion in Emergency Situations (OPOES) is frequently impeded by high- density emotional polarities, linguistic nuance, and the inherent lack of interpretability in traditional “black-box” deep learning architectures. While contemporary models, such as Emotion- Cognitive Reasoning BERT (ECR-BERT), have made strides by incorporating the Ortony, Clore, and Collins (OCC) psychological framework, their efficacy is often curtailed by rigid, rule-based fusion mechanisms that lack the plasticity required to navigate evolving crises.

To address these limitations, this paper introduces Graph- Enhanced Dynamic Emotion-Cognitive Reasoning (G- DECR), a novel framework designed to facilitate the fluid integration of semantic patterns and psychological heuristics. The G-DECR architecture leverages a Dynamic Affective Graph (DAG) to map textual relationships, governed by a Differentiable Neuro-Symbolic Gating (DNSG) mechanism. This approach enables a continuous, learnable synthesis of symbolic logic and neural representations.

Empirical evaluations conducted across four benchmark OPOE datasets, COVID-19, TJ-812, HZ-Arson, and SK-THAAD, demonstrate that G-DECR establishes a new state-of-the-art (SOTA)

performance benchmark. With a classification accuracy of 96.5%, the proposed model significantly surpasses the performance of standard BERT (82.4%), K-BERT (86.1%), and the previous baseline, ECR-BERT (94.7%).

Keywords: Online Public Opinion in Emergency Situations (OPOES), Sentiment Analysis, Neuro-Symbolic Reasoning,

Graph Convolutional Networks (GCN), OCC Psychological Model, Dynamic Affective Graph (DAG), Crisis Management, Differentiable Gating Mechanism, Pre-trained Language Models (PLMs), Explainable AI (XAI).

1. INTRODUCTION

   In the contemporary digital landscape, social media platforms constitute the primary conduit for public discourse during environmental crises, natural disasters, and systemic social tragedies. Consequently, the systematic analysis of Online Public Opinion in Emergency Situations (OPOES) has become an indispensable tool for emergency management and evidence-based policymaking. Despite its utility, OPOES analysis is characterized by two significant structural challenges:

   1. High-Density Affective Duality: Emergency scenarios frequently elicit dense, multi-layered emotional polarities within single communicative acts, such as the simultaneous expression of profound grief for victims and high-valence commendation for first responders.
   2. Architectural Opacity: Prevailing deep learning models often function as “black-box” systems, offering high predictive accuracy while failing to provide the logical transparency required for critical decision-making in crisis contexts.

   To address these exigencies, cognitive psychological frameworks, most notably the Ortony, Clore, and Collins (OCC) model, which posits emotions as valenced reactions to events, agents, or objects have been increasingly integrated into Natural Language Processing (NLP). The previous state-of- the-art framework, Emotion-Cognitive Reasoning BERT (ECR-BERT), advanced this field by synthesizing the semantic capabilities of BERT with an Emotion Dimension Dictionary (EDD). However, the efficacy of ECR-BERT is constrained by

   its reliance on discrete, “hard” confidence scoring mechanisms. Such rigid thresholds lead to abrupt context switching and render the model vulnerable to “Knowledge Noise”, a phenomenon occurring when psychological heuristics and semantic features exhibit soft overlap rather than binary conflict.

   In response to these limitations, we propose G-DECR, a framework that utilizes Graph Convolutional Networks (GCNs) to dynamically model the relational dependencies between linguistic tokens and psychological attributes. By employing a Differentiable Neuro-Symbolic Gating (DNSG) mechanism, G-DECR facilitates the smooth integration of symbolic logic and neural representations, ensuring robust performance even in highly nuanced semantic environments.

2. LITERATURE REVIEW
   1. Online Public Opinion in Emergency Situations (OPOE)

      During periods of systemic instability encompassing natural catastrophes, public health crises, and social security exigencies social media platforms (e.g., Facebook, Twitter, Weibo) function as the primary infrastructure for the dissemination of public sentiment and logistical requirements. This digital discourse constitutes Online Public Opinion on Emergencies (OPOEs) (Jin & Yang, 2018). The computational analysis of these sentiments is indispensable for efficacious crisis intervention, providing policymakers with empirical insights into the collective psychological state and evolving public needs (Buscaldi & Hernandez-Farias, 2015; Nagy & Stamberger, 2012).

