Geza Kovacs
Geza Kovacs Research Open-Source Teaching Contact Publications CV / Resume

Geza Kovacs

geza

Geza Kovacs

I'm Geza Kovacs, a Senior Research Scientist at Google, working on applications of LLMs. I was previously at Lilt, working on improving translators' productivity using interactive machine translation.
I did my PhD in Computer Science at Stanford, where I was advised by Michael Bernstein in the Human-Computer Interaction group, and undergrad and masters at MIT. See my resume for details.

Research

Automatic Correction of Human Translations Lilt
We develop Transformer-based models for performing automatic correction of translation errors made by human translators, and show that in a human-in-the-loop setting our system helps translators produce higher quality translations.

Jessy Lin, Geza Kovacs, Aditya Shastry, Joern Wuebker, John DeNero. Automatic Correction of Human Translations. NAACL 2022 (best paper award).


Effects of Translation on Website Engagement Lilt
In this 3 million user study we ran on UNetbootin and HabitLab, we found that human and machine translations improve website engagement, and in-browser machine translation is often used when English is shown.

Geza Kovacs, John DeNero. Measuring the Effects of Human and Machine Translation on Website Engagement. AMTA 2022 (best presentation award).


Interactive Machine Translation Use on Lilt Lilt
Lilt is an interactive Machine Translation platform used by thousands of translators. This study investigates how translators use interactive MT, quality and speed of MT suggestions, and how translators spend their time on Lilt.

Geza Kovacs. Interactive Machine Translation Use on Lilt. AMTA 2020 Workshop on the Impact of Machine Translation.

Samuel Läubli, Patrick Simianer, Joern Wuebker, Geza Kovacs, Rico Sennrich, Spence Green. The Impact of Text Presentation on Translator Performance. Target: International Journal of Translation Studies, 2021.


Adapting to Changing Difficulty Preferences HabitLab
In this 8,372 user study we ran on HabitLab, we found users choose easier interventions over time, but expect to return to harder interventions. Adapting to users' difficulty preferences via low-frequency prompting improves retention.

Geza Kovacs, Zhengxuan Wu, Michael Bernstein. Not Now, Ask Later: Users Weaken Their Behavior Change Regimen Over Time, But Expect To Re-Strengthen It Imminently. CHI 2021.


HabitLab: Online Behavior Change Experiments HabitLab
HabitLab is an open-source platform that helps users reduce their time online by conducting personalized behavior change experiments using 30+ interventions. It is available for Chrome and Android and has over 12,000 active users.

Geza Kovacs. HabitLab: In-the-Wild Behavior Change Experiments at Scale. Stanford University Department of Computer Science, PhD Thesis, 2019.


Conservation of Procrastination HabitLab
Do productivity interventions save time, or just redistribute it elsewhere? In this 5,230 user study we ran on our HabitLab platform, we found that interventions help save time not just on targetted sites, but on other sites as well.

Geza Kovacs, Drew Mylander Gregory, Zilin Ma, Zhengxuan Wu, Golrokh Emami, Jacob Ray, Michael Bernstein. Conservation of Procrastination: Do Productivity Interventions Save Time or Just Redistribute It? CHI 2019.


Effects of Rotating Behavior Change Interventions HabitLab
In this 1,654 user study we ran on our HabitLab platform, we found that rotating interventions improves effectiveness at the cost of increased attrition. We reduced this attrition by half with a dialog that explains the rotation system.

Geza Kovacs, Zhengxuan Wu, Michael Bernstein. Rotating Online Behavior Change Interventions Increases Effectiveness But Also Increases Attrition. CSCW 2018.


EduFeed: An Educational Social Feed for Children EduFeed
EduFeed is an educational social news feed for preliterate children.
We used EduFeed to teach reading and arithmetic skills to children in India.

Kiley Sobel, Geza Kovacs, Galen McQuillen, Andrew Cross, Nirupama Chandrasekaran, Nathalie Riche, Ed Cutrell, Meredith Morris. EduFeed: A Social Feed to Engage Preliterate Children in Educational Activities. CSCW 2017.


Crowd Research: Open and Scalable Laboratories Crowd Research
Crowd Research is a project through which we coordinated hundreds of volunteers to build and deploy Daemo, a crowdsourcing platform that implements novel solutions to trust and fairness issues in crowdsourcing.

Rajan Vaish, Neil Gaikwad, Geza Kovacs, Andreas Veit, Ranjay Krishna, Imanol Arrieta Ibarra, Camelia Simoiu, Michael Wilber, Serge Belongie, Sharad Goel, James Davis, Michael Bernstein. Crowd Research: Open and Scalable University Laboratories. UIST 2017.


Effects of In-Video Quizzes on MOOC Video Viewing InVideo
A data-mining analysis of how users interact with in-video quizzes on Coursera.
We discover that the presence of in-video quizzes results in quiz-driven navigation patterns such as seeking from one in-video quiz to the next.

Geza Kovacs. Effects of In-Video Quizzes on MOOC Lecture Viewing. Learning at Scale 2016.


FeedLearn: Microlearning in Facebook Feeds FeedLearn
FeedLearn helps users learn vocabulary while browsing Facebook by inserting interactive quizzes into the news feed. Users complete 4x more exercises with FeedLearn, compared to injecting links into the feed or sending daily emails.

Geza Kovacs. FeedLearn: Using Facebook Feeds for Microlearning. CHI 2015 Works in Progress.


