In modern Japanese, a vast array of loanwords are employed in everyday life. These words are commonly written in katakana, one of the three primary scripts used to write Japanese. Although some words such as コーヒー kōhī, and パン pan are from sources other than English (Dutch koffie and Portuguese pão, respectively), the majority of loanwords in use today are derived from English. In addition to calques from English, there are also terms and phrases that have been coined by the Japanese using English roots; these are known as 和製英語 wasei eigo ("English made in Japan") and typically do not translate directly: e.g., メールマガジン mēru magajin ("mail magazine") for an e-mail newsletter.
This repository contains primarily some Python (>= 3.6) scripts for scraping various resources in order to create a set of (primarily English) loanwords in Japanese and some SQL scripts to merge the data. At the moment, there are 221,588 English-Japanese vocabulary pairs.
This serves two main purposes:
- As a resource for learners and researchers of Japanese/English who wish to study the phonetic changes that occur for (English) calques in Japanese and as a reference for vocabulary study.
- As a resource for data science and machine learning practitioners who may wish to improve their systems (e.g., by allowing speech synthesis technologies to handle English text embedded in Japanese sentences).
On a personal note, the code here was developed primarily with the second purpose in mind and is being released in the interest of goodwill towards the community of researchers, practitioners and individuals who have made it possible.
| English | Japanese |
|---|---|
| NANOTREE | ナノツリー |
| MICHELANGELOS | マイカランジャローズ |
| LARAMIE | ララミー |
| THRUN | スラン |
| EMBRATEL | エンブラテル |
| CALMNESS | カムナス |
| PRESTI | プレスティー |
| EUBOEAN | ユービーアン |
| SYMMS | シムズ |
| LUBA | ルバ |
| FINESSE | フィネス |
| OUDONG | ウドン |
| YINGST | イングスト |
| COSMOPOLITE | コズマパライト |
| SIGN-OFF | サインオフ |
| HOLE | ホール |
| PRESENILIN | プレセニリン |
| COMPANYWIDE | コンパニーワイド |
| GRANINGEVERKEN | グラニンゲバーカン |
| MORALE | モラール |
| FLUBBING | フラビング |
| DILLMAN | ディルマン |
| KARNICKI | カーニキー |
| MILKENS | ミルカンズ |
| WORKER | ワーカー |
This project makes use of the following resources.
- Dumps from Wikipedia (EN, JA)
- Dumps from Wiktionary (EN, JA)
- The Japanese-Multilingual dictionary (JMdict)
- The CMU pronouncing dictionary (CMUdict)
- Jose Llarena's Britfone repository.
- The data created for the research paper Richardson J, Nakazawa T, Kurohashi S. Bilingual Dictionary Construction with Transliteration Filtering. In: Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC'14) (2014)
- The Japan Technical Communicators Association's (JTCA) publication 外来語（カタカナ）表記ガイドライン 第3版 (2015)
The resources above can be divided into two types: those that provide an English-katakana mapping and those that provide an English-phonetic mapping.
Resources of the first type (Wikipedia, JMdict, LREC'14, JTCA) remain largely unmodified. Some basic tricks are employed to align source and target words, but katakana are otherwise unchanged.
Resources of the second type (Wiktionary, CMUdict, Britfone) are heavily processed in order to systematically transform phonetic and surface forms into more or less natural katakana.
The simplest way is to import the final data from ./loanwords_gairaigo/db/merged.sql into a database.
If you wish to obtain the individual processed data before they are merged, you can import the data in the other SQL files britfone.sql, cmudict.sql, wiktionary.sql, lrec2014.sql, jtca.sql, jmdict.sql, wikipedia.sql. To generate the type one and type two data, use create_type_1.sql and create_type_2.sql.
