🌌 EmotionVerse‑2: Galactic Emotional Intelligence

EmotionVerse‑2 propels emotional AI into the cosmos—melding advanced BERT‑based multi‑task learning with psychologically‑anchored embeddings, neural plasticity gates, and memory consolidation. It doesn’t merely tag sentiment; it interprets the full spectrum of human and beyond‑human affect with scientific precision and artistic nuance.

✨ Revolutionary Emotional Paradigm

This is more than a model—it’s a new paradigm. EmotionVerse‑2 unifies six complementary tasks under one roof, sharing deep contextual representations while preserving task‑specific finesse. Through joint optimization, the system gains emergent capabilities, identifying subtle emotional shifts and cross‑task patterns impossible for single‑headed networks.

🚀Unified Multi‑Task Engine: One BERT encoder feeding specialized heads for Primary, Secondary, Meta, Sentiment, Interaction, and Context classification—trained end‑to‑end.
🧬Plutchik & Russell Fusion: Embeds every label within a combined categorical and valence‑arousal manifold, preserving family hierarchies and psychological distances.
🔗Dynamic Task Weighting: Adaptive loss balancing and focal‑loss components to ensure rare emotions get the focus they deserve without sacrificing overall stability.
📊Holistic Metrics: Beyond accuracy—balanced F1, precision, recall per task, plus cross‑task coherence scores to measure psychological consistency.

🔬 Architecture & Emotional Data Synthesis

At its core, EmotionVerse‑2 uses a transformer encoder fine‑tuned on an expanded EmotionVerse dataset. Task‑specific linear heads leverage shared embeddings, with synchronized gradient updates fostering transfer learning across emotional dimensions.

Core Modules:

Neural Plasticity Gate: Modulates hidden states based on emotion intensity signals, enabling context‑sensitive feature adaptation.
Amygdala‑Inspired Intensifier: Non‑linear intensity layer for capturing high‑arousal nuances.
GRU Memory Consolidator: Injects historical emotional context, anchoring predictions in prior sentiment sequences.
Circumplex Injector: Valence‑arousal vectors woven into the embedding fabric, preserving dimensional integrity.

🚀 Psychologically‑Informed Embeddings

Emotion labels become vectors, not IDs. Each of the 768 dimensions encodes:

Valence‑Arousal Anchors: First two dims reflect core positivity/negativity and activation.
Intensity Signals: Third dim quantifies emotional force, calibrated from dataset statistics.
Family Centroids: Emotions gravitate toward their psychological family centers, preserving conceptual clusters.
Relationship Pull: Secondary affinities nudge embeddings toward related emotions, creating a smooth emotional manifold.

📊 Evaluation & Performance Horizons

Robust evaluation spans single‑label accuracy and multi‑label F1 for each head, plus a novel Emotion Coherence Index measuring cross‑task consistency via valence‑arousal similarity.

Primary Head: Accuracy & Top‑2 recall.
Secondary & Meta Heads: Macro/Weighted F1 scores.
Interaction & Context: Precision & Recall for fine‑grained communicative styles.
Sentiment: Mixed‐class accuracy with polarity overlap analysis.

🛠️ Usage & Integration

Seamless Hugging Face support:

from transformers import AutoTokenizer, AutoModelForSequenceClassification
tokenizer = AutoTokenizer.from_pretrained("ayjays132/EmotionVerse-2")
model = AutoModelForSequenceClassification.from_pretrained("ayjays132/EmotionVerse-2")

text = "I’m thrilled yet anxious about tomorrow’s launch."
inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True)
outputs = model(**inputs)

	
		
	
	
		Parse logits into separate heads and map to labels as per config.id2label mappings

📂 The EmotionVerse Dataset: Cosmic Fuel

The EmotionVerse dataset underpins this model—a sprawling repository of 3K+ entries annotated across 6 dimensions, enriched with meta‑emotional and contextual narratives. It’s the cornerstone enabling AI to grasp the full tapestry of affect.

Explore it: ayjays132/EmotionVerse

📜 Licensing

EmotionVerse‑2 and its dataset are released under the Apache 2.0 License. Use, modify, and innovate freely—powered by open science.

🙏 Acknowledgements

Heartfelt thanks to Hugging Face (transformers, datasets) and the global research community whose innovations fuel this project’s reach for the stars.