Active Interpretation of Disparate Alternatives
University of Florida

• Project Overview:
  Played a dual role in the DARPA-sponsored "Active Interpretation of Disparate Alternatives" (AIDA) project, focusing on developing an advanced search engine for analyzing alternative hypotheses in event-centric knowledge graphs.
• Contributions:
  • Engineered a novel two-tier graph searching algorithm, optimizing knowledge graph exploration at both mention and cluster levels. This innovation led to a significant 25% improvement in the final F1 score.
  • Developed a sophisticated graph clustering algorithm. This algorithm enhanced the differentiation of alternative hypotheses by quantifying both structural and semantic distances between candidates, resulting in a 20% improvement in clustering quality (v-measure).
  • The system demonstrated exceptional performance, achieving top results in the NIST TAC SMKBP2020 evaluation, a benchmark test in knowledge-based performance.

"More Than Reading Comprehension: A Survey on Datasets and Metrics of Textual Question Answering (pdf) "
University of Florida

• Project Overview:
  Authored the comprehensive research paper titled "More Than Reading Comprehension: A Survey on Datasets and Metrics of Textual Question Answering." This research involved a thorough analysis of Textual Question Answering (QA), aiming to provide natural language answers using unstructured data, mainly focusing on machine reading comprehension (MRC).
• Contributions:
  • Developed a new taxonomy for textual QA, considering novel application scenarios.
  • Covered 7 major subtasks of textual QA, including classical MRC, four novel MRC tasks (conversational QA, multi-hop QA, long-form QA, and cross-language QA), ODQA, and commonsense QA.
  • Conducted comparative analysis across multiple facets of benchmark datasets.
  • Provided detailed statistics for each dataset and described evaluation metrics for each benchmark.
  • Addressed current trends and challenges in textual QA, suggesting directions for future research to tackle unsolved tasks and propose more challenging datasets.

Our work is accepted by SIGIR'23 ( (pdf) )
University of Florida

• Project Overview:
  Spearheaded the "MythQA" project at the University of Florida, an innovative endeavor in the realm of machine learning. This project was centered on the development of large-scale check-worthy claim detection using multi-answer open-domain question answering (QA).
• Contributions:
  • Led a team in the creation of a unique benchmark dataset, utilizing advanced data collection methods via the Twitter API.
  • Innovated in the development of new evaluation metrics and a specialized data schema, tailored specifically for this project, along with detailed annotation guidelines.
  • Successfully developed and implemented a user-friendly annotation tool and a comprehensive end-to-end pipeline. This pipeline played a crucial role in assessing the impact of various components, such as information retrieval, machine reading comprehension, and distinct answer selection, significantly enhancing the project's overall performance.

Open-Domain Multi-Hop Dense Sentence Retrieval (under submission)
University of Florida

• Project Overview:
  Contributed to the cutting-edge research paper "M3: A Multi-Task Mixed-Objective Learning Framework for Open-Domain Multi-Hop Dense Sentence Retrieval". This project aimed to enhance dense retrieval performance by merging contrastive learning with multi-task and mixed-objective learning frameworks.
• Contributions:
  • Developed the M3 system, an advanced, recursive multi-hop dense sentence retrieval system using a new method for dense sentence representation learning.
  • Proposed the M3-DSR method, vastly surpassing existing benchmarks in sentence-level retrieval.
  • Introduced an innovative heuristic hybrid ranking algorithm that effectively combines single-hop and multi-hop sentence evidence, thereby improving performance.
  • Achieved groundbreaking results in multi-hop retrieval and fact verification on the FEVER dataset, setting a new standard in open-domain fact verification.