EmoPerso: Emotion-Aware Personality Detection

• EmoPerso is a self-supervised framework for personality detection that integrates emotion-aware modelling to improve MBTI prediction from text.
• It employs generative augmentation, pseudo-labeled emotion modelling, and multi-task learning with cross-attention to fuse personality and emotion representations.
• Empirical results on Kaggle and Pandora datasets demonstrate significant Macro-F1 gains, underscoring the importance of emotion integration.

EmoPerso is a self-supervised framework for personality detection from text that improves personality detection through emotion-aware modelling. It first leverages generative mechanisms for synthetic data augmentation and rich representation learning, then extracts pseudo-labeled emotion features and jointly optimizes them with personality prediction via multi-task learning; a cross-attention module captures fine-grained interactions between personality traits and inferred emotional representations, and a self-taught strategy refines relational reasoning iteratively (Shen et al., 2 Sep 2025). Within the broader literature, it belongs to a family of systems that model the interaction between emotion and persona or personality across text, dialogue, speech, and vision, rather than treating them as independent signals (Zhang et al., 2018).

1. Problem setting and conceptual scope

EmoPerso addresses personality detection from text, where the objective is to infer stable personality traits from users’ social media posts. The framework focuses on MBTI, treating the four binary dimensions I/E, S/N, T/F, and P/J as four independent binary classification tasks over forum posts or Reddit comments (Shen et al., 2 Sep 2025). The motivating claim is twofold: existing methods rely heavily on large-scale annotated personality datasets, and many methods either ignore emotion or model it only shallowly, even though emotions and personality are cognitively linked.

The framework’s central premise is that emotion should be modeled as an auxiliary but structured signal for personality inference. In EmoPerso, emotions are not supplied through an external annotated emotion dataset; instead, they are inferred in a self-supervised fashion from the same personality-labeled posts via pseudo-labels, and then optimized jointly with personality prediction. This design contrasts with work that studies personality and emotion in isolation, and it aligns with older joint formulations such as PersEmoN, which investigated whether apparent personality, emotion, and their relationship can be jointly learned from face images in the wild (Zhang et al., 2018).

This suggests that EmoPerso is best understood not merely as a classifier, but as a representation-learning system in which personality prediction is improved by explicitly organizing affective information. A plausible implication is that its gains arise less from any single architectural novelty than from enforcing multiple complementary constraints on the latent space: stylistic augmentation, pseudo-emotion supervision, cross-modal interaction inside text representations, and reasoning-chain selection.

2. Core architecture

The framework is organized as a staged pipeline around a frozen LLM backbone, reported as DeepSeek-V3 in the main experiments. For each input batch of posts $X$ with personality labels $y_{\text{pers}}$, the system first applies Augment(X) to create style-conditioned paraphrases and contextual completions, yielding augmented text $\widetilde{X}$. The backbone then encodes $\widetilde{X}$ into token embeddings $H = (h_1,\dots,h_T)$, and attention pooling produces a global representation $z$ (Shen et al., 2 Sep 2025).

The shared representation is decomposed into personality-specific and emotion-specific vectors: $$z_{\text{pers}} = f_{\text{pers}}(z), \quad z_{\text{emo}} = f_{\text{emo}}(z).$$ Personality prediction treats the four MBTI dimensions as four binary tasks: $$p_c = \sigma\big((W_{\text{pers}} z_{\text{pers}} + b_{\text{pers}})_c\big), \quad c=1,\dots,C_p,$$ with binary cross-entropy loss

$$\mathcal{L}_{\text{pers}} = - \sum_{c=1}^{C_p} \left( y_c \log p_c + (1-y_c)\log(1-p_c) \right).$$

The attention pooling step itself is defined as

$$z = \sum_{i=1}^{T} \alpha_i h_i, \quad \alpha_i = \frac{\exp\big(v^\top \tanh(Wh_i+b)\big)}{\sum_{j=1}^{T}\exp\big(v^\top \tanh(Wh_j+b)\big)}.$$

After decomposition, the Interact module refines personality features through personality–emotion interaction. A multi-head cross-attention block uses $y_{\text{pers}}$0 as query and token embeddings $y_{\text{pers}}$1 as keys and values; emotion embeddings $y_{\text{pers}}$2 are then used to compute token-level weights and aggregate the attended token features into a refined personality vector $y_{\text{pers}}$3. The final prediction stage concatenates $y_{\text{pers}}$4 with a reasoning representation $y_{\text{pers}}$5 produced by the self-taught module and maps them through an MLP: $y_{\text{pers}}$6

The backbone LLM is frozen; only the pooling layer, MLP heads, cross-attention, and reasoning fusion layers are trained. This division of labor is architecturally important: generic linguistic competence is inherited from the frozen LLM, while the trainable layers specialize the representation toward joint personality–emotion inference.

