Deep Learning for Natural Language Processing (Recruiting)
Description
Deep learning methods have been applied for solving a wide range of natural language processing (NLP) tasks such as dialog, summarization, and question answering.
You can suggest any NLP tasks you want to attack. The
recent publications of our lab would help to choose the task.
Candidate Qualifications
- Strong attitude for the investigation
- Basic python skills (required)
- Understanding of recent algorithms in NLP literature. e.g., BERT.
- Experience in implementing algorithms using PyTorch or Tensorflow
Expected Internship Period
- Minimum three months - excluding period for the preliminary study (Pytorch, base ML, NLP basics)
Contact
Development of AI Technology for Process-Centered Assessment for Customized Education (Recruiting)
Description
The purpose of this research is to go beyond merely using the latest AI technologies, aiming to develop core technologies for providing customized and reliable education. Duplicate applications are not allowed, and candidates should have a foundational knowledge in the relevant fields, with a focus on research and publication of related technologies and papers. In each category, candidates have the flexibility to propose their own research topics.
1. Multi-Modal Hybrid Task-Oriented Dialogue System
Current large language models (LLMs) struggle to accurately capture user intent during interactions, which is a significant limitation preventing agents from providing appropriate assistance within specific contexts. Moreover, LLMs often blindly agree with demonstrations and find it difficult to accurately interpret user intent. This project aims to research a hybrid task-oriented dialogue system (TOD) that effectively utilizes LLMs to enhance existing TOD models and better achieve user goals.
2. Multilingual Training and Machine Translation Model
Although many multilingual models have emerged, they commonly perform poorly with low-resource languages and fail to reflect cultural differences. Additionally, as LLM-based machine translation models compete with traditional machine translation models (e.g., NLLB), this project aims to analyze the strengths and weaknesses of each and develop a hybrid model.
3. Mathematical Reasoning, Multi-Modal Reasoning, and Hallucination
Current LLMs show weaknesses in mathematical computations and complex reasoning due to issues like hallucinations and the characteristics of neural databases. This project aims to analyze the mathematical capabilities of current LLMs and enhance their mathematical reasoning abilities through various techniques, such as semantic parsing, beyond simple sampling and Tree of Thoughts (ToT) decoding.
4. Adaptation of Diffusion in Computational Linguistics and NLP
This program is designed for individuals passionate about advancing the field of Natural Language Processing (NLP) through the innovative use of diffusion models. The program focuses on exploring and integrating diffusion models into various NLP applications to enhance the controllability. This project aims to research new diffusion models that ensure text quality, diversity, and semantic quality by combining continuous and discrete diffusion models, with the goal of developing the next architecture beyond current models like ChatGPT.
5. AI Trustworthy: Fairness and Ethical Preference
As the anthropomorphism of LLMs can place users in perilous situations beyond discomfort, including those that may encourage suicidal thoughts, it is crucial to flexibly detect the toxicity of generated responses, moving beyond static definitions to accommodate various contexts. In this regard, we aim to develop models that generate responses that align with social norms and moral standards in response to sensitive input sequences.
Candidate Qualifications
- Basic python skills (required)
- Understanding of Basic Deep Learning, Probability Theory, Recent NLP (required)
- Paper analysis & review capability (required)
- Research experience relevant to internship topics (required)
Expected Internship Period
- Minimum three months - excluding period for the preliminary study (advanced NLP)
Contact
- If applicants have any questions regarding the topic or publication, please contact hyukhunkoh-ai@snu.ac.kr
Ensuring Privacy in Large Language Models: Designing NLP Algorithms with Differential Privacy Guarantees (Recruiting)
Description
As interest in Large Language Models keeps growing, concerns about privacy leakage also keep increasing. As differential privacy becomes the gold standard for privacy guarantees, we aim to design various NLP algorithms with differential privacy (DP) guarantees while producing high-quality texts. Furthermore, we aim to verify that LLMs equipped with DP guarantees are robust against various privacy threat models.
Related Papers
Candidate Qualifications
- Familiar with recent LLMs
- Familiar with deep learning tools (PyTorch, Transformers)
- Fundamental probability theory and linear algebra knowledge
Expected Internship Period
- Minimum three months- excluding period for preliminary study
Contact
- seonghojoo (seonghojoo@snu.ac.kr)
