Hsuehkuan Lu

Machine Learning Engineer

Highly focused, cooperative, and with strong learning ability. Passionate about machine learning, natural language processing/understanding, and data science. Practical experiences in ML projects with Python, and GraphQL API design. Proficient in implementing algorithms and researching. 

  Taipei, Taiwan      

[email protected]

Education

2016 - 2019

Tsinghua University

Master of Science, Computer Science

Knowledge Engineering Group (KEG) Lab

2012 - 2016

National Central University

Bachelor of Science, Computer Science and Information Engineering

Work Experience

Algorithm Engineer  •  HIPR PacSoft Technologies

August 2020 - Present

  • Model MySQL 30+ tables, and Elasticsearch 10+ indices.
  • Design GraphQL APIs with Flask, and write unit tests.
  • Design NLP data processing pipeline (500M+ data) ranges from Google Scholar, world university rankings, journal rankings to institution rankings.
  • Design distributed crawling and computing systems with Dask, and asynchronous data processing methods (3-4x faster).
  • Import object-relational mapping of Elasticsearch, improving the organization of search engine indices mappings, and simplifying queries.

Projects

Impactio

  • Develop social network applications for academic researchers.
  • Design large data pipeline for institutions (750K+), journals (35K+), authors (250M+), and publications (230M+).
  • Design distributed crawling and on-the-fly merging systems for Google Scholar profile with Dask (avg. 3-5 MIN per user).
  • Design cursor-based pagination method with GraphQL to improve built-in offset-based pagination method by 10%+ efficiency.
  • Main backend APIs developer in the project (70%+).

News Miner

  • Developed news centering and trend/topic analysis applications.
  • Designed English news data processing pipeline (5K+ per day), and topical news clustering with semantic representations.
  • Adopted Word2Vec distributed representations with k-means clustering to produce text features, and designed a single-pass clustering method to merge related news.
  • Handled 100K+ news in 10 MIN, largely improved the processing performance by 250%-300% comparing to k-means clustering.

POS Tagging and Dependency Parsing

  • Combined POS tagging with dependency parsing so as to alleviate the error propagation problems and enriched sequence tagging information across tasks.
  • Experimental results on dataset Universal Dependency 2.0 achieved 81.14 with LAS (Labelled Attachment Score) averagely, while baseline scored 72.14 with LAS (11% improvement).
  • The ablation tests indicated that the POS tagging information largely improved the performance of dependency parsing by 10%.
  • Tagging system was implemented in Tensorflow with Python, and was able to annotate 1K+ sentences per MIN.

Wikipedia Data Processing and Joint Representations of Entities and Texts

  • Tackled the parallel corpus reliance on the cross-lingual problems and enriched textual information with external knowledge information.
  • Designed a weakly-supervised algorithm to produce aligned cross-lingual corpus (avg. 300K+ paragraphs per language pair).
  • Jointly modeled knowledge entities and texts with the model derived from Skip-Gram with Negative Sampling algorithm and made word vectors (300d) public.
  • Achieved 44.99 Pearson-r in SemEval-2017 Track 4a task (En-Es), while end-to-end LASER model achieved 40.87 Pearson-r (10% improvement).
  • Proposed a cross-lingual information retrievement task, and our proposed method achieved 80% Top-10 Accuracy, while the strongest baseline acquired 61% Top-10 Accuracy (30% significant improvement).

Skills

Programming Languages

Python
C++
Java


Framework and Data Analysis

Tensorflow
PyTorch
Pandas
Jupyter Lab
Dask
GraphQL

Database and Search Engine

MySQL
Elasticsearch
MongoDB

VCS and CI/CD Tools

Git
Bitbucket
Alembic
Jenkins

Publications

  1. Hsuehkuan Lu, Yixin Cao, Lei Hou, and Juanzi Li. Knowledge-Enhanced Bilingual Textual Representations for Cross-Lingual Semantic Textual Similarity. International Conference of Pioneering Computer Scientists, Engineers and Educators (ICPCSEE), 2019. CCIS Volume 1058, pages 425-440.
  2. Hsuehkuan Lu, Lei Hou, and Juanzi Li. How Important Is POS to Dependency Parsing? Joint POS Tagging and Dependency Parsing Neural Networks. Chinese Computational Lingustics (CCL), 2019. LNCS, Volume 11856, pages 625-637.

