Keqin Peng (彭科钦)
|
|
tel: +44 7432 164777 / +86 195 2048 2118
About Me
I graduated in July 2022 from Beijing Normal University, School of Artificial Intelligence, earning a Bachelor of Science degree in Computer Science and Technology. Currently, I am pursuing a Master's degree in Robotics and Artificial Intelligence at University of Glasgow, UK.
My primary research interests include the application and improvement of natural language processing methods, with a focus on document-level information extraction and mining. My specific research areas encompass named entity recognition, entity relation extraction, event extraction, and text summarization.
Education
|
University of Glasgow (2024.09 ~ 2025.09)
|
|
Beijing Normal University (2018.9 ~ 2022.7)
|
Publications
-
U-Net with depthwise separable convolutions and big kernels for document-level relation extraction
Rongen Yan, Keqin Peng and Depeng Dang
2023 11th Transactions of the Association for Computational Linguistics. [Reviewing]destails
The article will be presented to you when it is published -
Document-level relation extraction with less redundant image-style feature
Rongen Yan, Keqin Peng and Depeng Dang
2023 11th Transactions of the Association for Computational Linguistics. [Reviewing]destails
The article will be presented to you when it is published
-
Research on Multi-Document Summary Generation Method Based on Event Structure
Keqin Peng and Xuesong Wang
Beijing Normal University Graduation Thesis[PDF]details
- Traditional multi-text summarization methods often suffer from the problem of losing information between text segments, which limits the output of traditional automatic summarization neural network models This limits the output effect of traditional automatic summarization neural network models. We introduce a hierarchical Transformer model to connect and encode the sorted and filtered paragraphs hierarchically, which improves the quality of the output summary. The quality of the output summary is improved.
- The sorting algorithm based on LSTM model can calculate the correlation between paragraphs and input headings, which can be used as the basis for sorting and filtering to improve the model's capacity for input size. The model's capacity for input size is improved.
- The hierarchical Transformer structure learns contextual information about the characters within and between paragraphs separately and portrays the event structure.
- Use Beam Search and length normalization methods to improve the length and fluency of output sentences.
Work Experience
Project Development
- Intelligent Financial and Taxation System
- Developed an automated accounting bot module using the Python and Selenium tech stack, enabling web automation and local file processing.
- Participated in requirements discussions to help define functionality scope and technical choices.
- Designed and implemented the logic of the automated preprocessing bot module, utilizing PyPDF2 and Pandas to process document data, improving automation and data accuracy.
- Data Butler System
- Developed modules such as Configuration Center, Project Workspace, Homepage, and Business Management using Java and the Spring Boot framework, ensuring high system performance and maintainability.
- Contributed to the logic design of the Project Center module, applying RESTful API design principles to enable efficient management of project lists and assigned tasks.
- Designed the database schema, optimized query performance, and used the MySQL database management system to support fast access to large amounts of data, achieving efficient ORM (Object-Relational Mapping) with MyBatis.
- Intelligent Financial and Taxation System
Research on Enterprise AI Transformation
- Designed theoretical knowledge and practical assessment processes for large model application capabilities.
- Drafted an AI development recommendation report for the company: analyzed the current situation and competitive landscape, selected suitable technical routes and development paths; provided detailed analysis of AI development strategies, investment scale, product output models, and cash flow cycles; and drafted basic development processes and standards for the company's AI development team.
NLP Technology R&D and Applications
- Created a medical experiment text dataset: collected 200,000 operation texts from laboratories in various fields of global medical experiments; used a combination of NER + RE + keylist methods to clean and organize data, ensuring uniform formatting and valid content; and generated a high-quality dataset at low cost using a combination of few-shot pre-training and automated labeling models.
- Conducted technical validation and pipeline development: validated existing technologies such as NER, RE, SRL, and KG to provide feasible technical solutions for tasks.
- Proposed new tasks and designed models: introduced the task of extracting key step flowcharts from natural language texts and designed corresponding models.
- Redesigned end-to-end models: optimized model design based on task requirements to address issues of error accumulation and error propagation.
- Deployed new technologies and collaborated with medical laboratories.
Hunan Changsha Bank Digital Technology Co., Ltd.
2024.01 ~ 2024.09
Position: Backend Development Engineer, Development Division 1
Institute for AI Industry Research, Tsinghua University
2022.07 ~ 2023.07
Position: NLP Research and Development Engineer
Patent
A document-level relationship extraction method, system, device and storage medium
Rongen Yan, Keqin Peng
National Invention Patent, Under Review
Other Projects
Research task publishing platform
Database leader; 2021.3~2021.5
Independent Exploration Project for Undergraduates of Beijing Normal University
American College Mathematical Modeling Competition
Programmer; 2021.01-2021.03