I am in search of a Ph.D. position in 2024, preferably in America.

I am fascinated by reasoning and dedicate to causality described in pre- and post-conditions (referring to the PROC2PDDL paper). If you are interested in this, feel free to email me.😃

Highest Education:

 MSE in Data Science at University of Pennsylvania; GPA: 3.97

Research Interests:

1. Reasoning in Natural/Symbolic Language
2. Interdisciplinary in NLP, CV and Robotics

Expertise:

1. Learning Science and Cognitive Science (6 years of experience, B.S., M.Ed)
   • Human Learning and Knowledge Storage
   • Learning Task Design
2. Natural Language Processing (3 years of experience, MSE)
   • Event Extraction
   • Event Reasoning in Natural and Symbolic Language
     e.g. Schema Generation, Event and Entity Pre- and Post-condition Generation

Publications

• Human-in-the-Loop Schema Induction. Tianyi Zhang*, Isaac Tham*, Zhaoyi Hou*, Jiaxuan Ren, Liyang Zhou, Hainiu Xu, Li Zhang, Lara J. Martin, Rotem Dror, Sha Li, Heng Ji, Martha Palmer, Susan Brown, Reece Suchocki, Chris Callison-Burch. ACL Demo 2023

• Argument Role Labelling Question-Answer Pair Data Augmentation. Tianyi Zhang, Elior Sulem, Dan Roth. 2023, In submission

• PROC2PDDL: Predicting Domain Definitions Based on Natural Language for Symbolic Planning. Tianyi Zhang*, Li Zhang*, Zhaoyi Hou, Ziyu Wang, Yuling Gu, Peter Clark, Chris Callison-Burch, Niket Tandon. 2023, In submission
  Not publicly available currently.

Education

• MSE in Data Science, Jan. 2021 - Dec. 2022
  University of Pennsylvania, Philadelphia, America
• M.Ed in Learning Science and Technology, Sept. 2018 - Dec. 2019
  University of Pennsylvania, Philadelphia, America
• B.S in Educational Technology, Sept. 2014 - Jun. 2018
  Beijing Normal University, Beijing, China

Research Experience

1. Research Assistant: NLP Group at UPenn, May. 2022 - Jul. 2023
   • Event schema generation using GPT-3

     • Design the scaffold prompts (cause, plan, procedure, effect, etc.) for GPT-3 to generate inclusive events
     • Apply SRL and constituency parsing to summarize and extract structured events
     • Build schema graphs by adding temporal relations to the events
     • Ground the graph nodes into the ontology through the semantic inference or the semantic similarity
     • Outcome: "Human-in-the-Loop Schema Induction" paper

   • Natural language to symbolic language translation for reasoning

     • Focus on open-domain natural language (wikiHow) to symbolic language (PDDL) generation with GPT-4
     • Decompose the task into three stages: extraction, inference, and translation
     • Identify strong text extraction and entity states inference abilities with increasingly complex wikiHow text (~5000 words)
     • Acknowledge a weak translation capability to predefined symbolic pattern
     • Improve the performance using CoT and instructions on translation.
     • Outcome:"PROC2PDDL" paper

2. Research Assistant: Cognitive Computation Group at UPenn, Mar. 2020 – Dec. 2022
   • Event trigger identification and classification using sequence tagging

     • Build 1-of-N (72) joint model and BIO identification + event type classification pipeline model
     • Transfer learning with target language auxiliary dataset, e.g. OntoNotes Arabic

   • Event argument identification and classification using QA

     • Design fixed questions for each argument role and convert argument role labeling task to question answering task
     • Build has-and-no-answer joint model and has/no answer classification + has answer identification pipeline model
     • Transfer learning with auxiliary QA dataset, e.g. SQuAD, QAMR

   • Event data augmentation using answer extraction (AE) and question generation (QG)

     • Train AE-QG T5 model to extract QA pairs from unlabeled event text
     • Train AEwSRL-QG Bert-T5 model to extract QA pairs from unlabeled event text
     • Evaluate on joint-QA and pipeline-QA model and prove effectiveness of the above data augmentation approach
     • Outcome: "Argument Role Labelling Question-Answer Pair Data Augmentation" paper