🧑‍🔬 Wasi Uddin Ahmad

Wasi Uddin Ahmad works on pretrained language models for settings that standard English text models handle poorly: source code written alongside developer prose, and languages with few NLP resources. Their papers here ask how pretraining objectives, careful data selection and explicit linguistic structure can make transformer encoders and encoder-decoders more useful in these settings. The work combines model design with new benchmarks for evaluation.

The most cited of these papers, Unified Pre-training for Program Understanding and Generation (2021), introduced PLBART. It is a BART-style encoder-decoder trained by denoising on unlabeled Java and Python functions together with English StackOverflow posts. The aim was to support code summarization, code generation and translation between programming languages without large parallel corpora. Also from 2021, Syntax-augmented Multilingual BERT adds a small graph attention network over Universal Dependencies trees to mBERT. It tests whether shared syntactic structure can narrow word-order differences between languages in zero-shot cross-lingual transfer.

The newest work, BanglaBERT (2022), moves toward a dedicated encoder for Bangla, a widely spoken language that is underserved by large multilingual models. The pretraining text comes from selected websites ranked by popularity and quality rather than from noisy public web dumps. The tokenizer's alphabet extends beyond the Bangla script so that romanized and code-switched input is still handled. The paper also builds an inexpensive inference dataset, using machine-translated training examples filtered by LaBSE similarity alongside human-translated evaluation sets. These choices connect the work to low-resource languages and evaluation and benchmarks.

Ahmad is a Senior Research Scientist at Nvidia in Santa Clara. Earlier, Ahmad was an applied scientist at AWS AI Labs from 2021 to 2024 and a graduate student researcher in computer science at the University of California, Los Angeles. Before that came a master's degree at the University of Virginia. Several papers also carry a Bangladesh University of Engineering and Technology affiliation, which fits the Bangla work. Co-authors on the papers here include Kai-Wei Chang and Md Saiful Islam. The work spans language models and pretraining and code and programming languages.

Wasi Uddin Ahmad has 3 papers in our library, in 2021–2022, most of them on Language models & pretraining, Multilingual & cross-lingual NLP, Low-resource languages; often with Kai-Wei Chang, Haoran Li, Md Saiful Islam.