CV

Ph.D. in Computer Science

• Thesis Topic
  • Enhance Machine Reasoning with Human Rationales

MS in Information Technology & Web Science

BS in Computer & Systems Engineering

Objective Human Explanation Evaluation

• Proposed a prompt-based unified data format for different tasks
• Enhanced the established Simulatability score with a novel metric to evaluate explanations' helpfulness at fine-tuning and inference
• Our metric can evaluate 5 datasets with 2 models (T5 and BART) consistently, while the established metric falls short

Active Learning for Human Labels and Explanations

• Proposed a dual-model AL framework to generate natural language explanations as additional information for the prediction model
• Designed a novel AL selection strategy based on the similarity between unlabeled data and human-annotated explanations
• Justified human explanations with our AL framework are beneficial for the prediction model to perform better and converge faster

QA-Pair Generation (QAG) for Children Storybooks

• Implemented a QAG system with 1). a heuristics-based answer extraction module, question; 2). a fine-tuned BART-based question generation module; 3). a DistilBERT-based ranking module to rank and select best QA-pairs
• Outperformed 2 SOTA QAG systems on the FairytaleQA dataset in terms of the mean average ROUGE-L precision at top 1, 3, 5, and 10 QA-pairs generated per story section
• Developed an interactive story-telling web application built upon our QAG system and justified its usefulness through a user study with 12 pairs of parents and children

FairytaleQA Dataset for Children Education

• Supervised education experts with carefully designed annotation schema to create 10, 580 QA-pairs on 278 fairytale stories
• Benchmarked FairytaleQA dataset can provide helpfulness on Question Answering and Question Generation tasks by fine-tuning various SOTA language models and performing in-depth analysis

Instruction-finetune LLM for Mental Issue Detection

• Curated high-quality human-annotated datasets for mental issue detection in online communities
• Designed prompts for augmented tasks with curated datasets and instruction-fine-tuned an Alpaca and FLAN-T5 model
• Benchmarked with SOTA models (e.g., fine-tuned Alpaca-Lora, mental-RoBERTa) and conducted extensive ablation studies on prompt selection, subsampling, and transfer learning

QA Annotation Framework with Extern KG Support

• The framework supports human annotators to select a preferred concept, then retrieve and rank the most relevant Commonsense Knowledge from ConceptNet to facilitate QA-pair annotation
• Human evaluation justified that expert-created QA-pairs with our framework are preferred to the ones generated by carefully prompted GPT-3.5 few-shot approaches

Member of Upsilon Pi Epsilon

• The International Honor Society for the Computing and Information Disciplines

Member of Gamma Nu Eta

• The National Information Technology Honor Society

Reviewer

• ACL 2023 [2], EMNLP 2023 [5], ACL ARR Aug [1]

Teaching Assistant

• Computer Organization

Teaching Assistant

• Introduction of AI

Teaching Assistant

• Computer Organization