CrisisKD: Five-Stage Knowledge Distillation for Aspect-Level Sentiment and Emotion Analysis in Crisis Discourse

Marko Haralović; Onat Akca; Salih Eren Yücetürk; Minsi Li; Mariët Theune

cs.CLSep 4, 2026

CrisisKD: Five-Stage Knowledge Distillation for Aspect-Level Sentiment and Emotion Analysis in Crisis Discourse

Authors: Marko Haralović, Onat Akca, Salih Eren Yücetürk, Minsi Li, Mariët Theune

Organizations: Faculty of Electrical Engineering, Mathematics and Computer Science, University of Twente · Faculty of Electrical Engineering and Computing, University of Zagreb · Faculty of Behavioural, Management and Social Sciences, University of Twente

Abstract

Identifying the target of emotional words or phrases in crisis situations, especially health-related ones, is important for understanding public concerns across cultural and linguistic contexts. We propose CrisisKD, a five-stage teacher--student knowledge distillation framework for aspect-level sentiment and emotion analysis on unannotated social media data. A teacher LLM generates aspect-level labels and reasoning traces that supervise a smaller student model across aspect extraction, syntactic parsing, opinion extraction, sentiment classification, and emotion classification. Using this framework, we construct and release a dataset containing 50,615 aspect-level labels, together with the annotation and fine-tuning scripts as open-source resources. The resulting student supports end-to-end ABSA and emotion detection at substantially lower inference cost than the teacher. On a manually annotated 500-tweet gold set, the 5-task Qwen2.5-7B student improves over the untuned model by 7.9 F1 points on aspect extraction, 17.0 points on emotion accuracy, and 6.5 points on sentiment accuracy. On the external ABEA benchmark, CrisisKD improves the same-model Qwen2.5-7B ICL baseline by 2.8 F1 points on ATE and 3.8 F1 points on joint ATE+AEC.

Figures & tables

Panel of Figure 1 — Figure 1: System overview of the proposed teacher–student framework. Each input sentence is parsed by spaCy into CoNLL-U format and, together with the raw sentence, is fed to a frozen teacher LLM ( $\theta_{T}^{\ast}$ ). The teacher produces multi-task reasoning traces $R_{1}$ – $R_{5}$ , which serve as knowledge-distillation targets for the student LLM (LLaMA-3-8B-Instruct or Qwen2.5-7B-Instruct, fine-tuned with LoRA). Task-specific generative losses $\mathcal{L}_{1}\ldots\mathcal{L}_{5}$ are averaged into $\mathcal{L}_{G}$ .

Model	Scenario	Aspect Extraction			Emotion Cls.				Sentiment Cls.
		Prec.	Rec.	F1	Prec.	Rec.	F1	Acc.	Prec.	Rec.	F1	Acc.
Llama-3-8B-Instruct	No fine-tuning	0.341	0.515	0.411	0.566	0.635	0.515	0.733	0.876	0.909	0.891	0.911
	Fine-tune (3 tasks, w/o opinions and syntax)	0.654	0.643	0.648	0.699	0.656	0.669	0.850	0.935 ∗	0.912 ∗	0.920 ∗	0.929
	Fine-tune (4 tasks, w/o opinion extraction)	0.653	0.626	0.639	0.710	0.640	0.649	0.836	0.933	0.906	0.918	0.927
	Fine-tune (4 tasks, w/o syntactic parsing)	0.629	0.646	0.638	0.724 ∗	0.657	0.675	0.852	0.933	0.899	0.915	0.926
	Fine-tune (5 tasks)	0.649	0.622	0.635	0.698	0.640	0.638	0.858	0.933	0.906	0.918	0.927

Table 1: Student–teacher fidelity across training scenarios. We report precision, recall, F1, and accuracy for sentiment and emotion classification, and precision, recall, and F1 for aspect extraction for each model against teacher outputs. Bold values indicate the best performance for a model, and ∗ denotes the best performance overall.

