publications

2024

1. REFeREE: A REference-FREE Model-Based Metric for Text Simplification

   Accepted at LREC-COLING 2024 (long paper)

   Yichen Huang, and Ekaterina Kochmar

   2024

   TL;DR arXiv GitHub

   • Text simplification lacks a universal criterion of quality, and annotated references are costly. We propose a pretrained TS metric that can be fine-tuned to suit different quality standards.
   • A 3-stage training curriculum: 1) Arbitrarily scalable pretraining on source-based objectives. 2) Pretraining on source-based and higher quality reference-based objectives. 3) Fine-tuning to specific criteria.
   • Outperforms reference-based metrics in predicting overall quality and reaches competitive performance in predicting specific quality.

2023

1. Robustness Tests for Automatic Machine Translation Metrics with Adversarial Attacks

   EMNLP 2023 Findings (short paper)

   Yichen Huang, and Timothy Baldwin

   2023

   TL;DR arXiv GitHub

   • MT metrics are inevitably used on out-of-distribution translations from unseen systems. We investigate the robustness of such metrics with off-the-shelf adversarial attacks.
   • Our human evaluations validate that BERTScore, BLEURT and COMET tend to overpenalize adversarially-degraded texts, especially those associated with unseen, strong MT systems.
   • We also devise a way to probe for self-inconsistency in distance-based metrics (e.g. BERTScore) without human ratings.
2. Learning Interpretable Low-dimensional Representation via Physical Symmetry

   NeurIPS 2023

   Xuanjie Liu, Daniel Chin, Yichen Huang, and Gus Xia

   2023

   TL;DR arXiv Video Demo GitHub

   • How can we learn low-dimensional, interpretable representations from perception (e.g. pitch from music audio) with minimal domain knowledge? What gives rise to these concepts in the first place?
   • Physical symmetry as a fundamental inductive bias, implemented as a self-consistency constraint (e.g. the representation of pitch should be temporally transposition-equivariant).
   • Our models learn linear pitch factors and 3D Cartesian coordinates in an unsupervised fashion.
   • My part in this: Parts of experiments in Sec 4.1 (linear pitch factor). Design, implementation and writing for experiments in Secs A.4.2 (pitch-timbre disentanglement) and A.6.1 (SSL with arbitrary natural melodies).
3. Causal Discovery on the Capacities and Specifications of Large Language Models

   Course project for NLP 702: Advanced NLP

   Yichen Huang, Sathyamoorthy Rajendran, and Duy Khai Doan

   2023

   TL;DR PDF Slides GitHub

   • Can we establish causal relationships between LMs’ performances in different categories (e.g. better algebraic performance causing better performance in physics) and between the performances and LM specifications (e.g. more parameters causing better algebraic performance)?
   • We apply off-the-shelf causal discovery algorithms on MMLU performances of 34 LMs.
   • No strong evidence that performances on conceptually close subjects are causally related (e.g. high school/college computer science).
   • Model size does not seem to causally determine the performances in all subjects.
4. JukeControl: Enhancing Jukebox for Audio-to-Audio Music Generation with ControlNet

   Unpublished research, course project for NLP 703: Speech Processing

   Yichen Huang, and Gus Xia

   2023

   TL;DR PDF Slides GitHub

   • Modifying Jukebox’s top-level transformer prior model with ControlNet to enable audio-to-audio conditional generation.
   • Early results on vocal-source separation, vocal-conditioned accompaniment generation and note-conditioned synthesis.
   • Future work: recasting more tasks as audio-to-audio (e.g. Recasting conditions: performance rendering, instrumentation. Recasting outputs: transcription).
5. BabyLM v.s. OpenWebText: How Does Child-Oriented Language Affect Syntax Acquisition for Language Models?