Alternatively, you can recreate the data from scratch by downloading the resources as explained in ./loanwords_gairaigo/data/download_instructions, processing them in the same order as in ./loanwords_gairaigo/python/process_all.sh which will create some SQLite 3 databases in ./loanwords_gairaigo/db/, which are then merged by ./loanwords_gairaigo/db/create_type_1.sql, ./loanwords_gairaigo/db/create_type_2.sql and finally ./loanwords_gairaigo/db/merge_clean_db.py.
Care should be taken for the data from JMdict; data were initially extracted mechanically and then judged by three human reviewers to determine if they were acceptable loanwords (e.g., many mimetic words (擬音・擬態語 gion/gitai-go) made it through the initial pass). The words that were eligible for the final merge are inidicated with a value of 1 in the column ok of table gairaigo_combined in jmdict.sql. The judging criteria are available in ./loanwords_gairaigo/docs/外来語を判断する.pdf (Japanese only).
The situation is similar for JTCA, where the table from the PDF was copy-pasted and reviewed for errors. As such, there exists only the SQL file jtca.sql.
More information on how the data is processed is below. If you decide to process the data from scratch, you will require MySQL for the langlinks Wikipedia file, Python 3.6 or greater and the 3rd party Python libraries lxml and MySQL Connector and Unidecode which can all be easily pip-installed with pip3 install lxml mysql-connector-python Unidecode.
One may wish to imagine that English words are transcribed perfectly phonetically into katakana. Alas, the unfortunate truth is that this is often not the case. Phenomena such as the introduction of loanwords when Japanese and/or English phonology was different, misreadings or misinterpretations of the English source words or precedence given to the written form of the English word over its pronunciation have led to cases such as モンキー monkī for English "monkey", which would be phonetically transcribed more accurately as マンキー mankī.
Therefore, we begin by transcribing phonetically and then correct these transcriptions based on some common graphemes. For example, English words starting with MO that have been transcribed as マ are corrected to モ (except for words beginning in MOTHER).
Another complicating factor is the presence of multiple transcription systems and therefore transcriptions for a single English term. Some systems employ ウィ and ウェ to transcribe wi and we, while others may prefer ウイ or ウエ. Similarly, newer transcriptions allows ヴ vu to represent the voiced labiodental fricative v which does not exist in Japanese (it is still pronounced like a b) or combinations like クィ kwi or スィ swi for words like "queen" or "sweets".
We let the data speak for themselves and accept mappings based on the reliability of the resource.
For resources of type one, we choose the ordering JTCA > LREC'14 > JMdict > Wikipedia. Firstly, both Wikipedia and JMdict are open source, however JMdict with its etymology and wasei tags allow us to make better judgements. Furthermore, while Wikipedia gives us a large resource it is relatively low-quality, suffering from inconsistent transcriptions and nicknames, etc. The relative ordering of LREC'14 and JTCA is debatable: LREC'14 provides many good scientific and technical vocabulary but is only backed by three authors and many English-katakana pairs collapse to a single item after normalization, while JTCA provides far fewer vocabulary items but is endorsed by an industry association.
For resources of type two, we choose the ordering Britfone > CMUdict > Wiktionary. Wiktionary's transcriptions were so volatile and low-quality that, despite the large number of terms it contributed, it may be worth dropping the resource altogether. Both Britfone and CMUdict provide high-quality, canonical transcriptions for British and American English (received pronunciation (RP) and general American (GA), respectively). However, RP is typically the model for the import of loanwords and so transcriptions from RP tend to be better. In more recent years, however, loanwords from GA have seen a rise, but this is unlikely to overtake the large bulk of RP-based terms that already exist in Japanese in the near future.
After extracting transcriptions from resources of both types, we then merge them to create the final dataset. In this final merge, we prefer type one to type two. The full ordering is then JTCA > LREC'14 > JMdict > Wikipedia > Britfone > CMUdict > Wiktionary.
Aside from this, there remain English homonyms like "wind", which could be ウィンド windo or ワインド waindo. We do not deal with this and simply list whichever transcription appears first in the ordering.
GNU GENERAL PUBLIC LICENSE Version 3
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS “AS IS” AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.