呂學寬

機器學習工程師

Machine Learning Engineer

專注力強，樂於合作，學習能力強。對機器學習、自然語言處理/理解以及資料科學充滿熱情。擅長以Python做數據分析，實作機器學習專案，以及設計 GraphQL API。具備熟練實作算法模型以及研究能力。

  Taipei, Taiwan      

Email: [email protected] 

Phone: (+886) 905193233

工作經驗

算法工程師  •  HIPR PacSoft Technologies

八月 2020 - Present

  • MySQL 30+表架構模型建立，以及 Elasticsearch 10+索引架構模型建立。
  • 利用 GraphQL + Flask 設計後端 API，並撰寫單元測試。
  • 自然語言處理資料處理流程設計（合計約5億筆資料），包含Google Scholar、世界大學排名、期刊排名，以及學術機構排名資料。
  • 設計 Dask 分布式計算、爬蟲（16+ workers），以及 Asyncio 非同步資料處理流程（約提升3-4倍效能）。

專案

Impactio

  • 針對學術研究者設計的社交網絡。
  • 設計大量資料處理流程：學術機構 (750K+)、期刊 (35K+)、作者 (250M+) 以及學術論文 (230M+)。
  • 利用 Dask 設計 Google Scholar 使用者資料分布式爬蟲方法，以及實時合併處理系統 (平均每個使用者 3-5 MIN)。
  • 設計 GraphQL cursor-based 分頁方法，優化內建 offset-based 分頁方法，效率提升約 10%+。
  • 專案後端 API 服務的主要開發者 (70%+)。

News Miner

  • 開發跨新聞媒體及趨勢、主題分析應用。
  • 設計英文新聞的資料處理流程 (每日5K+)，以及新聞主題聚類方法設計、文本語義表示學習。
  • 採用 Word2Vec 詞向量 + k-means 聚類方法生成文本特徵，並設計單向聚類方法合併相似新聞。
  • 系統平均 10 分鐘能處理 100K+ 筆新聞，而採用 k-means 方法則需要 10x 以上的時間才能達到收斂。

詞性標註與依存句法分析

  • 結合詞性標註到依存句法任務中，以降低彼此任務間的錯誤信息傳遞，並增強跨任務間的結構信息。
  • 在 Universal Dependency 2.0 的數據集上，我們提出的方法達到 81.14 LAS (多語言平均)，baseline 為 72.14 LAS，提出方法提升約 11%。
  • 對照實驗顯示詞性信息對於依存句法分析的預測有約 10%+的影響。
  • 標注系統由 Tensorflow + Python 實作，處理速度約每分鐘1K+。

維基百科數據處理及聯合知識文本表示學習

  • 解決跨語言方法對於平行語料的依賴，並嘗試結合知識實體到文本中以增強文本信息。
  • 設計弱監督算法生成維基百科跨語言語料 (平均每語言對 300K+ 段落)。
  • 算法基於負採樣 Skip-Gram 算法聯合學習知識實體、文本表示，並公開實驗訓練出的詞向量 (300d)。
  • 在 SemEval-2017 Track 4a 跨語言 (En-Es) 文本語義相似度任務中達到 44.99 Pearson-r，對比端到端模型 LASER 40.87 Pearson-r 提升近 10%。
  • 提出跨語言信息檢索任務，我們的方法達到 80% Top-10 Accuracy，對比 baseline 實驗 61% 有大幅提升 (30%)。

技能

程式語言

Python
C++
Java


框架及資料分析工具

Tensorflow
PyTorch
Pandas
Jupyter Lab
Dask
GraphQL

資料庫及搜尋引擎

MySQL
Elasticsearch
MongoDB

版控及CI/CD工具

Git
Bitbucket
Alembic
Jenkins

論文發表

  1. Hsuehkuan Lu, Yixin Cao, Lei Hou, and Juanzi Li. Knowledge-Enhanced Bilingual Textual Representations for Cross-Lingual Semantic Textual Similarity. International Conference of Pioneering Computer Scientists, Engineers and Educators (ICPCSEE), 2019. CCIS Volume 1058, pages 425-440.
  2. Hsuehkuan Lu, Lei Hou, and Juanzi Li. How Important Is POS to Dependency Parsing? Joint POS Tagging and Dependency Parsing Neural Networks. Chinese Computational Lingustics (CCL), 2019. LNCS, Volume 11856, pages 625-637.

End

Powered By
CakeResume