Worked on Writing Assistant related projects in two main parts:

• AI Writer: An application which aims to increase diversity of an article as well as reduces efforts of writing “filling text” for human. (This was the project used for the intern conversion)

  • Continuous Writing:

    • GPT2

  • Rewriting:

    • Paraphrasing

      • UniLM (SimBERT)

    • Back-translation

      • Bing Translator
      • Google Translation

    • Information-Retrieval-based

      • Elastic Search
      • Approximate Nearest Neighbor (annoy)

    • Style Transfer

      • Style Transformer

• Value Understanding: Built a numerical extractor which can extract quantity fact from raw text.

  • Designed an annotation guideline especially for Chinese quantity extraction.
  • Communication with labeling company and annotate more than 2000 article data to construct the training dataset from scratch.

  • Designed two major approaches namely “NER Combine” and “Quantity MRC”.

    • NER Combine: Combine spans with label extracted from NER model with an scope-based rule-based algorithm
    • Quantity MRC: Construct query for each slots based on extracted Quantity

  • Post-processing modules that able to deal with complex sentences especially the “respectively cases”.

  • Got used as back-end of three different projects

    • Writing Assistant (mainly finance): Including value consistency and value recommendation
    • Medical thesis analyser
    • A WeChat mini program
      

Interviewed 5 internship candidates (after getting the return offer).

• Generation of slides from academic paper
• Math word problem generation
  

Doing case of Anti-healthcare fraud and Medical record analysis.

Including research of:

• Information Extraction
• Named-entity Recognition
• Relation Classification
• Knowledge Graph

PKU Thesis: Design and Implementation of Chinese Document Numerical Fact Extraction

Ecojoy

We want to solve the problem of “Toy waste”. Excessive pollution not only affects the physical environment of future generations but also cultivates children who do not cherish resources, which has a major impact on the world. We hope that through a very simple way, every old toy will no longer be piled up at home or enter the landfill, but can also become a resource for others. We have software engineering, social education, and economics background. Observing that the problem of toy waste is becoming more and more serious, it is readily available and cheap, becoming a quick solution for most parents to deal with their children. We believe that as long as the sharing and acquisition methods are simple enough, it can immediately improve the situation of excessive waste. Through subscription to become members of Ecojoy App, you can easily share excess toys at home, and through the perfect toy information and rating system on APP, users can easily find suitable toys to meet their needs and achieve toy sharing and reuse.

Facebook Page

In this competition, each sample is a time-series data of a concrete pump vehicle. The goal is to predict the likelihood of each data sequence that whether a machine might fail. I used LightGBM and reach Top 5% in rank.

Source Code