Decision Intelligence for the Enterprise

Graduate Algorithms Course

- Investigated the impact of causal direction data for 2 Natural Language Processing tasks of Neural Machine Translation and Syntax Discovery.- The research was well-received as an oral presentation at EMNLP 2021 conference.

Research Areas: Deep Learning, Natural Language Processing, Transformers, Noisy Data- Achievements: - Best Bachelor's Thesis Finalist - 1 Paper at NAACL 2021.- Summary: - Experiment and developed on a 3 level explainable Rumour Detection and Tweet Verification - Explored improving stance detection using semantic dependency graphs.Diagnosed the presence of spurious cues in the datasets.

Research Areas: Computer Vision, Multimodal Learning, Video Captioning- Conceptualized novel task of Spatio-temporal video dense captioning and curated a 1.2M+ Captions dataset.- Outperformed referring expression generation SoTA, with a show-attend-tell inspired model delivering 1.29 CIDEr.

• Achieved 23% and 42% F1 on token-level and sentence-level bias classification respectively, over state of the art.• Developed a novel solution for bias removal by deletion or rewrite using language generation model preserving information.• Designed an end to end pipeline with a token-level interpretable report for the 18 Bias types tagged and removed by model.

• Contributor & Maintainer for NLP Research packages WordTokenizers and TextAnalysis; total 300+ GitHub Stars.• Formulated novel regex-free approach for custom multilingual tokenization upto 4 times faster than SpaCy, NLTK.• Proposed end-to-end CRF-LSTM model for general-purpose NER, POS tagger with 93.1, 91.2 F1 on CoNLL’03.• Presented at JuliaCon 2020 Conference as a full talk; WordTokenizers published in JOSS journal.