   Mini course project for NLP 705: Current Topics in NLP

   Yichen Huang

   2023

   TL;DR PDF Slides GitHub

   • Under the same small data budget, how does developmentally plausible data (child-oriented conversational texts) compare with web scraps in terms of LMs’ syntax acquisition?
   • Our web-scrap-trained LM demonstrates stronger syntactic performance, the acquisition of which happens early on in the training process.
   • Controlled experiments show that this superior performance is strongly associated with vocabulary and, to a lesser extent, with sentence length.

2022

1. UNITER-Based Situated Coreference Resolution with Rich Multimodal Input

   DSTC10 workshop at AAAI 2022

   Yichen Huang, Yuchen Wang, and Yik-Cheung Tam

   2022

   TL;DR arXiv GitHub

   • In a conversational situation, how can we determine which object is mentioned based on rich, multimodal information (text, vision and object knowledge base)?
   • Based on a UNITER vision-language backbone, we inject dialogue, object and scene information through adaptors.
   • We devise a method to encompass relative positions between objects through attention bias.
   • Our solution ranks second in the official evaluation at DSTC10.
2. Faster Sequence-to-Sequence Symbolic Music Generation with Rule-Augmented Edit-Based Models and Knowledge Distillation

   Unpublished research

   Yichen Huang, and Gus Xia

   2022

   TL;DR Slides GitHub

   • Specific tasks on symbolic music (e.g. harmonic style transfer) can be approximated with rule-based transformations and refined with minor edits. Can we use this property to accelerate autoregressive models for closer-to-real-time applications?
   • Rule-based transformation + semi-autoregressive edit model with redesigned edit operations, trained with a distillation objective from a slower autoregressive model.
   • Early experiments: 4x faster inference than the autoregressive teacher model, but there exists a large performance gap.
3. An AI-Empowered Piano Performance Interface for Non-Pianists

   Unpublished research

   Yichen Huang, and Gus Xia

   2022

   TL;DR PubPub Demo GitHub

   • Harmonic style transfer allows interactions with music based on high-level chord and texture controls. This can be leveraged to empower music amateurs.
   • An intuitive real-time interface for piano performance and a guided explorative interface for learning.
   • Small-scale user studies validate that our system is expressive and engages music skills.
4. <a id= style="display: block !important; position: relative !important; top: -90px !important; visibility: hidden !important;"></a>

   A Frankensteined Gaggia Classic Pro

   Accepted on my kitchen counter, 2022

   Yichen Huang, and Daniel Chin

   2022

   TL;DR

   • The Gaggia Classic Pro is an excellent budget single-boiler espresso machine. It, however, lacks precise monitoring and control for temperature and flow.
   • We modified one such machine with a PID controller, pressure gauge, piggy-backed flow control knob, and several quality-of-life features, achieving consistently superior shot quality and outperforming budget single-boiler and heat exchanger baselines.
   • Additionally, we implement an over-caffeination prevention mechanism where the pressure gauge’s connector starts to leak after the third consecutive shot.

   

5. <a id= style="display: block !important; position: relative !important; top: -90px !important; visibility: hidden !important;"></a>

   Music X Lab led by Gus Xia: Expressive Music Performance with AI

   A concert at the Steinway Dubai Community Concert Hall

   Gus Xia, Ziyu Wang, Daniel Chin, Yichen Huang, and Other Music X Lab Members

   2022

   TL;DR YouTube

   • A concert showcasing some of the earlier works at Music X Lab, including the piano interface.
   • Gus made me program the UAE national anthem into the piano interface.
   • The melody of the aforementioned anthem is stuck in my head to this day.

2021

1. Meme Caption Generation

   Course project for CSCI-SHU 376: NLP

   Yichen Huang, and Yuchen Wang

   2021

   TL;DR PDF GitHub

   • Meme generation differes from general image captioning in that one image can have correspond to a diverse range of captions. How can we generate captions that are both specific and diverse?
   • We propose a captioning pipeline with built-in stochasticities for diverse outputs and a CLIP-based reranking mechanism to ensure specificity.