3. Generative augmentation and pseudo-labeled emotion modelling

A distinctive part of EmoPerso is its generative front end. Given an input post sequence $y_{\text{pers}}$7, the system uses an LLM with style control signal $y_{\text{pers}}$8 such as formality, expressiveness, or conciseness to generate paraphrases: $y_{\text{pers}}$9 It also performs contextual feature completion by masking phrase-level spans and prompting the LLM to reconstruct them. To maintain semantic consistency between original and augmented text, EmoPerso imposes a KL regularization term between token-level output distributions: $\widetilde{X}$0 A style-classification head predicts the prompted style labels, and the total generative loss is

$\widetilde{X}$1

Emotion modelling is likewise self-supervised. Rather than relying on external emotion annotations, EmoPerso derives pseudo-labels from affective adjectives, intensifiers, punctuation and exclamation marks, emotive $\widetilde{X}$2-grams, and other psycholinguistic markers. These cues form a binary vector $\widetilde{X}$3 over emotion categories $\widetilde{X}$4, described as following basic emotion theory, including joy, anger, sadness, fear, disgust, surprise, and contempt. Emotion prediction uses sigmoid outputs: $\widetilde{X}$5 with multi-label binary cross-entropy

$\widetilde{X}$6

The joint objective combines personality and emotion losses: $\widetilde{X}$7 In the reported experiments, the ratio is about $\widetilde{X}$8 for personality versus emotion, reflecting emotion’s auxiliary role. This design gives emotion two roles simultaneously: a supervision source for shared representation learning and a conditioning signal for later cross-attention and reasoning stages.

4. Personality–emotion interaction and self-taught reasoning

The cross-attention module is the mechanism by which EmoPerso converts emotion from a side task into a conditioning variable for personality inference. For each attention head $\widetilde{X}$9,

$\widetilde{X}$0

and scaled dot-product attention is computed as

$\widetilde{X}$1

Outputs from all heads are concatenated and projected to obtain attended token representations. Emotion-conditioned modulation then computes token-level importance weights

$\widetilde{X}$2

and the final personality vector is

$\widetilde{X}$3

A cosine-based alignment loss encourages compatibility between the refined personality representation and emotion embedding: $\widetilde{X}$4

The second major mechanism is the self-taught reasoning module, inspired by STaR. For each input $\widetilde{X}$5, the LLM decoder generates multiple candidate reasoning chains $\widetilde{X}$6. Each chain is scored by information gain,

$\widetilde{X}$7

and by mutual information with emotion features,

$\widetilde{X}$8

The preference score is

$\widetilde{X}$9

and the best chain is selected by the corresponding argmax criterion. The selected chain is embedded as $H = (h_1,\dots,h_T)$0 and fused with $H = (h_1,\dots,h_T)$1. To avoid uniform weighting over chains, EmoPerso adds an entropy loss

$H = (h_1,\dots,h_T)$2

The full training objective is

$H = (h_1,\dots,h_T)$3

This architecture makes explicit that EmoPerso does not treat rationales as post hoc explanations. Instead, reasoning chains are operational latent variables used to improve prediction by selecting chains that both reduce uncertainty in personality prediction and correlate with inferred emotion features.

5. Datasets, baselines, and empirical results

EmoPerso is evaluated on two MBTI-labeled social media datasets: Kaggle MBTI, drawn from PersonalityCafe forum posts with 8,675 users, and Pandora, a Reddit-based MBTI dataset with 9,084 users. Both are split $H = (h_1,\dots,h_T)$4 into train, validation, and test. The principal metric is Macro-F1 across the four MBTI dimensions (Shen et al., 2 Sep 2025).

