Hwiyeol Jo

Currently at NAVER Search US

* Reach me by email: hwiyeolj@gmail .com

Research Interests

Artificial Intelligence with Cognitive Science

Machine Learning for Natural Language Processing to investigate human minds

• Research Keyword
  • Knowledge Transfer through Representation Learning
    • incorporating (lexical) semantics, information retrieval,
  • General machine learning frameworks like meta-level representation learning.

Education

• M.S. in Computer Science & Engineering (2015 – 2017) , Seoul National University, Republic of Korea
  • Master Thesis: Data Augmentation Technique with Knowledge Extraction for Text Question Answering by Deep Neural Networks [Advisor: Byoung-Tak Zhang]
• B.S. in Computer Science & Engineering (2008 – 2015), Konkuk University, Republic of Korea
• Military service at Third Republic Of Korea Army; TROKA (2011 – 2013)

Research Experience

Google Scholar

(Accepted)

Jo, H. (2023). Self-supervised Post-processing Method to Enrich Pretrained Word Vectors. EMNLP2023(Findings) [Selected for Presentation]

• Extended extrofitting without requiring any external lexicon; Self-supervised extrofitting
• Improved on various benchmarks like word similarity tasks in various language, DST, and text classification, showing the quality of the approach

Jo, H. (2023). A Self-Supervised Integration Method of Pretrained Language Models and Word Definitions. ACL2023(Findings)

• Dissected BERT representation with respect to the basic form of semantics, formal definition
• Developed DefBERT; The method to inject formal definition into contextualized models
• Proposed DefRank, SenseRank tasks to measure the ability to incorporate formal definitions

Kim, J., Kim, H. J., Cho, H., Jo, H., Lee, S. W., Lee, S. G., … & Kim, T. (2022). Ground-Truth Labels Matter: A Deeper Look into Input-Label Demonstrations. EMNLP2022

• Investigated input-label importance of large-scale pretrained language model (e.g., GPTs) in few-shot settings

Jo, H., Kang, D., Heads, A., & Hearst, M. (2021). Modeling Mathematical Notation Semantics in Academic Papers. EMNLP2021(Findings)

• Proposed two notation prediction tasks to test models’ mathematical semantics understanding
• Presented a fine-tuned model mathBERT for the tasks, which shows much better performance than pretrained BERT
• Showed in-depth analyses of the results

Jo, H.*, Lim, J.*, & Zhang, B. (2021). Devil’s Advocate: Novel Boosting Ensemble Method from Psychological Findings for Text Classification. EMNLP2021(Findings) (*Equal Contribution)

• Applied a novel ensemble method based on group discussion theory, Devil’s Advocate

Lim, J.*, Jo, H.*, Zhang, B., & Park, J. (2021). Passive Versus Active: Frameworks of Active Learning for Linking Humans to Machines. The 43rd Annual Meeting of the Cognitive Science Society. (*Equal Contribution)

• Analyzed active learning on both Humans and Machines
• Proposed two experiments (human and machine, respectively) in order to compare active and passive learning.

Lim, J.*, Jo, H.*, Zhang, B. & Park, J. (2020). Human-Like Active Learning: Machines Simulating Human Learning Process. NeurIPS 2020 Workshop on BabyMind (*Equal Contribution) [Spotlight Presentation]

• Extended to CogSci2021

Jo, H., & Cinarel, C. (2019). Delta-training: Simple Semi-Supervised Text Classification using Pretrained Word Embeddings. EMNLP2019

• Suggested a semi-supervised training framework using knowledge gaps between two classifiers

Jo, H., & Ryu, J. (2018). Psychological State in Text: A Limitation of Sentiment Analysis. IJCAI-ECAI Workshop on AI and Computational Psychology: Theories, Algorithms and Applications (CompPsy). [Extended Abstract]

• Presented the limitation of current sentiment analysis method; The model is not able to model psychological state of the person.

Jo, H., & Choi, S. J. (2018). Extrofitting: Enriching Word Representation and its Vector Space with Semantic Lexicons. In Proceedings of the 3rd ACL Workshop on Representation Learning for NLP (RepL4NLP).

• Developed extrofitting that injects lexical semantics into word vectors
• Extrofitting results in word vector generalization

(Preprint/Under Reviews)

Jo, H., & Zhang, B. (2019). Ruminating Word Representations with Random Masker. submitted

• Proposed an iterative training method for better word representation and for model regularization

Full Publication Lists with Psychology related papers are omitted.

Work Experiences

[2024.02 – Current] Research Scientist & Engineer, Search CIC, NAVER, Republic of Korea

• Improving Search Engine results with self/unsupervised-method by using LLMs’ internal knowledge.

[2022.11 – 2024.02] Research Scientist & Engineer, Search US, NAVER, Republic of Korea

• Large Language Model backbone (with deepspeed+Megatron) and its applications (with finetuning+prompting)
• Improving Search Engine results, using word/model semantics
• Input-level sensitivity checking for AI safety

[2021.07 – 2022.11] Research Scientist & Engineer, Clova CIC, NAVER, Republic of Korea

• General MRC Models to extract information related to query from given documents
• Make it component model of information retrieval filtering and question generation model

[2019.07 – 2021.07] Research Assistant, Institute of Computer Technology, Seoul National University, Republic of Korea

• Leader of Video Intelligent Platform Team; Build full stack of AI Demo system
• Demo video: https://youtu.be/ElzgeQIZF0s

[2020.03 – 2021.02] Lecturer, Department of Computer Education, Seoul National University of Education, Republic of Korea

[2020.03 – 2020.08] Lecturer, Computer Engineering Department, Hongik University, Republic of Korea

[2019.02 – 2019.07] Researcher, AI Division, LG Sciencepark, Republic of Korea

• Research regarding data augmentation
• PoC on recent models/algorithms

[2017.07 – 2019.02] Researcher, Artificial Intelligence Lab, LG Electronics CTO, Republic of Korea

• Word vector generation and post-processing for target tasks (e.g., Chatbot embedded in TV products, and disorder symptom classifier for customer service center)
• PoC on recent models/algorithms and related company (e.g., Luminoso)
• Research regarding word vector specialization and document representation

Teaching Courses

Lecturer, Deep Learning and its Application (NLP), Hanyang University, Spring 2024.

Lecturer, App Programming, Seoul National University of Education. Fall 2020.

Lecturer, Introduction to Computer Engineering, Hongik University. Spring 2020.

Lecturer, Computing and Mathematics, Seoul National University of Education. Spring 2020.

Academic Services

ACL: 2023, 2024

EMNLP: 2021, 2022, 2023

CogSci: 2022, 2023, 2024

COLING: 2022, 2024 (with LREC),

NAACL: 2024

CoLM: 2024

Talks

Post-processing Approaches Toward Better Embeddings, Bird-of-Feather session on Embeddings in EMNLP2023, Dec 9th, 2023

AI using Python, Seoul National University of Education, Jan 16-17th, 2020.

Computational Linguistic Approaches for Sentiment Analysis, Cognitive Science Colloquium in Seoul National University, June 3rd, 2019.

Large-Scale Text Classification with Deep Neural Networks, Korea Institute of S&T Evaluation and Planning, Feb 1st, 2017.

Introduction to IoT, Drone, and Cognitive Science, National Association of Cognitive Science Industries, Seojeong University, Feb 23–25th, 2016.

Computer Skills

Python [Fluent] (I have a teaching experience as a lecturer to undergraduate students)

Pytorch [Fluent] (I have been doing research)

Tensorflow [Advanced] (I am able to understand how the codes are working, and modify them)

C, C++, Java [Intermediate] (It’s been for a long time I use them)