User:Eiji

Contact Information

   Name:Eiji Miyamoto
   E-mail address:motopon57@gmail.com
   University: University of Manchester
   IRC:thelounge72
   github:https://github.com/yypy22
   Language: Japanese, English

Why is it that you are interested in Apertium?

I am intrigued by natural language processing and its usage. NLP is widely used and it improves human and machine communication better. We can see many applications of NLP in real life. Apertium is open-source and free software for machine translation, so apertium matches my interest. The community here is very welcoming and supportive. I have always wanted to contribute to open-source projects, and I believe apertium is the best open-source software and community to work with.

Which of the published tasks are you interested in?

Tokenization for spaceless orthographies in Japanese https://wiki.apertium.org/wiki/Task_ideas_for_Google_Code-in/Tokenisation_for_spaceless_orthographies

What do you plan to do?

Investigating the suitable tokenizer for east/south Asian languages which usually do not use spaces and implementing it. Besides, improving Japanese-related files.

Reasons why Google and Apertium should sponsor it

Apertium translates European languages into other European languages mainly and my proposal for Google Summer of Code 2023 will open up the possibility for future translation in Asian languages which usually do not have space between words in sentences. The Japanese repository in apertium has not been updated for over two years and I believe my task for GSoC2023 will make this repository more active. I also believe my proposal will expand the possibility of creating Asian Language and Other language pairs in the future.

Work plan

Time period	Weekly Goal	Details
Week 1 May 29 - June 4	Investigating word segmentation and tokenization from paper, and summarising useful findings into a report	Looking up NLP conference papers on word segmentation or tokenization, and checking popular Japanese tokenizers such as MeCab and Juman.
Week 2: June 5 - June 11	Testing possible algorithms for tokenization and becoming aware of the pros and cons of them	N-gram, Maximal matching, Viterbi, SentensePiece, etc... Discussing the algorithm with mentors.
Week 3: June 12 - June 18	Analyzing drawbacks of current tokenization for apertium-iii and apertium-jpn	Drawing up a report and thinking of what the hybrid model needs to cover. Making a pseudocode for the model.
Week 4: June 19 - June 25	Producing a hybrid model	Trying to minimize the drawbacks of the current tokenizer and make the model more efficiently. Coding the model according to the pseudocode in python.
Week 5: June 26 - July 2	Analyzing the model and improving it	Checking efficiency of memory usage and speed of the model. If it has room for improvement, improve it. Testing speed and accuracy with texts, comparing manually tokenized sentences and the one tokenized with the hybrid model.
Week 6: July 3 - July 9	Converting the model into a faster language	Python model to C++ or faster language.
Mid-term Evaluation
Week 7: July 10 - July 16	Continue to work on converting	Fixing bugs and checking logic and function. Testing the model with simple sentences.
Week 8: July 17 - July 23	Converting the hybrid model into apertium-jpn	Sending a pull request and refactoring the code.
Week 9: July 24 - July 30	Improving twol file and lexicon file	Adding more kanji, hiragana, and katakana in it and organizing the file for better use. Including roman numeric characters. Evaluation of coverage improvement.
Week 10: July 31 - Aug 6	Improving disambiguation for Japanese	Adding grammar to lexical selection, determiner, and remove reading noun. Evaluation of Coverage improvement.
Week 11: Aug7 - Aug13	Testing and fixing bugs	Testing accuracy and speed with texts and refactoring the code.
Week 12: Aug 14 - August 20	Finalise GSOC project	Writing reports and completing tests.
Final Evaluation

Coding Challenge

Integration apertium-iii transducer into apertium-jpn. https://github.com/yypy22/apertium-jpn I modified MakeFile.am and modes.xml to use tokenized.py from apertium-iii. I added tokenizer-related code such as $(LANG1).autotok.hfst into makefile and program tag in modes.xml, dropping tokenized.py file into jpn repository.

Skills

Python, Java, C++(intermediate), JavaScript(intermediate), HTML, CSS, Docker, Django, XML, Git, jQuery I am currently a second-year student at the University of Manchester and have taken algorithm, data science, AI, programming, and software engineering lectures. I also have taken summer school at Tokyo University on NLP online and got a solo silver medal(top 4%) from the Kaggle competition. I worked as an intern last summer for a data feed system and used python, js, docker, git, HTML, CSS, and SQL mainly. I believe my basic NLP knowledge of Japanese and my experience equip me to work on this project successfully.

Other Summer Plans

I do not have any other plan and can work on the project 40hrs/week or more. My summer vacation will start from 29th May to 18th September.