Home Page of Wang Xuancong
Affiliation History:
[1] MOH Office for Healthcare Transformation (MOHT), Singapore
Previous:
[2] Human Language Technology, Institute for Infocomm Research (I2R), Singapore
[3] Ph.D candidate, Natural Language Processing Group, Department of Computer Science, School of Computing, National University of Singapore
[4] Ph.D candidate, NUS Graduate School for Integrative Sciences and Engineering
Email: xuancong84@gmail.com
QQ/Wechat: 476578019
Home Page: http://sites.google.com/site/xuancong84/
Office: Human Language Technology, Institute for Infocomm Research (I2R), Singapore
IT knowledge blog: https://xuancong84.blogspot.sg/
Github: https://github.com/xuancong84
Education
Ph.D in Computer Science, NUS Graduate School for Integrative Sciences and Engineering, 2008-2015
Bachelor of Science in Physic (major) and Mathematics (minor), National University of Singapore, 2004-2008
Hwa Chong Junior College, 2002-2004, Singapore
River Valley High School, 2000-2002, Singapore
Primary and Middle school affiliated to University of Science and Technology of China (USTC)
Bio
Wang Xuancong has obtained Ph.D in Computer Science at the National University of Singapore (2008-2015). He received a Bachelor of Science in Physics (1st class honor) from the National University of Singapore, in 2008. His research focuses on natural language processing, information retrieval and speech processing. In National University of Singapore (NUS), he has done some research of Automatic Speech Recognition (ASR), Machine Translation (MT), punctuation prediction and disfluency detection.
Currently, he is a senior data scientist in MOHT, working on data platform development for digital phenotyping, data analytics and healthcare machine learning models.
Research Interests
Natural Language Processing, Speech Processing, Machine Translation, Physical Sciences (Electro-magneto-gravitics)
Applications
0. The mathematical principle of traditional ethics. https://sites.google.com/site/xuancong84/arithmetic-ethics
0.5 Use smart-phone to do complex arithmetic, e.g., perform arbitrary-precision complex-number arithmetic, symbolic and numeric integration and differentiation, linear algebra, matrix arithmetics, etc. [link]
I was in charge of almost everything including GPU server selection and purchase, OS and deep learning software installation, text data extraction, pre-processing, post-processing, data purification, data relevance selection, deep neural network model construction/enhancement, model training, model adaptation, overall system integration, front-end back-end linkage and the web demo UI. The web demo was originally created by Lang Jun, it is enhanced by me to support asynchronous translation output display, line number display, line wrapping, synchronized scrolling, etc. However, lately, it was hidden from outside by the research institute due to the Cyber security act.
2. NUS Speech to Speech translation system (Link 1: NUS news PDF (too old, link taken down already)):
3. Augmented toolkit for Max-margin Markov Networks (M3N), which supports multiple layers of labels, i.e. Factorial M3N, https://github.com/xuancong84/m3n-ext
4. Solid directional rubix cube (super-super-cube). For a normal Rubix cube, we only solve the outermost layer. However, when the outermost layer is solved, the inner layers may not be solved if they are present. Moreover, for a normal cube, we only solve for colors, when the colors of each face are the same, the directions may not be consistent if they are distinguishable. Thus, in terms of the total number of configurations, a solid-directional Rubik's cube (super-super-cube) is more complex than a directional Rubik's cube (super-cube), which in turn is more complex than a normal Rubik's cube, as shown in the diagram below from right to left:
My Rubix Cube simulator together with a sample 15x15x15 solution (incomplete but sufficient to illustrate) can be downloaded freely at https://github.com/xuancong84/solid-rubix-cube
5. How to use auto-stereogram vision to spot the difference between two nearly identical images, e.g., spot the 3 distinct pixel pairs in the following image pair of 2D pixel arrays.
See the explanation at https://sites.google.com/site/xuancong84/spot-diff (bilingual), and download the source code, https://github.com/xuancong84/spotdiff
6. How can any primary school students mentally compute the product of two large integers each ranging from 0~9999999999, mul.jpg (the trick is that you do it digit by digit: starting from the unit digit, you mentally compute and memorize each digit, or write down each digit if you cannot memorize so many)
7. The standard 5-string score sheet (5-line staff/stave) can be modified slightly to improve the learning rate for piano beginners. Paired-triplet-line score sheet requires much less memory than the standard 5-string score sheet. It also has slightly larger pitch coverage and thus requires fewer ledger lines.