Task	Model	Acc.	F1
Aspect Extraction	Qwen2.5-7B Base	–	0.549
	Qwen2.5-7B FT (3)	–	0.583
	Qwen2.5-7B FT (5)	–	0.628
	Teacher: Qwen2.5-32B	–	0.709
Emotion Classification	Majority class	0.551	0.051
	Uniform random	0.305	0.077

Table 2: Performance on the manually annotated gold set. We report accuracy and macro-F1 where applicable. Majority-class and uniform-random baselines are included for the classification tasks.

Task	Model	Rec.	Prec.	F1
ATE	Qwen3-14B ICL	0.468	0.427	0.447
	Qwen2.5-7B ICL baseline	0.303	0.529	0.385
	Ours: Qwen2.5-7B	0.359	0.486	0.413
ATE + AEC	Qwen3-14B ICL	0.358	0.327	0.342
	Qwen2.5-7B ICL baseline	0.179	0.313	0.228
	Ours: Qwen2.5-7B	0.231	0.313	0.266

Table 3: Performance on the ABEA test set. ATE denotes Aspect Term Extraction, AEC denotes Aspect Emotion Classification, and ICL denotes in-context learning. We compare our Qwen2.5-7B model against a same-size ICL baseline and a larger Qwen3-14B ICL model as an upper-bound reference.

Emotion	Top aspects (log-enrichment)
Proud	doctor (4.04), people (1.69), vaccine (1.48), quarantine (1.47), shopping (1.45)
Trustful	doctor (3.24), hand sanitizer (2.53), World Health Organization (2.52), mask (2.38), government response (1.89)
Optimistic	shopping (3.14), health (1.75), cure (1.74), food (1.68), price (1.66)
Thankful	doctor (3.13), health (2.48), news (2.26), grocery/supermarket/store (1.98), shopping (1.82)
Empathetic	doctor (3.06), people (2.43), food (2.15), health (2.10), shopping (1.78)
Hopeful	cure (3.04), vaccine (2.48), news (1.48), shopping (1.30), world (1.08)

Table 4: Top aspects per emotion ranked by log-enrichment, defined as

\log\!\left(P(\text{emotion}\mid\text{aspect})/P(\text{emotion})\right)

, over the top 50 aspects (minimum support

n\geq 200

). Positive values indicate that the emotion is more common for that aspect than in the dataset overall, negative values indicate underrepresentation, and 0 indicates baseline-level prevalence. Rows are ordered by the strongest top-ranked log-enrichment value in each emotion category.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 2 — Figure 2: Reasoning example for aspect-based sentiment analysis and emotion detection on a single sentence via our teaching framework.

Panel of Figure 3 — Figure 3: Example reasoning traces for two COVID-19 tweets demonstrating the complete annotation pipeline. Left: sarcastic reference requiring contextual sentiment disambiguation. Right: anxious expression of genuine health concern with negative sentiment.

Figure 4: Illustrative reconstruction of a real tweet. Display name, profile image, username, timestamp, and engagement metadata are fictionalized. Reasoning trace across multiple aspects for a COVID-19 tweet. The model identifies two distinct aspects and performs the complete annotation pipeline for each, resulting in shared negative sentiment but divergent emotions: angry toward media censorship versus scared regarding the outbreak itself.

Metric	Value
Total sentences	21,937
Total aspects	50,615
Average aspects per sentence	2.307

Appendix

Table 5: Corpus summary based on Qwen2.5-32B annotations.

Sentiment	Count	Percent
Neutral	26,249	51.87%
Negative	18,288	36.13%
Positive	6,078	12.01%

Appendix

Table 6: Aspect-level sentiment distribution (

N=50{,}615

Emotion	Count	Percent
Neutral	21,608	42.70%
Annoyed	10,488	20.73%
Anxious	8,877	17.54%
Hopeful	1,640	3.24%
Scared	1,552	3.07%
Empathetic	1,436	2.84%

Appendix

Table 7: Aspect-level emotion distribution (

N=50{,}615

). The no emotion label indicates that the teacher did not assign an emotion to the aspect.

