I want to develop
high-precision AI translation
specialized in a specific field
I want to improve the
accuracy
of a multilingual chatbot
I want to predict
utterances during conversations or Q&A
I am looking for
parallel corpora
that support
the languages of Southeast Asian countries.
I want to create
terminology dictionaries
specialized
for specific industries or fields of expertise.
Developing machine translation since the 1980s,
we have a proven track record of delivering parallel corpora to major companies.
With our many years of experience and know-how, we provide
multilingual corpora that contribute to solving your problems.
We possess parallel corpora in various fields such as
tourism, medical, legal, financial, and intellectual property.
We also provide parallel corpora in dialogue formats,
such as press conferences and Q&A sessions.
We can provide over 1 million pairs of parallel corpora!
They can also be used as
translation memory for the translation support tools you are currently using.
In addition to data translated from Japanese to English,
we possess parallel corpora
translated from English to Southeast Asian and rare languages.
Since the parallel corpora are translated by native speakers of each language,
the nuances and idiomatic expressions
specific to each language are reflected.
In addition to fields and data types, we manage the translation direction of the source and target texts,
so you can extract the parallel corpora you need.
We also accept requests for the creation of multilingual technical term dictionaries.
We provide safety precautions for working in factories and at work sites in a conversational format tailored to each industry.
We cover a wide range of fields, from primary industries such as agriculture, forestry, and fisheries, to reports on economics, finance, IT, and nuclear power.
We provide data in multiple languages necessary for foreigners to enter and live in Japan, such as laws regarding immigration and residence status, and medical questionnaires for clinical settings.
Research departments of major telecommunications companies
/broadcasting station research laboratories
/AI-based machine translation engine development companies
, etc.
The necessity of parallel corpora is increasing in the development and tuning of machine translation and multilingual generative AI. Furthermore, even when a translator starts working in a new field, having translation memory accumulated within the company can significantly improve translation efficiency. On the other hand, it is not easy to collect a systematic parallel corpus on your own, so it is efficient to procure data from external sources and focus on the development of the target deliverables or translation work.
There are six points to consider when selecting a parallel corpus as listed below.
1) Language combination
2) Field
3) Quality
4) Quantity
5) Data type
6) Presence or absence of context
The "language combination" in 1) refers to whether the language pair that the translator is translating or the language pair to be trained in machine translation development is Japanese and English, or Japanese and Chinese, etc. The same applies when registering in the translation memory of the CAT tool used; "language combination" can be said to be the most important element. Furthermore, if you are looking for more natural expressions, which language is the source text is also an important factor. As a concrete example, even if we just say "Japanese-English parallel corpus," it is natural that there will be differences in the fluency of English expressions and the background of the translation between cases where the Japanese source is translated into English and cases where the English source is translated into Japanese.
The "field" in 2) refers to areas such as tourism, medicine, law, economics, and science and technology. Machine learning with data from the field for which you wish to increase accuracy will lead to accelerated development. When building a language model to improve machine translation accuracy in a specific field, a benchmark for seeing results is considered to be around 100,000 pairs. It goes without saying that the field is also important when creating specialized terminology dictionaries from parallel corpora. In our company's experience, the fields in demand within Japan have shifted over the years: the tourism field was in demand during the era when we were focusing on inbound support in the name of becoming a tourism-oriented nation, and the medical field was in demand to strengthen medical care for foreign visitors and residents. Furthermore, needs have shifted to the business field, which includes presentation materials used for announcing and explaining business content, as well as lectures at meetings and events and the subsequent Q&A sessions.
The "quality" in 3), in other words, refers to accuracy and depends on how the parallel corpus was created. It is desirable for the data to have been translated by humans; data translated using machine translation that has not undergone human review or correction will naturally be of lower quality. Even in the case of human-translated parallel data, there are instances where one source sentence is translated into two or more target sentences, which affects quality if a one-to-one correspondence between source and target sentences is required. Furthermore, translations that are excessively abbreviated cannot be considered high-quality parallel corpora.
Regarding the "quantity" in 4), for example, if the purpose is to register them in the translation memory or terminology dictionary of a CAT tool used in-house, a parallel corpus of several tens of thousands of pairs in a specific field can be considered a sufficiently useful amount for a translator. When developing a machine translation engine, or performing machine learning for a specific field, it is said that 100,000 pairs have a certain effect. On the other hand, it is said that tens of millions of pairs are required when developing a general-purpose machine translation engine. As such, the amount of parallel corpus required varies greatly depending on the application.
The "data type" in 5) indicates whether the file used to create the parallel corpus is a report/white paper, a presentation/press conference, or a Q&A session. It is more effective to use a parallel corpus created from reports/white papers if you want to machine-learn written language, and from presentations/press conferences/Q&A sessions if you want to machine-learn spoken language. Since our company manages attributes in detail, it is possible to extract parallel corpora by data type.
The "presence or absence of context" in 6) refers to whether there is a semantic connection between multiple sentences. Specifically, example sentences listed in a dictionary only require the inclusion of a specific headword and have no connection to other example sentences. Therefore, it is judged as "no context." In contrast, a report consists of multiple sentences describing a certain event or occurrence, so it has "context." Similarly, for press conferences and Q&A sessions, since multiple speakers converse alternately, it can be said that there is "context." Not only for machine translation, but also for generating more accurate answers in chatbots, it has become more important than ever for the parallel corpus used for machine learning to have "context."
Summary
When selecting a parallel corpus, it is necessary to consider points 1) through 6) above depending on the intended use. In particular, when performing machine learning for a specific field, it is essential to consider the quality, data type, and the necessity of context. To utilize a parallel corpus that suits your purpose, please first check our free sample data.