QuizCram: Question-Driven Video Viewing QuizCram
QuizCram is a MOOC lecture viewer that uses quizzes to direct navigation and reviewing. It automatically generates materials from existing Coursera lectures.

Geza Kovacs. QuizCram: A Question-Driven Video Studying Interface. CHI 2015 Student Research Competition.


Smart Subtitles for Foreign Language Learning Smart Subtitles
Smart Subtitles helps learners learn vocabulary while watching foreign-language videos. It features interactive transcripts with hover-over definitions and dialog-based navigation. Users learn 3x more vocabulary with Smart Subtitles than with dual-language subtitles.

Geza Kovacs, Robert C Miller. Smart Subtitles for Vocabulary Learning. CHI 2014.


Visualizing Grammar to Learn Foreign Languages GrammarVis
GrammarVis helps foreign language learners interactively explore the syntactic structure of sentences to better understand their word order and grammar.
I built it as part of a foreign-language manga reader application.

Geza Kovacs, Robert C Miller. Foreign Manga Reader: Learn Grammar and Pronunciation while Reading Comics. UIST 2013 Demo.

Geza Kovacs. Multimedia for Language Learning. MEng Thesis, MIT, 2013.


ScreenMatch: Visual Context for Translators ScreenMatch
ScreenMatch provides software translators with screenshots illustrating the contexts in which the strings they are translating are used in the app.
It collects screenshots while users run the app, and matches them via OCR.

Geza Kovacs. ScreenMatch: Providing Context to Software Translators by Displaying Screenshots. CHI 2012 Student Research Competition (1st place).


Open-Source

UNetbootin: LiveUSB Creator UNetbootin
UNetbootin is a graphical tool to create bootable live USB drives.
It allows users to install Ubuntu and other Linux distributions on tablets and netbooks without CD drives.

I built and maintain UNetbootin. It has been downloaded over 40 million times.

Wubi: Ubuntu Installer for Windows Wubi
Wubi is a tool that allows users to install Ubuntu from Windows.
It installs and boots Ubuntu from a file, so it does not require any repartitioning.

I built the original versions of Wubi, which is now part of Ubuntu.

Chinese/Cantonese Dictionary for Firefox/Chrome Cantonese and Chinese Dictionaries for Chrome and Firefox
Extension for helping you learn Chinese while reading webpages.
Hover over a word, and a definition and pronunciation will be displayed.
Shows both Standard Chinese (Mandarin) and Cantonese pronunciations.

Chinese and Cantonese Dictionaries for Kindle Kindle Cantonse and Chinese Dictionaries
Dictionaries for helping you learn Chinese while reading on your Kindle.
Highlight a word, and a definition and pronunciation will be displayed.
Shows both Standard Chinese (Mandarin) and Cantonese pronunciations.

Input Methods for Learning Chinese/Cantonese Cantonse and Chinese Input Methods
Input methods that help Chinese speakers learn Cantonese and vice versa.
Can type words using pronunciations in either language, and shows pronunciations in the other language. Based on RIME.

2-Key Cantonese Input Method 雙粵拼輸入法 Cantonse and Chinese Input Methods
A phonetic input method for Cantonese that allows you to unambiguously type a syllable with 2 keystrokes, or 3 if including tones. Designed to be easy to learn by users of Double Pinyin (雙拼) for Mandarin. Based on RIME.

My other open-source projects are on Github, Npm, PyPI, and Launchpad.


Teaching

Understanding Users (CS 377U) - Stanford - TA, Spring 2019

Human-Computer Interaction Research (CS 376) - Stanford - TA, Fall 2018

Natural Language Processing (6.863) - MIT - TA, Fall 2012

Introduction to C++ (6.096) - MIT - Instructor, MIT, IAP 2011

Introduction to C# and the .NET Framework - MIT - Instructor, IAP 2011

Maslab Autonomous Robotics Competition (6.186) - MIT - Software Director, IAP 2011


Contact

For personal correspondence, I prefer email:

my first name at cs stanford

If asking about UNetbootin, please report an issue on Github, and consider making a donation.


FAQ / Misc

Can I have the source code for one of your projects?

Check my GitHub account.

What programming languages and libraries do you use and recommend?

I used to contribute to the LiveScript language but it's not very well maintained anymore, so mostly use Python and JavaScript these days.

For data science I prefer using Python in a Jupyter notebook, using rpy2 to invoke R libraries.
For machine learning I prefer PyTorch.
For backend development I prefer Node.js with Koa if writing in JavaScript, or Flask if writing in Python.
For frontend development I am experienced with React, Polymer, and AngularJS.
For mobile development I've had good experiences with NativeScript and React Native.

What natural languages do you know?

I'm fluent in English and Standard Chinese (Mandarin). I also know Hungarian by virtue of having spent my childhood in Hungary, and Vietnamese from speaking it with my family. I have also studied Spanish and Japanese, and can still read and understand them but haven't spoken them in a while.

How do I pronounce your name?

Kovács Géza (IPA: kovaːt͡ʃ ɡeːzɒ) The "e" sound in Geza is the same as in "hey". The "cs" in Kovacs is pronounced "ch", not "ks". Hungarian puts the surname first.

I also use my Vietnamese name, Minh Trung (漢字: 明忠) (IPA: miŋ˧˧ ṯɕʊwŋ˧˧)