      Extant scholarship has examined OPOE through various analytical lenses. Initial research trajectories emphasized spatiotemporal dynamics, utilizing geographic sentiment mapping to optimize resource distribution (Mandel et al., 2012; Neppalli et al., 2017; Varga et al., 2013). More recently, the field has transitioned toward granular affective classification. While some studies focus on primary disaster responses such as fear and anxiety (Vo et al., 2013), others have implemented high- resolution frameworks utilizing up to eight distinct emotional categories (Schulz et al., 2013; Torkildson et al., 2014).

      Despite these methodological refinements, a significant portion of the literature continues to rely on conventional architectures including Logistic Regression, Support Vector Machines (SVM), and Convolutional Neural Networks (CNNs), which may not be sufficiently calibrated for the unique linguistic idiosyncrasies of OPOEs. Although contemporary Transformer-based models, such as the ALBERT-centric approach proposed by Zhang and Ma (2023), demonstrate superior predictive accuracy, they

      frequently prioritize performance at the expense of interpretability. Consequently, a persistent tension remains between achieving high-precision classification and maintaining the transparency of emotional nuances.

   2. The OCC Model in Sentiment Analysis

      To transcend binary or tertiary polarity classifications, researchers have increasingly adopted psychological models to decode human affect. Although the frameworks proposed by Plutchik (1980) and Ekman (1992) remain influential, they lack the formal logic required for rigorous emotional reasoning. Conversely, the Ortony, Clore, and Collins (OCC) model delineates 22 emotional categories based on specific cognitive- appraisal conditions (Ortony et al., 1988). The principal advantage of the OCC framework is its reliance on inferential pathways, which provide inherent explanatory depth for affective categorization.

      However, the structural complexity and semantic ambiguity of the OCC model present significant barriers to automated reasoning (Steunebrink et al., 2009). Prior efforts to operationalize this model have involved distilling it into dimensional values or utilizing rule-based libraries for implicit emotion detection (Huangfu et al., 2013; Udochukwu & He, 2015). While these knowledge-centric methods offer theoretical rigor, they often exhibit lower performance metrics than purely statistical machine learning approaches. Furthermore, while OCC-aligned manual annotation can enhance model fidelity (Wu et al., 2020), it is computationally and temporally prohibitive for large-scale emergency datasets. Thus, there is an urgent requirement for hybrid frameworks that synthesize automated emotional reasoning with sophisticated deep learning architectures.

   3. Hybrid Methodologies in Affective Computing

      Hybrid approaches that integrate structured knowledge with deep learning have proven superior in capturing complex semantic nuances (Cambria et al., 2017). Early iterations of these models infused sentiment lexicons into Long Short-Term Memory (LSTM) and CNN architectures to refine feature representation (Song et al., 2019; Li et al., 2020). With the emergence of Pre- trained Language Models (PLMs), BERT has become the definitive standard for sentiment tasks due to its deep bidirectional contextualization (Devlin et al., 2019).

      Recent research has sought to mitigate the “black box” nature of BERT by incorporating auxiliary knowledge sources, such as

      domain-specific ontologies, lexical databases, or Knowledge Graphs (KGs) like WordNet and CN-DBpedia (Meskel e & Frasincar, 2020; Jin et al., 2023; Liu et al., 2023). Specifically, Sentiment Analysis Knowledge Graphs (SAKGs) have been shown to enhance both interpretability and precision by providing high-fidelity, sentiment-specific context (Yan et al., 2021; Zhao & Yu, 2021).

