Telugu Sentiment Summary Engine
Budget: ₹12,500 – ₹37,500 INR
I have a growing collection of Telugu customer-review text that needs to be distilled into concise, one-line summaries and tagged as Positive, Negative, or Neutral. The result I need is a clean JSON output per record, so each review comes back with its summary and sentiment label in a machine-readable format.
Because the language is highly nuanced, I’d like you to blend both rule-based and machine-learning techniques: think lexicon cues for idiomatic Telugu alongside a fine-tuned transformer or any other classifier that lifts accuracy. Feel free to draw on pretrained Telugu-BERT, FastText, spaCy, custom dictionaries—whatever combination you believe delivers the most reliable hybrid model.
Deliverables
• Python or notebook script that ingests raw Telugu text and produces the JSON format
• Trained model files (and any custom lexicons) with version control
• README explaining setup, dependencies, and how to retrain or update the model
• Brief validation report: precision, recall, and overall accuracy on a held-out test set of the same domain
Acceptance criteria
The pipeline must run end-to-end on my sample dataset and reach a sentiment-classification F1 score of at least 0.80 while generating legible summaries that preserve key points from each review.
Because the language is highly nuanced, I’d like you to blend both rule-based and machine-learning techniques: think lexicon cues for idiomatic Telugu alongside a fine-tuned transformer or any other classifier that lifts accuracy. Feel free to draw on pretrained Telugu-BERT, FastText, spaCy, custom dictionaries—whatever combination you believe delivers the most reliable hybrid model.
Deliverables
• Python or notebook script that ingests raw Telugu text and produces the JSON format
• Trained model files (and any custom lexicons) with version control
• README explaining setup, dependencies, and how to retrain or update the model
• Brief validation report: precision, recall, and overall accuracy on a held-out test set of the same domain
Acceptance criteria
The pipeline must run end-to-end on my sample dataset and reach a sentiment-classification F1 score of at least 0.80 while generating legible summaries that preserve key points from each review.