Emotion-Enhanced Multi-Task ACSA Framework

• The paper presents an emotion-enhanced multi-task ACSA framework that integrates discrete emotional dimensions into sentiment polarity prediction using a unified encoder–decoder approach.
• It employs parallel sentiment and emotion classification heads built on a Flan-T5 architecture, leveraging joint training and VAD-based refinement for consistent labeling.
• Experimental results on benchmark datasets demonstrate improved F1 scores, validating the effectiveness of combining discrete Ekman emotions with continuous VAD information.

Emotion-enhanced multi-task Aspect Category Sentiment Analysis (ACSA) frameworks advance sentiment analysis by integrating discrete emotional dimensions into the sentiment polarity modeling process. The explicit joint modeling of sentiment polarity and category-specific emotions—anchored in Ekman’s six basic emotions and neutral—enables LLMs to capture fine-grained affective cues for each aspect category within text. This paradigm addresses the limitation of traditional ACSA, which typically restricts annotation and prediction to coarse sentiment polarity, thereby neglecting the underlying affective dimensions that shape sentiment expressions. Coupled with a Valence-Arousal-Dominance (VAD)–guided refinement mechanism, the framework ensures emotional label fidelity, elevating both the reliability and the interpretability of aspect-sensitive sentiment predictions (Chai et al., 24 Nov 2025).

1. Model Design and Multi-Task Framework

The core of the emotion-enhanced multi-task ACSA approach is a unified, generative encoder–decoder architecture based on Flan-T5 (3B parameters). The model processes an input sentence $S$ and a predefined set of aspect categories $C = \{c_1, ..., c_m\}$, and employs two parallel heads formulated as different target output sequences:

• Sentiment Polarity Head: Generates ordered category–polarity pairs $(c_i, y_i)$, where $$y_i \in \{\text{positive}, \text{neutral}, \text{negative}\}$$.
• Aspect Emotion Classification Head: Generates ordered category–emotion pairs $(c_i, e_i)$, with $$e_i \in E = \{\text{anger}, \text{disgust}, \text{fear}, \text{joy}, \text{sadness}, \text{surprise}, \text{neutral}\}$$.

Both heads share all encoder–decoder parameters; differentiation arises solely through target serialization. The multi-task design allows simultaneous learning of coarse sentiment and fine-grained emotion, capturing more nuanced affective features associated with each aspect (Chai et al., 24 Nov 2025).

2. Task Specification and Training Objective

The formalization comprises two generative tasks:

• Aspect-Category Sentiment Analysis (ACSA): Predicts $Y^{\mathrm{sen}} = \{ (c_i, y_i) \}$ for all $c_i \in C$.
• Aspect Emotion Classification (AEC): Predicts $Y^{\mathrm{emo}} = \{ (c_i, e_i) \}$ for all $c_i \in C$.

Negative log-likelihood loss is used for both tasks: $$\mathcal{L}_{\mathrm{sen}}(\theta) = -\mathbb{E} \left[ \sum_{t=1}^T \log p_\theta(y_t^{\mathrm{sen}} \mid y_{<t}^{\mathrm{sen}}, t^{\mathrm{sen}}(x)) \right]$$

$$\mathcal{L}_{\mathrm{emo}}(\theta) = -\mathbb{E} \left[ \sum_{t=1}^T \log p_\theta(y_t^{\mathrm{emo}} \mid y_{<t}^{\mathrm{emo}}, t^{\mathrm{emo}}(x)) \right]$$

The joint training objective is

$$\mathcal{L}(\theta) = \alpha\,\mathcal{L}_{\mathrm{sen}}(\theta) + (1-\alpha)\,\mathcal{L}_{\mathrm{emo}}(\theta)$$

where $\alpha \in [0,1]$ balances the two tasks. Empirically, $\alpha = 0.6$ yields optimal F1 on validation data. No task-specific regularization beyond AdamW weight decay is used (Chai et al., 24 Nov 2025).

3. Generative Emotion Description and Labeling

Emotion assignment leverages LLM-driven generative capabilities. Each sentence is decomposed (via GPT-4o-mini prompts) into aspect-focused sub-sentences $\{s_i\}$, each aligned with a specific $(c_i, y_i)$. The emotion generation prompt for each $s_i$ is:

“Given sub-sentence: ‘…’, the aspect category: $<$c_i$>$, the sentiment polarity: $<$y_i$>$. From $$\{\text{anger}, \text{disgust}, \text{fear}, \text{joy}, \text{sadness}, \text{surprise}, \text{neutral}\}$$, select the most appropriate emotion label.”