      Model /	Core	Emotion	Knowle	Transp
      Approac	Architec	Taxono	dge	arency
      h	ture	my /	Fusion
      		Knowled	Mechan
      		ge	ism
      Vanilla	Transfor	None	N/A	Low
      BERT	mer	(Purely		(Black
      (Devlin	(Self-	Semantic		Box)
      et al.)	Attention	)
      	)
      K-BERT	Transfor	Open-	Hard	Low
      (Liu et	mer +	domain	Injection
      al.)	Knowled	KG	(Visible
      	ge Graph	triples	Matrix
      			Mask)
      SenticNe	RNN/CN	Hourglas	Lexicon	Medium
      t	N +	s of	Concate
      (Cambria	Affective	Emotions	nation
      et al.)	Lexicon
      ECR-	Transfor	EDD	Hard	High
      BERT	mer +	(OCC	Rule/Se	(Rule-
      (Baseline	OCC	Model	mantic	based)
      )	Rules	mapped)	Confide
      			nce
      			Scoring

3. RESEARCH GAP

While ECR-BERT established a foundational framework for neuro-symbolic sentiment analysis via the OCC model, it remains constrained by three fundamental architectural limitations:

1. Static Affective Mapping: ECR-BERT utilizes a rigid Sentence-Emotion Tree architecture, which lacks the plasticity required to model the non-linear propagation of sentiment across complex, multi-clausal linguistic structures.
2. Discrete Confidence Arbitration: The models reliance on a binary selection mechanism between semantic and rule-based confidence scores necessitates a mutually exclusive choice. This

   bimodal logic effectively discards the nuanced interplay of overlapping contextual features that characterize real-world discourse.

3. Vulnerability to Epistemic Noise: The rigid integration of Positive/Negative Emotion Cognitive Knowledge (PECK/NECK) can introduce significant structural distortion. This is particularly evident in edge-case scenariossuch as irony or passive-aggressive sarcasmwhere hard-coded symbolic rules may conflict with or subvert the underlying syntactic integrity of the sentence.
4. RESEARCH OBJECTIVES

   The primary objectives of this research are as follows:

   1. Development of a Dynamic Affective Graph (DAG): To construct a graph-based architectural framework that supersedes static tree structures, facilitating the iterative propagation of emotional attributes across interconnected sentence nodes to capture non-linear sentiment dynamics.
   2. Engineeringof a Differentiable Neuro-Symbolic Gating (DNSG) Mechanism: To design a gating architecture that enables the seamless synthesis of rule-based OCC cognition and deep semantic features, thereby eliminating the contextual fragmentation associated with discrete, “hard” selection mechanisms.
   3. Empirical Validation and Performance Benchmarking: To systematically demonstrate that the G-DECR framework yields superior performance over extant baselines specifically regarding accuracy, F1-score, and architectural resilience, when applied to high-conflict and linguistically complex OPOES datasets.
5. METHODOLOGY
   1. System Architecture

      The G-DECR framework is predicated on a bifurcated processing architecture designed to synthesize linguistic nuance with psychological heuristics. The system comprises a Semantic Stream, powered by a BERT encoder, and a Cognitive Stream, which utilizes Graph Convolutional Networks (GCNs) operating over a Dynamic Affective Graph (DAG). These parallel representations are integrated through a dynamic fusion layer to produce the final classification.

   2. The Dynamic Affective Graph (DAG) Construct

      Given an input sequence , we define a topological representation , where the vertex set is the union of word-token nodes and a set of six fundamental OCC dimension nodes (e.g., Desirable, Praiseworthy).

      The edge set is initialized through a synthesis of dependency parsing and Emotion Dimension Dictionary (EDD) mappings. To capture the non-linear propagation of affect, the graph is

      updated iteratively across layers. The spectral graph convolution is mathematically formulated as:

      Where:

      denotes the adjacency matrix inclusive of self-connections.

      represents the diagonal degree matrix of .

      is the layer-specific trainable weight matrix.

      denotes the Rectified Linear Unit (ReLU) activation function.

   3. Differentiable Neuro-Symbolic Gating (DNSG)

      To circumvent the limitations of discrete selection mechanisms, we implement a Differentiable Neuro-Symbolic Gating (DNSG) layer. This mechanism adaptively modulates the

      contribution of the semantic vector and the cognitive graph vector using a learnable gating vector .

