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: Your Way of Speaking Leaks Your Mood

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

But consider your recent posting history:

• 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 genuine happiness. It might represent emotional recovery — an attempt to return to normal after a low period.

Human emotion is not an isolated label ("happy"/"sad"). It is a dynamic process — with ups and downs, turning points, and inertia.

2. The 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 → Agitated

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

Together, these two dimensions form an emotion's spatial coordinates.

3. Modeling Emotion Dynamics from Text

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

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

• Temporal sequences — emotion has inertia, but also turning points
• Individual differences — the same words express different emotions for different people
• Context — a single post's sentiment can be altered by its context

4. Fusing Three Approaches

The research explores three complementary methods:

1. LLM Prompting

• Large language models predict emotion directly
• Both "user-aware" (considering user history) and "user-agnostic" settings were tested
• User-aware prompting significantly improved accuracy

2. Maximum-Entropy Models

• Ising-style interactions model transitions between emotional states
• Captures "emotional inertia" — if the previous post was very negative, the next is more likely negative too

3. Structured Transitions

• Explicitly models transition probabilities between emotional states
• Learns which events are more likely to cause emotional shifts

5. Application Scenarios

What is emotion dynamics modeling useful for?

• Mental health monitoring: tracking emotional trends to detect early signs of depression or anxiety
• Customer service: understanding a customer's emotional journey for more empathetic responses
• Content recommendation: suggesting content based on a user's emotional state
• Public opinion analysis: tracking how public sentiment toward events evolves
• Human-AI interaction: enabling AI assistants to perceive and respond to user emotional states

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

Though a physicist, Feynman understood people deeply:

> "Knowing where a particle is tells you nothing about its trajectory."

Emotion works the same way. Knowing that someone is "happy right now" does not mean understanding their emotional journey.

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

7. Takeaways

If you are designing an emotional AI system, ask yourself:

1. "Am I classifying emotional states, or modeling emotional dynamics?" 2. "Have I considered temporal context?" 3. "Have I captured individual differences?" 4. "Can my model predict emotional trends, not just the current state?"

The next step for affective AI is not more accurate recognition of "happy" or "sad", but understanding how emotions flow, transform, and are shaped.

Only when we can trace the "trajectory" of emotion can we truly understand the human mind.