User 9593 1475392706

Kai-Chou, Yang

As a Kaggle Competition Master and a winner of international data science challenges, I am experienced in machine learning, deep learning and related frameworks such as PyTorch.


My research focuses on natural language processing (NLP), where I have released 11 open-source projects such as MianBot (700+★ on Github) and presented certain academic papers on top conferences like ACL, AAAI, CIKM, and WSDM.

International Awards

For the following achievements, I am the first author as well as the team leader.

2nd Place, CIKM Cup: Cross-lingual Short-text Matching Challenge

  • Proposed two densely-connected architectures, CPRNN and DACNN, for sentence pair modeling.
  • Fused semantic features from different levels to create diversity intra-models.
  • The solution has been oral presented on CIKM 2018 in Turin, Italy.

3rd Place, WSDM Cup: Fake News Classification Challenge

  • Implemented various NLI networks like ESIM and injected world knowledge using BERT.
  • Proposed a disagreement-aware model based on the single-word attention.
  • The paper has been oral presented on WSDM 2019 in Melbourne, Australia.

4th place, Google AI: Gendered Pronoun Resolution Competition

  • Leveraged the information redundancy from BERT and extracted features from the optimal layer.
  • Proposed a multi-heads Siamese semantic scorer for answer selection.
  • The paper has been presented on ACL 2019 in Florence, Italy.

Kaggle Competition Master, Ranked top 0.2% (233/114,366)

  • Top 1% (4/838), Gendered Pronoun Resolution Competition.
  • Top 1% (27/4,550), Toxic Comment Classification Challenge.
  • Top 3% (30/1,449), CareerCon 2019 - Help Navigate Robots.
  • Top 4% (103/3,165), Jigsaw Unintended Bias in Toxicity Classification.
  • Top 6% (223/3,633), CommonLit Readability Prize. 
  • Top 10% (384/3,946), TalkingData AdTracking Fraud Detection Challenge.

Work Experience

Taiwan AI Labs, Machine Learning Engineer, Sep 2019 ~ Now

Question Answering System

  • Propose a conditional question generator with mT5 for controllable QA data augmentation and as the base of dense retrieval, which improves recall@50 from baseline model by 16%.
  • Build a generative pseudo labeling pipeline using a open-domain passage retriever and machine reader, which improve the nDCG@10 by 4.2 - 9.7, on various domains.
  • Build an efficient passage re-ranker based on tiny-bert with a time-series based clustering framework for effective negative passage sampling.
  • Leverage FinBERT on QA analysis and slot filling for fintech dialogue system.

Natural Language Understanding 

  • Implement a document encoder with self-contrastive learning and a document clustering algorithm, which is scalable for million scale of streaming data.
  • Implement a GROVER-like generator as the backbone for topic detection, article rewriting, and tag generation.
  • Propose a semi-automatic framework for fake-news identification, which gathers evidence from event properties, user behavior and textual features.
  • Propose a SOTA Chinese typo correction system based on a boosting loop of automatic speech recognition and text to speech for weak supervision.
  • Build a general-purpose NLP training pipeline for team use involving data augmentation, data regularization, and unsupervised domain adaptation.

Education

Master in Department of Computer Science, NCKU                                         GPA: 4.30

  • Honorary member of the Phi Tau Phi Scholastic Honor Society. (Ranked 1st among all graduates.)
  • As a teaching assistant for Introduction to Data Science, Data Mining and Discrete Mathematics.
  • As a speaker / teaching assistant for introduction lectures of machine learning.

Bachelor in Department of Computer Science, NCKU                                      GPA: 3.92






  • Academic excellence awards 2016.
  • Academic excellence awards 2015.
  • Honorable mention on the graduation exhibition.
  • Research assistant on a question answering system project for the Ministry of Science and Technology.

Side Projects

I list some of my project experiences. You can refer to my Github for the other interesting ideas.

Mianbot

  • Got 700+ stars and 200+ forks on Github.
  • Implemented the hierarchical keywords matching using word2vec.
  • Implemented the IR-based searching module to support chit-chat.
  • Allow user to define customized scenarios with JSON.
  • The extracted QA pairs were released in PTT-Gossiping-Dataset, a widely-used Chinese chit-chat corpus.
Paragraph image 00 00@2x

NCKU Smart-Life LineBot

  • A Linebot that helps solve trivial matters such as restaurant recommendation.
  • The dialogue system is based on LUIS for intent classification.
  • The backend was built with Django / Flask (new version) and host on Heroku.
  • The backend is connected with Line server using the web API.
Paragraph image 00 00@2x

Knowledge & Skills

  • General Machine Learning
    • Classification, Regression, Clustering, Boosting, Feature Engineering.
  • Natural Language Processing
    • Sentence Pair Modeling: Natural language Inference, Machine Reading Comprehension, Sentence Similarity
    • Text Classification / Regression / Clustering
    • Deep contextual representation (ELMO / BERT / XLNet / ELECTRA / RoBERTa / ERINE2.0 / BigBird / T5)
  • Recommendation System
    • Factorization: Matrix Factorization, Factorization Machine, DeepFM
    • Graph Embedding: DeepWalk, Node2Vec, item2Vec

Publication

  1. Fake News Detection as Natural Language Inference. Kai-Chou Yang; Timothy Niven; Hung-Yu Kao. WSDM Cup 2019
  2. Fill the GAP: Exploiting BERT for Pronoun Resolution. Kai-Chou Yang; Timothy Niven; Tzu Hsuan Chou; Hung-Yu Kao. ACLWS'19
  3. Generalize Sentence Representation with Self-Inference. Kai-Chou Yang; Hung-Yu Kao. AAAI 2020
  4. The Prevalence and Impact of Fake News on COVID-19 Vaccination in Taiwan: A Retrospective Study of Digital Media. Yen-Pin Chen; Yi-Ying Chen; Kai-Chou Yang; Feipei Lai; Chien-Hua Huang; Yun-Nung Chen; Yi-Chin Tu. JMIR
Powered By
CakeResume