Score what sentiment analysis can't: complexity.
Fossier scores the complexity of any emotion pair on a 0–198 scale. 28 named states, from guilt to ambivalence, each with an empathic difficulty level. Built on 6 years of research, validated on 28,218 texts.
Emotional distance computed in 3D Pleasure-Arousal-Dominance space with cross-validated axis weights. Not a simple sentiment score.
Every emotional pair mapped and named, from Mepris to Culpabilite.
Complexity maps to cognitive demand for empathic response.
Pearson correlation between predicted complexity and observed co-occurrence across two independent multilabel datasets.
Your best friend is moving abroad. You're happy for them. You're devastated for yourself. Both feelings are real, and neither cancels the other.
You discover your business partner has been siphoning funds for months. You trusted them completely. Now every shared memory feels contaminated.
He survived and came home. She's overjoyed. But every car backfire makes him flinch, and she's terrified it will never be the same.
She walks across the stage with honors. Her parents didn't come. She proved everyone wrong. And she's furious she had to.
Someone cuts in line and spits on the floor. You feel exactly what you'd expect. No ambiguity, no inner conflict.
No credit card required. Full model access.
Send any text to /v1/analyze and get back detected emotions, named dyads, complexity scores, and empathic levels.
Submit any text up to 5,000 characters in any language supported by the LLM.
Claude or GPT identifies emotions present with intensity and textual evidence.
Each detection maps to the 8 base emotions with PAD coordinates.
Weighted PAD distance yields a complexity score from 0 to 198 with an empathic level.
Standard JSON over HTTP. No SDK required. Works from any language, any platform. Auth via Bearer token.
From $0/mo. Scale when you're ready.
We present C_dyade, a continuous scoring model for emotional complexity. Each of Plutchik's 8 primary emotions is projected into the PAD (Pleasure-Arousal-Dominance) affective space using empirically derived coordinates. Complexity between any emotion pair is computed as a [...] function of their geometric distance in ℝ³ and combined intensity, producing a score on a [0, 198] scale. Model parameters were estimated through k-fold cross-validation on [...] expert-annotated items, with stable convergence across all folds. The model identifies 28 named dyadic states, each assigned an empathic difficulty level.
| Corpus | n | Spearman r | p |
|---|---|---|---|
| XED English | 17,528 | −0.854 | < 0.001 |
| SemEval-2018 | 10,690 | −0.839 | < 0.001 |
No credit card required. Upgrade anytime.
For prototyping and exploration. Full model access with daily limits.
No credit card • No expiration
For production applications. Higher limits, priority support, webhook notifications.
14-day free trial • Cancel anytime
Unlimited volume, dedicated infrastructure, SLA guarantees, and custom integrations.
Sentiment analysis tells you positive or negative. Fossier tells you how complex the emotional state is. Joy + Fear (guilt) and Joy + Trust (love) both read as "mixed" in sentiment analysis. Fossier gives them distinct scores (148 vs 27), named dyads, and empathic levels.
The /v1/compute endpoint is language-agnostic: it takes emotion labels directly. The /v1/analyze endpoint processes raw text via LLM (Claude or GPT), so it works in any language your LLM supports.
< 5ms for /v1/compute (pure math, no ML inference). ~1-2s for /v1/analyze (includes LLM call for emotion detection). The compute endpoint is ideal for real-time applications like games and chatbots.
Plutchik's wheel is the most empirically validated discrete emotion taxonomy. It provides 8 base emotions with clear oppositions and graded intensities. Our PAD axis weights are cross-validated on 28,218 texts from two independent datasets (XED & SemEval).
The Enterprise plan includes self-hosted deployment options with dedicated support and custom model tuning. The compute engine is a single Python module with zero external dependencies. Contact us for details.
The /v1/compute endpoint takes two emotion labels and two intensity values. No LLM needed. It's pure math at < 5ms. Use your own emotion detection pipeline and let Fossier handle the complexity scoring.
GPT gives you a label: “mixed emotions.” Fossier gives you a score (C=148), a named dyad (guilt = joy + fear), and an empathic level (creation). That granularity lets your system decide how to respond, not just that there are emotions. Plus, /v1/compute runs in < 5ms with zero LLM cost.