As shown in the figure above, every 3 lines in a group represent one octave, i.e., C D E F G A B. This modification make the score sheet conform to the periodic pattern in musical notes. Thus, the pianist only needs to remember 7 positions instead of all the 23+ positions (11 line positions and 12 intervening spaces). Moreover, ledger lines follow the same periodic pattern. All other features such as sharp and flat remain the same.
8. Nowadays, motion sensors using infra-red or microwave are very cheap and easily available on Lazada and Shopee. It is very easy to install motion sensors for pantry and wash-room lights. However, depending on the brand and cost, different sensors have different kinds of weakness and flaws, none are perfect.
9. I am a general-purpose disassembler, like to disassemble anything when needed
10. I have just started learning piano since 2013 and managed to play 6-7 songs. See my YouTube page. For some of my piano videos, I have used the music tempo visualization program created by myself in one class during my PhD 3rd year.
Music Tempo Visualization for Windows Media Player free download: https://github.com/xuancong84/TempoVis (binary is in the ./bin folder)
How to install:
1. copy TempoVis.dll into some fixed location (preferably your "\Program Files (x86)\Windows Media Player" folder)
2. Run command prompt with admin right and cd into that folder
3. Run "regsvr32 tempovis.dll". To uninstall, run "regsvr32 /u tempovis.dll"
4. Run your 32-bit Windows Media Player to activate this visualization
I have also created a Python version that can run in Linux (https://github.com/xuancong84/PyTempoVis )
11. I developed the world's best (up to 2022.5) open-source YouTube-based home Karaoke system by combining @vicwomg's Github repository (https://github.com/vicwomg/pikaraoke/ , Python 3 YouTube-based Karaoke player) and @tsurumeso's Github repository (https://github.com/tsurumeso/vocal-remover , DNN-based (deep-neural-network) vocal remover). It can search and download videos (with audio) from YouTube and many other video website, and run the PyTorch DNN model to split the vocal and instrument sound for the Karaoke player. It is fully open-source, at https://github.com/xuancong84/pikaraoke/ . Both the front-end and back-end support multiple languages (by Google Translate). And it supports Linux, Windows and Mac OS.
12. Out of my curiosity in electronics, I have also built an open-source smart-light sensor module for home ceiling lighting control. It uses 24GHz microwave micro-motion sensor to detect human presence and turns on the light when it is dark. I used EasyEDA to draw the PCB circuit for fabrication and wrote the Arduino source code at https://github.com/xuancong84/OpenSmartLight . Moreover, since in the mid-night, turning on the full light disturbs sleep, this device will turn on the onboard LED light smoothly in the mid-night.
Publications
Xuancong Wang, Nikola Vouk, Creighton Heaukulani, Thisum Buddhika, Wijaya Martanto, Jimmy Lee, Robert JT Morris, "HOPES -- An Integrative Digital Phenotyping Platform for Data Collection, Monitoring and Machine Learning", [doi: 10.2196/23984], Journal of Medical Internet Research 2021
Wenqiang Lei, Xuancong Wang, Meichun Liu, Ilija Ilievski, Xiangnan He, Min-Yen Kan, "SWIM: A Simple Word Interaction Model for Implicit Discourse Relation Recognition", IJCAI 2017
NIST Open Machine Translation 2015 evaluation (OpenMT15), I2R obtained 3rd position.
Xuancong Wang, Khe Chai Sim, Hwee Tou Ng, "Combining Punctuation and Disfluency Prediction: An Empirical Study", EMNLP 2014 (Oral)
Xuancong Wang, Hwee Tou Ng, Khe Chai Sim, "A Beam-Search Decoder for Disfluency Detection", COLING 2014 (Oral)
Xuancong Wang, Hwee Tou Ng, Khe Chai Sim, "Dynamic Conditional Random Fields for Joint Sentence Boundary and Punctuation Prediction", Interspeech 2012 (Oral)
Guangsen Wang, Bo Li, Shilin Liu, Xuancong Wang, Xiaoxuan Wang, Khe Chai Sim, "Improving Mandarin Predictive Text Input By Augmenting Pinyin Initials with Speech and Tonal Information", ICMI 2012 (Oral)