Sep 2022 – Dec 2022 · 3 mos · Pittsburgh, Pennsylvania, United States

• Advanced NLP by Dr. Graham Neubig (Fall '22)
• Natural Language Processing by Dr. David Mortensen (Spring '22)

Jun 2022 – Aug 2022 · 2 mos · Cambridge, Massachusetts, United States

• – As a part of Siri Understanding (SaLT), designed, implemented and trained different models at scale to improve robustness of automatic speech recognition (ASR) in Siri.

Sep 2020 – Feb 2021 · 5 mos

• • Worked on improving dense retrieval performance in cross-domain QA systems through synthetic query generation using transformer-based seq2seq models (such as T5 and BART.) The approach can substitute expensive gold annotation for specialised domains using a synthetically created silver training set.

Jan 2020 – Jun 2020 · 5 mos · Bangalore, India

• Presented a detailed case study of Indian Twitter for language and script usage. Showed the extent of script-mixing online, and catalogued the functions of script-alternation in different contexts. We show how script-mixing can intentionally be employed to emphasize certain entities, and also, pronounce sarcasm in certain cases.
• Designed a two-stage user study to explore when and where should code-mixing incorporate script-mixing. The first stage measures the likelihood of mixed-script text occurring in different domains, while the second measures the direct preference for the different versions. Results can motivate the design of multilingual interfaces in personal digital assistants and elsewhere.

May 2019 – Jul 2019 · 2 mos · Paris 12, Île-de-France, France

• Built a pipeline for extending a small Arabizi (code-mixed French-Arabic dialect) dataset by filtering out sentences from a web crawled corpus. Using language identification scores from a bag of n-grams model (fastText) as features, we trained a linear SVM classifier to extract actual Arabizi sentences.
• Built a normalisation module for code-mixed word variants leveraging distributional (word2vec), lexical (Levenshtein distance) and phonetic (Soundex) similarities. Fine-tuned mBERT for POS tagging, and showed that it can adapt to code-mixed text by continuing the MLM pre-training given sufficient unsupervised data.
• In a different project, generated representations for French Twitter users using their tweets by modifying the existing sentence embedding approaches. Showed through plots that the embeddings captured the regional information about the user suggesting the presence of region-specific linguistic features in tweets.

Feb 2019 – Mar 2019 · 1 mo · NIT Calicut, Kerala, India

• Using the R-CNN algorithm, we built a vehicle detection module to count the number of cars in the provided aerial footage. We then used the frequencies to train a linear regression model to predict the regular traffic patterns throughout the day in the city.
• Treating the junctions as nodes in a graph, we modified the max-flow min-cut algorithm to find the least busy (i.e. quickest) route for a vehicle.
• Finally, proposed an adaptive traffic light design that can guide emergency vehicles towards an optimized green corridor whenever requested.