Figure 5: Emotion distribution across the 50 most frequent aspects. Each bar represents the normalized distribution of emotion categories for a given aspect.

Pair	Aspect (F1)	Sent.	Emo.
A1–A3	0.7493	0.8635	0.8229
A1–A4	0.7147	0.8289	0.7909
A2–A4	0.7386	0.7676	0.7042
A2–A3	0.7101	0.8516	0.8438
Mean	0.7282	0.8279	0.7905

Appendix

Table 8: Inter-annotator agreement (aspect F1, sentiment accuracy, and emotion agreement).

Aspect-Based Sentiment Analysis

Aspect-Based Sentiment Analysis (ABSA) identifies what people express opinions about and the sentiment or emotion attached to each aspect.

See all 52 papers

Can LLMs Follow the Pulse of a Crisis? Evaluating Crisis Sentiment in Bangladesh's July Uprising
Sep 15, 2026Md. Samiul Alim, Mahir Shahriar Tamim, Tanvir Ahmed Khan +4LLM Evaluation Sentiment Analysis
North South University, Dhaka, Bangladesh · Charles Sturt University, Australia
Empath: Tracing Multi-Level Emotion Dynamics in Crisis Counseling Dialogues
Sep 24, 2026Ziwei Gong, Yuchen Huang, Wen Liang +4Mental Health Counseling Emotional Support Conversation
Columbia University, USA · Barnard College, USA
LLM-guided Semi-Supervised Approaches for Social Media Crisis Data Classification
May 8, 2026Jacob Ativo, Bharaneeshwar Balasubramaniyam, Anh Tran +4Social Media Analysis Semi-Supervised Learning
Department of Computer Science University of Illinois at Chicago

Sep 15, 2026J/K move · Enter open · S save

Can LLMs Follow the Pulse of a Crisis? Evaluating Crisis Sentiment in Bangladesh's July Uprising

Md. Samiul Alim, Mahir Shahriar Tamim, Tanvir Ahmed Khan, Sharjil Khan+3

North South University, Dhaka, Bangladesh · Charles Sturt University, Australia

LLM Evaluation Sentiment Analysis Low-Resource Language Processing Language Model Robustness

Open paper AskPDF arXiv

Crisis sentiment analysis is especially challenging for low-resource languages such as Bangla, where language, context, and public reaction shift rapidly. We introduce UNRESTSENT200K, a Bangla crisis sentiment dataset with approximately 200K Facebook and YouTube comments from the July-August 2024 Bangladesh uprising. The dataset covers five event-aligned phases, from early escalation and internet blackout to regime transition and a later flood crisis. Each comment is linked to its parent post, enabling evaluation with and without discourse context. All comments are annotated through a fully human process involving 14 native Bangla-speaking annotators and senior validation, achieving substantial agreement (kappa = 0.73, alpha = 0.71) and 94.2% blind-audit agreement. We benchmark fine-tuned encoders, prompted LLMs, and LoRA-tuned LLMs. Results show that parent-post context consistently improves performance, while temporal shift across phases causes large performance drops. Strong LLMs perform well, but still struggle with sarcasm, implicit political references, and phase-dependent meaning. UNRESTSENT200K provides a benchmark for studying context-aware and temporally robust sentiment analysis in low-resource crisis discourse. UNRESTSENT200K is available at https://sami0055.github.io/UNRESTSENT200K/

Abstract

Aspect-Based Sentiment Analysis

Explore similar work

Can LLMs Follow the Pulse of a Crisis? Evaluating Crisis Sentiment in Bangladesh's July Uprising

Empath: Tracing Multi-Level Emotion Dynamics in Crisis Counseling Dialogues

LLM-guided Semi-Supervised Approaches for Social Media Crisis Data Classification