Duong Tung Ta PHD student

Large Language Model Auto-metacognitive prompting

LLM and Metacognition

In this project, we tasked GPT-3.5-turbo with generating metacognitive prompts for mathematical and logical reasoning, enabling the model to solve problems by breaking them down into step-by-step solutions. Furthermore, we crafted GPT-3.5-turbo prompts with 1-shot metacognitive demonstration to evaluate the provided step-by-step math and reasoning solutions against reference solutions from the corresponding solvers. This approach enhanced the model's accuracy on these evaluation tasks when compared to the baseline model employed in the Factuality Evaluation of large Language Models (FELM) benchmark.

Supervised by: Dr. Tim Oates

Deep Reinforcement Learning Multi-Agent Hierarchical Representation

Hierarchical Reinforcement Learning

This work was the focus of my summer research project. I conducted an in-depth study of 15 multi-agent and graph reinforcement learning papers to conceptualize the system. Additionally, I analyzed 3 graph reinforcement learning domains, namely transportation, coach management, and job scheduling. I implemented a deep reinforcement learning (DRL) model for job scheduling and interpreted the results to facilitate multi-agent decision-making in the scheduling of multiple tasks.

Supervised by: Dr. Tim Oates

Natural Language Processing Natural Language To SQL

Natural Language to SQL

Our project involved constructing a Vietnamese transaction dataset and tokenizing the query questions and SQL statements within it. We executed the Data-Agnostic RoBERTa Text-to-SQL model on an example English dataset to benchmark its evaluation performance. Subsequently, we advanced our experimentation by employing the Rat-SQL model and devising a BERT-multilingual model tailored to the Vietnamese dataset. The results demonstrated robust real-time query processing capabilities and considerable accuracy in translating Vietnamese natural language questions into executable SQL commands.

Supervised by: Thai Van Doan

Knowledge Graph Similar Company Searching

Similar Company Searching

This project involved constructing an industry tree and generating a Floyd Warshall matrix to evaluate the distance between industries, establishing relationships for this feature. Initially, we implemented the Random Walk with Restart algorithm to rank the top companies similar to a given query company. To accelerate the ranking process, we integrated company features and labeled relations into a knowledge graph, obtaining triplets. We then trained the TransE model on these triplets to retrieve graph embeddings and utilized the HNSW search algorithm. Furthermore, we improved search accuracy by 5% by employing the ComplEx model and the scaNN search algorithm.

Supervised by: Thai Van Doan

Teaching Teaching Assistant

Teaching Assistant

As a graduate student at UMBC, I had an honor to be a teaching assistant for courses instructed by Dr. Frank Ferraro. In this role, I was responsible for grading 4 assignments in CMSC 678 - Introduction to Machine Learning. Additionally, I conducted office hours to provide guidance and support to the students in this course. My duties extended to another course, CMSC 673 - Introduction to Natural Language Processing, where I evaluated 4 assignments, a mid-term examination, and held office hours to address students' queries and concerns for students in the class.

Supervised by: Dr. Frank Ferraro