This process operationalizes category-level emotion annotation, providing explicit emotional context per aspect that complements coarse polarity judgments (Chai et al., 24 Nov 2025).

4. VAD-Based Emotion Refinement and Consistency

To enhance emotion label fidelity, a VAD-based refinement mechanism is introduced:

• Continuous VAD Prediction: For each $s_i$, a DeBERTa model fine-tuned on EmoBank predicts $(\hat v_i, \hat a_i, \hat d_i) \in [1,5]^3$, which are normalized to $[-1,1]$:

$$v = (\hat v_i - 3)/2,\quad a = (\hat a_i - 3)/2,\quad d = (\hat d_i - 3)/2$$

• Centroid Mapping: Each Ekman emotion $e$ has a fixed VAD centroid $(v^*_e, a^*_e, d^*_e)$.
• Discrete VAD-Mapped Emotion:

$$e^{\mathrm{vad}}_i = \arg\min_{e\in E} \| (v,a,d) - (v^*_e, a^*_e, d^*_e) \|_2^2$$

• Consistency and Refinement: If the initial LLM-generated $e_i$ matches $e^{\mathrm{vad}}_i$, it is retained. Otherwise, the LLM is prompted for revision given $s_i$, $c_i$, $y_i$, the original $e_i$, and $e^{\mathrm{vad}}_i$, directing alignment with both context and VAD space. This process ensures emotion labels are semantically and affectively consistent (Chai et al., 24 Nov 2025).

5. Training Setup and Benchmark Evaluation

The framework is evaluated on SemEval-2015/2016 Restaurant (Rest15/Rest16) and Laptop (Lap15/Lap16) datasets. Each is split into 90% train, 10% validation (held out), with official test splits retained. Experimental results are averaged over three random seeds.

Hyperparameter details:

• Backbone: Flan-T5 (3B)
• Emotion generator: GPT-4o-mini
• VAD predictor: DeBERTa fine-tuned on EmoBank
• Optimization: AdamW, learning rate $3 \times 10^{-5}$, batch size 4, epochs 10
• Task balance parameter: $\alpha$ selected via grid search, optimal at 0.6
• Early stopping: based on validation ACSA F1

Each batch computes $\mathcal{L}_{\mathrm{sen}}$ and $\mathcal{L}_{\mathrm{emo}}$ from the same inputs with separate serialization. Parameters $\theta$ are updated to minimize the joint loss $\mathcal{L}(\theta)$ (Chai et al., 24 Nov 2025).

6. Quantitative Analysis and Ablation Studies

Empirical evaluation demonstrates that the emotion-enhanced multi-task ACSA framework yields superior performance over baseline models, including backbone Flan-T5 and PBJM. F1 score improvements are observed on all benchmarks:

Dataset	Full Model	Flan-T5	PBJM	$\Delta$ (Flan-T5)	$\Delta$ (PBJM)
Rest15	81.65	80.21	67.58	+1.44	+14.07
Rest16	85.01	84.58	75.03	+0.43	+9.98
Lap15	78.17	75.89	62.41	+2.28	+15.76
Lap16	67.95	66.36	55.62	+1.59	+12.33

Ablation results indicate:

• Removal of emotion task supervision reduces F1 (e.g., Lap15: –2.28).
• Removal of VAD-based emotion refinement induces further declines (e.g., Rest15: –2.86). This demonstrates that both explicit emotion label learning and VAD-guided consistency contribute to the observed improvements in sentiment polarity prediction (Chai et al., 24 Nov 2025).

7. Integration of Discrete and Dimensional Affect

The framework synthesizes discrete (Ekman’s emotion taxonomy) and continuous (VAD) affective modeling. The dual approach fortifies the model's capacity to represent not only high-level sentiment polarity, but also the nuanced emotional landscape underpinning aspect-category judgments. The explicit VAD-based consistency check, followed by LLM-mediated label revision, reduces subjective misalignment that may arise from generative emotion prediction. A plausible implication is that such hybrid modeling may facilitate more interpretable and robust sentiment analysis for applications requiring psychologically grounded affective interpretations (Chai et al., 24 Nov 2025).