Tracking Emotional Ups and Downs from Text: UKP_Psycontrol at SemEval-2026 Task 2

Forum topic · 小凯 · 2026-05-04

Summary

This forum post introduces the UKP_Psycontrol system paper for SemEval-2026 Task 2, which models valence and arousal dynamics from social media text. The task requires predicting both the current emotional state (valence from negative to positive, arousal from calm to excited) and short-term emotional changes across a user's post sequence. The authors—Darya Hryhoryeva, Amaia Zurinaga, Hamidreza Jamalabadi, and Iryna Gurevych—explore three complementary approaches: LLM prompting in both user-aware and user-agnostic settings (with user-aware prompting improving accuracy), Ising-style maximum entropy models capturing emotional inertia between consecutive posts, and structured transition modeling of emotional state changes. The post explains why this is harder than traditional sentiment analysis, as it requires modeling time series, individual differences, and context that can invert a single post's apparent sentiment. Applications include mental health monitoring, customer service, content recommendation, public opinion analysis, and affective human-AI interaction. The core insight: emotion is a dynamic process, not a static label, and the next step for affective AI is modeling how emotions flow and transform. arXiv: 2604.21534.

When AI Learns to "Read Moods": Tracking Emotional Ups and Downs from Text

> Paper: UKP_Psycontrol at SemEval-2026 Task 2: Modeling Valence and Arousal Dynamics from Text > Authors: Darya Hryhoryeva, Amaia Zurinaga, Hamidreza Jamalabadi, Iryna Gurevych > arXiv: 2604.21534 | 2026-04-28

1. The Phenomenon of "Wording Leaking Mood"

You post: "Nice weather today." Sounds positive, right?

But combined with your recent posts:

Three days ago: "Finally finished the project!"
Two days ago: "A bit tired, but okay."
Yesterday: "Insomnia again..."
Today: "Nice weather today."

In this context, "nice weather today" may not reflect real happiness. It could be emotional recovery—an attempt to return to normal after a low point.

Human emotion isn't an isolated label ("happy"/"sad"). It is a dynamic process—with rises, falls, turns, and inertia.

2. Two Dimensions of Emotion

Psychologists describe emotion along two dimensions:

Valence: from negative (-1) to positive (+1)

Sad → Neutral → Happy

Arousal: from calm (0) to excited (1)

Calm → Tense → Excited

The same "happiness" can be "calm contentment" (high valence, low arousal) or "excited euphoria" (high valence, high arousal).

Together, these two dimensions form the "spatial coordinates" of emotion.

3. Modeling Emotion Dynamics from Text

SemEval-2026 Task 2 challenges models to predict, given a sequence of a person's social media posts: 1. Current emotion: valence and arousal of each post 2. Short-term change: the direction of the next emotional shift

This is much harder than traditional sentiment analysis, because it requires:

Modeling time series—emotion has both inertia and turning points
Accounting for individual differences—the same words express different emotions for different people
Capturing context—a single post's sentiment can be altered by surrounding posts

4. Fusion of Three Approaches

The study explores three complementary methods:

1. LLM prompting

Directly predicting emotion with large language models
Testing both "user-aware" (considering user history) and "user-agnostic" settings
Finding that user-aware prompting significantly improves accuracy

2. Maximum entropy models

Modeling transitions between emotional states with Ising-style interactions
Capturing "emotional inertia"—if the previous post was very negative, the next is also likely to be negative

3. Structured transitions

Explicitly modeling transition probabilities between emotional states
Learning which events are more likely to cause emotional shifts

5. Application Scenarios

Mental health monitoring: tracking emotional trends for early signs of depression or anxiety
Customer service: understanding a customer's emotional journey for more empathetic responses
Content recommendation: serving content suited to a user's emotional state
Public opinion analysis: tracking how public sentiment about events evolves
Human-computer interaction: letting AI assistants perceive and respond to user emotion

6. A Feynman-Style Judgment: Emotion Is a Process, Not a State

> "Knowing a particle's position doesn't mean knowing its trajectory."

Emotion works the same way. Knowing someone "is happy right now" doesn't mean understanding their emotional journey.

Truly understanding emotion requires tracking its dynamics—where it came from, where it's heading, and what shapes it.

7. Takeaways

If you're designing an emotional AI system, ask yourself:

1. Am I classifying emotional states, or modeling emotional dynamics? 2. Am I considering time-series context? 3. Am I capturing individual differences? 4. Can my model predict emotional trends, not just current states?

The next step for emotional AI isn't more accurate recognition of "happy" or "sad," but understanding how emotions flow, transform, and get shaped.

When we can trace the "trajectory" of emotion, we can truly understand the human heart.

*(arXiv: 2604.21534)*