Code-Mixed (Hin-Eng) Sentiment Analysis System with Synthetic Data Augmentation
Budget: ₹1,000 – ₹3,000 INR
I have completed the literature survey for a post-graduate project entitled “Enhancing sentiment analysis for Hindi-English code-mixed text using synthetic data augmentation.” Now I need the full system brought to life.
Scope and goals
• Build or adapt a sentiment classifier for code-mixed (hi-en) social-media style text.
• Inject novelty through synthetic data augmentation so that the final model demonstrably outperforms a plain baseline trained on the raw corpus.
• Deliver clean, well-commented source code and set everything up on my M3 MacBook running macOS 15.6.
What I already have
• A starter dataset downloaded from the web and a PPT that details candidate architectures, augmentation ideas (back-translation, contextual synonym replacement, etc.), and evaluation protocol.
• Openness to stronger datasets or alternative techniques if you can justify the gain.
What I need from you
1. Data preprocessing tailored for noisy code-switching (tokenisation, transliteration handling, spelling mistakes, duplication cleaning).
2. Custom or fine-tuned model implementation—TensorFlow, PyTorch, or Scikit-learn, whichever best meets the accuracy target.
3. Robust synthetic data generation pipeline integrated into training.
4. A lightweight local web interface (Flask, Streamlit, or similar) so I can demo predictions live.
5. Hands-on assistance to install dependencies, run notebooks/scripts, and reproduce results on my machine. A quick screen-share session at the end is fine.
Acceptance criteria
• At least one augmentation strategy that lifts F1 or accuracy beyond the non-augmented baseline on a held-out test split.
• Reproducible training script, saved model weights, README with exact commands, and a requirements.txt/poetry.lock.
• Local web demo reachable at http://localhost:8501 (or similar) returning sentiment labels in real time.
Deadline - 15th October 2025.
When you apply, please highlight past work on multilingual NLP, code-mixing, or data-centric model improvements so I can gauge fit quickly. I’ll share the PPT and current dataset once we connect.
Scope and goals
• Build or adapt a sentiment classifier for code-mixed (hi-en) social-media style text.
• Inject novelty through synthetic data augmentation so that the final model demonstrably outperforms a plain baseline trained on the raw corpus.
• Deliver clean, well-commented source code and set everything up on my M3 MacBook running macOS 15.6.
What I already have
• A starter dataset downloaded from the web and a PPT that details candidate architectures, augmentation ideas (back-translation, contextual synonym replacement, etc.), and evaluation protocol.
• Openness to stronger datasets or alternative techniques if you can justify the gain.
What I need from you
1. Data preprocessing tailored for noisy code-switching (tokenisation, transliteration handling, spelling mistakes, duplication cleaning).
2. Custom or fine-tuned model implementation—TensorFlow, PyTorch, or Scikit-learn, whichever best meets the accuracy target.
3. Robust synthetic data generation pipeline integrated into training.
4. A lightweight local web interface (Flask, Streamlit, or similar) so I can demo predictions live.
5. Hands-on assistance to install dependencies, run notebooks/scripts, and reproduce results on my machine. A quick screen-share session at the end is fine.
Acceptance criteria
• At least one augmentation strategy that lifts F1 or accuracy beyond the non-augmented baseline on a held-out test split.
• Reproducible training script, saved model weights, README with exact commands, and a requirements.txt/poetry.lock.
• Local web demo reachable at http://localhost:8501 (or similar) returning sentiment labels in real time.
Deadline - 15th October 2025.
When you apply, please highlight past work on multilingual NLP, code-mixing, or data-centric model improvements so I can gauge fit quickly. I’ll share the PPT and current dataset once we connect.