Dataset	Best baseline	EmoPerso
Kaggle MBTI Avg Macro-F1	TAE 72.07%	81.07%
Pandora Avg Macro-F1	TAE 63.05%	68.60%

On Kaggle, EmoPerso reports I/E 80.05, S/N 79.27, T/F 87.03, P/J 77.91, with an average of 81.07. On Pandora, it reports I/E 66.84, S/N 68.15, T/F 71.90, P/J 67.51, with an average of 68.60. The paper states that it is the first reported model to exceed $H = (h_1,\dots,h_T)$5 average Macro-F1 on Kaggle, and that it improves especially under class imbalance on Pandora.

The ablation study attributes most of the gain to emotion-aware modelling. Removing emotions reduces the average Macro-F1 to 73.03 on Kaggle and 63.21 on Pandora. Removing the generative mechanism yields 76.52 and 64.77. Removing multi-task learning yields 74.84 and 63.96. Removing the shared encoder yields 78.58 and 66.50. Removing the cross-attention mechanism yields 76.10 and 64.51, while replacing it with gated fusion yields 78.87 and 66.12. Removing emotion modulation yields 77.84 and 65.70. Removing reasoning chains yields 75.42 and 64.15, and replacing STaR with CoT templates yields 77.68 and 65.91. Using reasoning chains without IG and MI still performs strongly at 79.21 and 67.13, but remains below the full model.

The reported backbone comparison shows that replacing DeepSeek-V3 with GPT-4o yields 80.75 on Kaggle and 69.08 on Pandora, described as comparable performance. Additional analyses report that heatmaps show varying importance of emotions across MBTI dimensions, t-SNE plots show overlapping and structured clusters for shared personality and emotion representations, and training with STaR reaches the best epoch faster while maintaining comparable overall training time. These results support the paper’s claim that emotion modelling is the most critical single component and that explicit interaction modeling is more effective than coarse fusion.

6. Related paradigms, limitations, and broader significance

EmoPerso sits within a wider research program that couples affect with persona, personality, or appraisal. PersEmoN jointly analyzes apparent personality, emotion, and their relationship from face images in the wild using a shared feature extraction module, two task-specific branches, and an adversarial-like loss to promote representation coherence across heterogeneous datasets (Zhang et al., 2018). Meta-PerSER adapts speech emotion recognition to each listener’s subjective labeling style using MAML, Combined-Set Meta-Training, Derivative Annealing, and per-layer per-step learning rates, emphasizing listener-personalized rather than globally aggregated emotion labels (Shen et al., 22 May 2025). In dialogue, PAL studies persona-augmented emotional support conversation generation and proposes a framework for dynamically inferring and modeling seekers’ persona (Cheng et al., 2022), while PESS introduces completeness loss and consistency loss for persona extraction through semantic similarity and shows that higher-quality inferred persona improves emotional-support response generation (Han et al., 2024). Persona-E$H = (h_1,\dots,h_T)$6 supplies a human-grounded dataset in which emotional responses to the same event vary across measured MBTI and Big Five profiles, and reports that Big Five information alleviates “personality illusion” in LLM-based appraisal (Yang et al., 10 Apr 2026).

Within that landscape, EmoPerso is specifically a text-based personality detector rather than a dialogue system or a direct emotion recognizer. Its stated limitations are that the datasets are imbalanced, the data are mostly English social media text, and self-reported MBTI together with social media verbal behavior may not precisely reflect underlying personality (Shen et al., 2 Sep 2025). These are substantive constraints: MBTI labels are sparse and noisy, and language or cultural bias can affect both pseudo-emotion heuristics and personality classification.

This suggests that EmoPerso’s significance lies less in solving personality inference definitively than in establishing a concrete template for joint emotion–personality modelling under limited supervision. A plausible implication is that the framework’s modular design—LLM-based augmentation, pseudo-labeled affective supervision, token-level personality–emotion interaction, and reasoning-chain selection—can be transplanted into adjacent tasks such as emotional-support dialogue, reader-conditioned appraisal, or personalized affective generation. In that sense, EmoPerso is both a specific 2025 framework for MBTI prediction from text and a representative instance of a broader shift toward modelling personality and emotion as mutually informative latent factors rather than isolated variables (Shen et al., 2 Sep 2025).