Yekyung Kim

I am a first year Ph.D candidate at UMass NLP, advised by Mohit Iyyer with a research focus on natural language processing.

Before joining UMass, I worked at Hyundai Motors Group and LG Electronics as a research engineer. I was fortunate to be selected as a specialist in AI and researched at CMU LTI as a visiting scientist mentored by Jaime Carbonell.

Email  /  Google Scholar  /  Github

My research goal is to build an efficient and trustworthy system that connects humans and machines leveraging language as a bridge. I am interested in systems that make reliable communication between humans and machines and learn knowledge with as little human supervision as possible. In particular, my research directions include:

1) Efficiency: Not all data are equally valuable in the real world for learning. How do we accurately calibrate the importance of data? How can we adjust models to new distribution resources efficiently?

2) Generalization and Robustness: Models prefer to learn by shortcuts, but often face unexpected wild inputs. How can we detect various out-of-distribution data and make models generalized better from limited data?

3) Trustfulness: Logical reasoning is a human’s hallmark capability that enables reliable communication. How can we control the model to give reliable answers expressing uncertainty? Is it possible for models to solve logic puzzles (ex., Einstein's Riddle) by comprehending problems and utilizing logical deduction? How can we teach models to reason logically with human guidance?

We propose VERISCORE, a metric for diverse long-form generation tasks that contain both verifiable and unverifiable content. VERISCORE can be effectively implemented with either closed or fine-tuned open-weight language models, and human evaluation confirms that VERISCORE’s extracted claims are more sensible than those from competing methods across eight different long-form tasks.

While long-context large language models (LLMs) can technically summarize book-length documents (>100K tokens), the length and complexity of the documents have so far prohibited evaluations of input-dependent aspects like faithfulness. In this paper, we conduct the first large-scale human evaluation of faithfulness and content selection on LLM-generated summaries of fictional books.

We propose an unsupervised learning approach to text interpolation for the purpose of data augmentation that does not require any heuristics nor manually crafted resources but learns to interpolate between any pair of natural language sentences over a natural language manifold.

We propose a data-centric framework to construct a new feature space that can capture various characteristics of datasets and novel sampling method to select a set of data points that maximizes the group informativeness.

We propose Meta-Crafting: a unified method that can capture both background and semantic shifts. Specifically, we construct a new discriminative feature space to detect OOD samples

I demonstrate that the amount of labeled training data can be reduced using active learning when it incorporates both uncertainty and diversity in the sequence labeling task.

We propose a improved unigram-level Korean NER model with sub-character level representation, jamo, which can represent a unique linguistic structure of Korean and its syntactic properties and morphological variations.

We collect users' music listening behavior from Twitter using music-related hashtags (eg,# nowplaying). We then build a predictive model to forecast the Billboard rankings and hit music.

We build on key principles of visual analytics and describe an end-to-end, visual exploration system for tweets that both presents overall summaries and supports analysis of any variations that exists in the activity