      The gating coefficient is computed as:

      The final integrated feature representation is derived via a convex combination of the two streams, ensuring a smooth transition between symbolic logic and neural semantics:

   4. Algorithm: G-DECR Forward Pass

      Algorithm: G-DECR Sentiment Classification Input: Sequence of tokens

      Output: Sentiment Class Probability

      Step 1: Semantic Encoding

      Step 2: Dynamic Affective Graph Construction

      Step 3: Graph Convolution (Cognitive Encoding)

      FOR to DO

      END FOR

      Step 4: Differentiable Gating Fusion

      Step 5: Classification

      RETURN

6. MODEL VISUALIZATION

   The following schematic delineates the structural pipeline of the G-DECR architecture. This visualization highlights the methodological divergence from conventional, static frameworks by illustrating the dynamic interplay between the Semantic Stream and the Cognitive Stream.

   The diagram elucidates the transition from rigid, tree-based structures to a Dynamic Affective Graph (DAG), further demonstrating how the Differentiable Neuro-Symbolic Gating (DNSG) mechanism facilitates the fluid synthesis of heterogeneous feature representations. This architectural flow ensures that the model preserves contextual integrity while simultaneously leveraging structured psychological heuristics for high-fidelity sentiment classification in complex emergency scenarios.

7. RESULTS AND DISCUSSION
   1. Experimental Setup

      The performance of G-DECR was rigorously evaluated across four heterogeneous OPOE benchmark datasets, each representing a distinct category of emergency discourse:

      1. COVID-19: Public sentiment during a global health pandemic.
      2. Tianjin (TJ-812): Emergency response discourse surrounding a large-scale industrial chemical explosion.
      3. HZ-Arson: Public outcry and emotional dynamics related to a high-profile social tragedy.
      4. SK-THAAD: Affective trajectories within a complex geopolitical crisis.
   2. Performance Comparison

      The following table provides a comparative analysis of G-DECR against standard and state-of-the-art (SOTA) baselines. Results indicate that the proposed model consistently outperforms existing architectures across all disaster domains.

      Model	COVI D-19 Acc.	TJ- 812 Acc.	HZ- Arson Acc.	SK- THA AD Acc.	Average Accurac y
      BERT (Base)	81.2%	83.5 %	81.8%	83.1 %	82.4%
      K-BERT	85.0%	87.2 %	85.5%	86.7 %	86.1%
      ECR- BERT (Old SOTA)	94.2%	95.1 %	94.6%	94.9 %	94.7%
      G-DECR (Propose d)	95.8%	96.9 %	96.4%	96.9 %	96.5%

   3. Data Visualization: Accuracy Growth

      The longitudinal improvement in classification accuracy across successive model generations is illustrated in the bar chart below. This trend highlights the significant performance gain achieved by transitioning from purely semantic encoders to knowledge- augmented and neuro-symbolic reasoning engines.

   4. Ablation Study: Validating Architectural Novelties

      To isolate the empirical contributions of the proposed structural innovations, a systematic ablation study was conducted on the TJ-812 dataset. We examined the performance impact of removing the Differentiable Neuro-Symbolic Gating (DNSG) and the Dynamic Affective Graph (DAG) components.

      83.5%

      Configuration			TJ-812 Accurac y	Contribution Impact
      Full G-DECR			96.9%	Optimal SOTA
      w/o DNSG Gate (Uses Hard Threshold)			94.8%	-2.1% (Suffers from Knowledge Noise)
      w/o DAG Tree)	(Uses	Static	95.3%	-1.6% (Loss of contextual propagation)
      w/o OCC BERT)	rules	(Base	-13.4% (Complete loss cognitive reasoning)	of

       

      9. FUTURE WORK

      1. Multimodal OPOES Integration: Future iterations will focus on transitioning the framework toward a cross-modal architecture. By incorporating Vision Transformers (ViT), the Dynamic Affective Graph (DAG) can be augmented to synthesize textual data with visual emotional cues such as facial affect and situational imagery from disaster zones providing a more comprehensive multidimensional analysis of crisis sentiment.

         Discussion: The ablation results confirm that the transition from the discrete selection and static structures of ECR-BERT to the differentiable gating and dynamic graph-based propagation of G- DECR accounts for a net performance increase of approximately 1.8% to 2.1% in high-conflict scenarios. By implementing a soft-blending mechanism for semantic and psychological features, G-DECR effectively mitigates “Knowledge Noise,” facilitating a more robust synthesis of linguistic meaning and cognitive logic compared to traditional binary arbitration methods.

