Semantic Mappings (based on GPT-2 embedding directions)
slow ← long-term thinking | short-term → fast
minor ← negative | positive → major
Temporal: immediate, instantly, now ↔ relationships, factors, future
Sentiment: worse, hurting, ugly ↔ wonderful, excellent, beautiful
← Short-term
Long-term →
Louder notes for attended tokens
Generated text will appear here...
Temporal → Tempo
fast (short-term) → slow (long-term)
Sentiment → Scale
minor (negative) → major (positive)
How it works
Four trained probes analyze each token's semantic properties:
Temporal → Pitch (short-term = high, long-term = low)
Sentiment → Timbre brightness (negative = dark, positive = bright)
Arousal → Note speed (calm = slow, excited = fast)
Certainty → Reverb (uncertain = wet, certain = dry)
Phrase smoothing averages nearby tokens for more musical contours.