MT Summit 2023

Shuyue Stella Li, Philipp Koehn

We develop a multilingual MT system from non-native source sentences to investigate the differences between human language acquisition and machine language acquisition. We create synthetic L2 data to enhance the robustness of L2 MT systems.

Arxiv

Shuyue Stella Li, Cihan Xiao, Tianjian Li, Bismarck Odoom

We propose a residual deep CNN model and a multi-task learning model to identify segment-level languages in English-Mandarin code-switched utterances. Data augmentation and fine-tuning methods are also explored to improve performance.

EMNLP 2023

Haoran Xu, Weiting Tan*, Shuyue Stella Li*, Yunmo Chen*, Benjamin Van Durme, Philipp Koehn, Kenton Murray

We introduce Language-Specific Matrix Synthesis (LMS) and Fuse Distillation (FS) to condense knowledge and improve model efficiency in multilingual language models.

ICNLSP 2023

Shuyue Stella Li*, Beining Xu*, Xiangyu Zhang*, Hexin Liu, Wenhan Chao, Leibny Paola Garcia

We propose a linguistically motivated metric: Phonetic-Syntax Ratio (PSR) to quantify the composition of phonetic and syntactic contents in learned representations.

Arxiv

Shuyue Stella Li & Kneton Murray

We establish a language-agnostic synthetic code-mixing algorithm that generates code-mixing datasets in a zero-cost fashion.

ICASSP 2023

Shuyue Stella Li*, Xiangyu Zhang*, Shu Zhou, Hongchao Shu, Ruixing Liang, Hexin Liu, Leibny Paola Garcia

We propose a quantum language model that can easily transmit learned information to downstream classical models, demonstrating ad hoc model portability.

ICASSP 2023

Yu Xuan, Xiangyu Zhang, Shuyue Stella Li, Zihan Shen, Leibny Paola Garcia, Roberto Togneri

This project uses a convex combination of adaptive filters to detect the fetal heart rate from the ECG signals from the mother's abdominal and thorax. Our algorithm out-performs both single adaptive filters and convex combination of RLS filters

Arxiv

Xiangyu Zhang, Shuyue Stella Li, Zhanhong He, Roberto Togneri, Leibny Paola Garcia

We design an end-to-end lyrics recognition model using self-supervised models trained on spoken speech and out-performed previous SOTA.

GECCO 2022 Best Presentation Award

Shuyue Stella Li, Hannah Peeler, Andrew Sloss, Kenneth Reid, Yuan Yuan, Wolfgang Banzhaf

We present the novel use of genetic improvement to find problem-specific optimized LLVM Pass sequences. This demonstrates the flexibility of the Shackleton Framework in being applied to a variety of use cases.

GECCO 2022

Hannah Peeler, Shuyue Stella Li, Andrew Sloss, Kenneth Reid, Yuan Yuan, Wolfgang Banzhaf

We design and implement a novel GP algorithm for LLVM compiler flag optimization, achieving a 20% runtime improvement over strong default baseline.

JHU Design Synposium 2022

Weicheng Hu, Shuyue Stella Li, Shanelle Cao, Bohan Hou, TJ Bai, Anton Dahbura

We design an expansion-based clustering algorithm for community detection using covid-19 mobility data to optimize disease simulations model.

UC Davis COSMOS Synposium 2018

Shuyue Stella Li, Janani Sekar, Jeffrey Yang

We model the activity of DNA recombinase CRE-lox and predict the topological transformations of DNA knots treated with Cre Recombinase. This research has potential pharmaceutical applications that can be furthered to improve the efficiency of enzyme activity in transforming circular DNA chains into the unknotted form.