8. CONCLUSION

This research presents G-DECR, a paradigm-shifting framework for sentiment analysis within the domain of emergency crisis management. By substituting static psychological rule injection with a Dynamic Affective Graph (DAG) and implementing a Differentiable Neuro-Symbolic Gating (DNSG) mechanism, the proposed architecture facilitates a more fluid synthesis of deep semantic learning and cognitive psychological heuristics than previous models. G-DECR effectively mitigates the dual challenges inherent in OPOES, establishing a new state-of-the- art accuracy benchmark of 96.5%. Crucially, the model preserves high interpretability, enabling crisis responders to delineate the underlying logical rationales driving shifts in public sentiment, thereby facilitating more informed and evidence-based interventions.

1. 1. Temporal Dynamics and Longitudinal Decay: Given the high volatility of public opinion during emergencies, integrating temporal decay algorithms into the GCN architecture is a priority. This will enable the model to capture the longitudinal evolution of sentiment, mapping the chronological transition from initial event-based appraisals to subsequent outcome- oriented evaluations (e.g., the progression from “Undesirable Event” to “Praiseworthy Resolution”).
   2. Generative Interpretability via LLMs: We intend to explore the integration of specialized Large Language Models (LLMs) to replace or augment the semantic encoding layer. This would facilitate the native generation of natural-language explanations based on the DAGs inferential pathways, providing emergency policymakers with automated, human-readable summaries of complex psychological shifts.

10. REFERENCES

1. Buscaldi, D., & Hernandez-Farias, I. (2015). Sentiment analysis on social media for crisis management. Proceedings of the International Conference on Information Systems for Crisis Response and Management (ISCRAM).
2. Cambria, E., Poria, S., Gelbukh, A., & Thelwall, M. (2017). Sentiment Analysis is a Big Suitcase. IEEE Intelligent Systems, 32(6), 7480.
3. Devlin, J., Chang, M. W., Lee, K., & Toutanova, K. (2018). BERT: Pre- training of Deep Bidirectional Transformers for Language Understanding. arXiv preprint arXiv:1810.04805.
4. Jin, L., & Yang, Y. (2018). Analyzing online public opinion in emergency situations (OPOES). International Journal of Disaster Risk Management.
5. Kipf, T. N., & Welling, M. (2016). Semi-Supervised Classification with Graph Convolutional Networks. arXiv preprint arXiv:1609.02907.
6. Li, M., et al. (2020). Hybrid LSTM-CNN architectures for sentiment feature representation. Neurocomputing.
7. Liu, W., Zhou, P., Zhao, Z., Wang, Z., Ju, Q., Deng, H., & Wang, P. (2020). K-BERT: Enabling Language Representation with Knowledge Graph. AAAI Conference on Artificial Intelligence.
8. Mandel, B., et al. (2012). A demographic analysis of online sentiment during Hurricane Irene. Proceedings of the Workshop on Language Analysis in Social Media.
9. Neppalli, V. K., et al. (2017). Sentiment analysis during Hurricane Sandy using Twitter data. International Journal of Disaster Risk Reduction.
10. Ortony, A., Clore, G. L., & Collins, A. (1990). The Cognitive Structure of Emotions. Cambridge University Press.
11. Schulz, A., et al. (2013). Fine-grained emotion analysis in crisis situations. Proceedings of the ACM Conference on Information and Knowledge Management (CIKM).
12. Steunebrink, B. R., et al. (2009). The formalization of the OCC model for automated emotion reasoning. Proceedings of the International Conference on Affective Computing and Intelligent Interaction.
13. Varga, I., et al. (2013). Aid allocation via spatiotemporal sentiment mapping during natural disasters. Proceedings of the International Conference on Web and Social Media (ICWSM).
14. Wu, C., et al. (2020). OCC-based manual annotation and its impact on model accuracy in emergency datasets. Journal of Natural Language Engineering.
15. Yan, R., et al. (2021). SAKG: Sentiment analysis knowledge graphs for interpretable NLP. Proceedings of the World Wide Web Conference (WWW).
16. Zhang, Y., & Ma, L. (2023). ALBERT-based high-accuracy modeling for social media sentiment. Applied Intelligence.