Mumbai, Maharashtra, India
• Developed Few Shot Named Entity Recognition model to extract company relevant entities on hinglish call transcripts • Engineered prompt to curate labeled dataset consisting of 70k+ sentences with 23+ classes using GPT-3.5+ turbo for NER • Trained FSNER model that uses similarity function; Achieved 60% accuracy for 5-way 10-shot and 40% on 10-way 10-shot • Improved F1 score to 60% for5-way10-shot, 57% for10-way10-shot using SDnet by only fine-tuning on support examples