Arkajyoti Chakraborty

Worked on activity recognition problem on video-based data, building a model to classify different activities using CNN feature extractor and fine-tuned transformer encoder model (84% accuracy). Developed an abnormal breathing detection pipeline using YOLO for masking and the activity pipeline (79% accuracy).

Worked on domain adaptation techniques over fake news data and hypothesized in relation of fake news to emotion features. Studied about gradient reversal method to apply domain adaptation over the cross-data to enhance the accuracy. Worked on two short papers getting accepted at the AAAI’23 student abstract track and ICON’22 short paper track.

Implemented three zero-shot (Bert, Distilbert, Roberta) spark-nlp annotators in the NLU pipeline. Designed test scripts for sequence classifier models (Longformer, Xlnet, Albert, and Debarta) and implemented demo notebooks for zero-shot classifiers. Successfully merged PR’s regarding annotators Bert, Distilbert, Roberta, and Sequence Classifiers for the upcoming updated release of the NLU library.

Research on clarification question generation via retrieval-based prompting large language models (LLMs) on Legal clauses and contracts. Proposed a retrieval-augmented prompting framework designed explicitly for clarification question generation for contracts. Designed and performed experiments on open-source large language models: Vicuna, Alpaca-Lora, and Dolly-V2 over zero-shot and few-shot setups. Paper titled ”Generating Clarification Questions for Disambiguating Contracts” under review at EMNLP’23.

Built a pipeline of custom-trained OCR models from scratch, focusing on lab reports and their accuracy over units, ranges, and numeric values. Tested different models for inference over the edge cases and trained a TrOCR model over 50k images, achieving an average CER of 0.04, outperforming AWS Tesseract.