🦎 Research Collaboration & Funding Opportunity | Explainable NLP, Domain Adaptation & Cross-Domain Sentiment…

🦎 Research Collaboration & Funding Opportunity | Explainable NLP, Domain Adaptation & Cross-Domain Sentiment Analysis

Why does a sentiment model with 93.8% accuracy on movie reviews collapse to 86.2% — or 76% — the moment you show it a product review?

Most of the field blames missing vocabulary. Our work says the bigger damage comes from words that are present in both domains and mean the opposite thing in each.

We call them chameleon words. 🦎

"The plot was predictable" → 👎 criticism

"The battery drains predictably" → 👍 praise

Same token. Reversed polarity. Standard pivot-based alignment selects these words as reliable bridges — and then carries the wrong signal straight across the domain gap.

🚀 What we've built so far (IMDB → Amazon, zero target supervision):

📉 Quantified the transfer gap across 4 classical models, a soft-voting ensemble, DistilBERT & RoBERTa — statistically confirmed (Z = 17.43, p < 0.0001)

🔍 A simple, interpretable coefficient-comparison algorithm that surfaced 18 chameleon words (predictable, formula, solid, loud, clean, heavy…)

🧠 SHAP attributions + RoBERTa attention weights explaining why contextual models partly recover and bag-of-words models cannot

✍️ LLM-based sentiment-preserving style transfer (Flan-T5, Qwen2.5-1.5B-Instruct) to generate target-style data without a single target label

🏆 Best cross-domain accuracy: 86.20% RoBERTa vs 76.00% TF-IDF logistic regression

🔹 Where we're heading — and where you come in:

🌍 Scaling chameleon-word detection across many domain pairs to build a generalised lexicon

🛠️ Chameleon-aware adaptation — masking or re-weighting polarity-flip features at inference

📊 Quantifying the downstream gain from training on style-transferred corpora

🇧🇩 Extending the analysis to Bangla and other low-resource review corpora

🤝 Joint publications in IEEE / ACL-track venues

📢 We're seeking:

Researchers & graduate students in NLP, domain adaptation and explainable AI

Co-authors for multilingual and multi-domain extensions

Funding partners, compute sponsors and lab collaborations (our transformers were fine-tuned on a free-tier T4 — imagine what's possible with real resources 💡)

If interpretable domain adaptation, LLM-driven data augmentation, or low-resource NLP excites you — let's connect and build something that actually explains the failure instead of just patching it.

#NLP #SentimentAnalysis #DomainAdaptation #ExplainableAI #SHAP #RoBERTa #LLM #ResearchCollaboration #Bangla #MachineLearning