Software engineer at SwiftKey. Expert in natural language processing, machine learning and data compression. Skilled at C++, Python, and Java.
Working on natural langauge processing for text input method.
Developed a prototype of Japanese input method engine using C++ including kana kanji conversion, next word prediction and spelling correction. Developed high speed and memory-efficient dictionary compression algorithm, reducing memory consumption by 40%. Achieved 93% conversion accuracy (F score by character), outperforming Google Japanese input.
Developed N-gram language model trainer for 1 TB Japanese web corpus based on Hadoop MapReduce, combining word segmentation, pronunciation inference and spam filtering. Improved the speed of N-gram counting algorithm twice faster than a naive approach, achieved 5.65 bit cross entropy in the best case. Designed a phrase extraction technique for a predictive input method; achieved 0.90 precision and 0.81 recall, reduced size of dictionary by 80%.
Developed Japanese input method based on cloud computing. It enables users to effectively share dictionaries on servers. Achieved input time reduction by 21% and keystrokes saved by 26% with predictive input method. Achieved 25 million API accesses per month with over 290,000 unique users.
Developed various computer games for Windows. Developed entire part of Tobi-marisa, a curtain fire shooting game. The program consists of over 20,000 lines of C++ and sold to more than 1,000 customers and earned 1 million yen.
Keio University, Faculty of Science and Technology, Department of Information and Computer Science, Hagiwara Laboratory, Japan