Our goal is publication quality output, not gist quality where only the basic meaning can be understood. With the Asia Online tools, this can be achieved.

Traditionally machine translation used a technique called Rule Based Machine Translation (RBMT). RBMT systems have the rules of a language coded by programmers. They have limited flexability and are difficult to customize and improve. Stylistic control is not easily possible and each language pair must be built independently of other language pairs.

Asia Online uses a technique called Statistical Machine Translation (SMT). SMT translation platforms use parrallel corpora (original text with a matching translation in the target language) to learn statistically which phrase or partial phrase is matches the other. The parallel corpora is analyzed and the statistical result of the analysis is used to determine new translations. With data available a new language pair can be built very quickly. This has enabled Asia Online to develop over 500 language pairs. In addition to the ease of developing a new language pair, with the right parallel corpora, SMT can give far higher translation quality than RBMT.

Asia Online also differentiates itself from others who have created SMT platforms by using only very clean parallel corpora. Others such as Google scrape the Internet and other content sources to get their corpora. Relying on public and uncontrolled sources for corpora means that there is a significant amount of dirty data in the parallel corpora. This results in lower quality. Asia Online's language pairs have far less parallel corpora than Google, but the data is clean, so when an error occurs, it is due to gaps in the data. Data can be submitted to fill the gaps when they present themselves. With dirty data, many thousands of examples would be needed to override the dirty statistics. With Asia Online between 5-20 examples usually fixes a problem and it will then not occur again.

Contact Asia Online for a free demonstration  and see the quality difference for yourself.

Asia Online provides a range of charging models to suit a variety of situations. The Asia Online Enterprise Translation Portal offers a solution for small-mid size organizations where no infrastructure is required and all translation is performed using the web portal. Using this model, Asia Online charges based on volume of translation and language pairs. There are a number of charging models to choose from.

Asia Online also licenses its software to large organizations that can support the infrastructure required to run a translation platform. Asia Online's translation platform is designed to scale to millions of words per day. The smallest implimentation requires 2 servers, with the ability to scale to thousands of servers. Asia Online provides a variety of onsite server licenses depending on customer requirements.

For each language pair Asia Online has developed a Baseline Engine. Baseline is the translation engine before it has been customized to focus on specific domains of knowledge. This is for general translation, but ideally you would add your custom data to tune the engine to your specific requirements. A baseline will typically translate general text such as news reasonably well, but would not have the same level of quality if the text was specific to a domain such as football or medical.

The topic of the text or writing style is called the Domain. The specific domain that you wish to translate shall be developed and layered on top of the baseline. Many domains are currently under development such as politics and IT, but others may need to be developed specifically to meet customer needs and to get the highest levels of quality possible.

The tuning for a specific domain of knowledge requires gathering additional data within that domain so that the knowledge in the engine is more focused and biased towards a particular subject. A domain may also be focused on the customers own preferred terminology.

Using the Asia Online Enterprise Translation Portal is the easiest means. This requires no special infrastructure at the client site. The Portal is accessed over the Internet through any common browser. Charges are based on a use model, or a monthly model depending on needs and give access to all available language pairs.

For large scale translation projects, the Asia Online translation environment can be installed at a customer’s site. This requires a minimum of 2 PCs to do translation. The performance will typically be many thousands of words per minute. Pricing for these licenses varies depending on the number of language pairs and domains. There are a number of charging models to choose from.

This question falls into two categories:

1. Data being translated - Using the Asia Online Enterprise Translation Portal, data is encrypted using SSL between client and the Asia Online servers. No data is kept by Asia Online after translation and all data is secure at all times. Alternatively the Asia Online translation environment can be installed at a customer’s site so the customers data never leaves its own work environment. Both options provide runtime security at differing levels based on customer needs.

2. Training Data – the data used to create an engine or domain. Training is very CPU intensive and occupies a number of high end servers for multiple weeks where the training data is analysed and statistics extracted. This is not required for most customers. In the training environment the number and power of servers determines how fast you can train. Typically an 8 CPU machine with 32GB RAM can take a number of days to train and tune and engine built from 10 million sentences. As this is very complex, we typically do this in house for our customers. When done in house, every measure is taken to ensure customer data is secure. The resulting engine or domain data is then used in the customer’s engine and no other company or users have access to the data or engine.

Enterprise Translation Portal

charging models

Asia Online also licenses its software to large organizations that can support the infrastructure required to run a translation platform. Asia Online's translation platform is designed to scale to millions of words per day. The smallest implimentation requires 2 servers, with the ability to scale to thousands of servers. Asia Online provides a variety of onsite server licenses depending on customer requirements.

In cases such as government or military where the training data may be highly sensitive, training can be done on the customer’s site with the purchase of special server licenses. Contact Asia Online for more information 

Quality is always subjective and there are many ways to measure it. There is seldom agreement on one absolute measure. The most accurate quality measurements will typically be produced when humans are given a clear set of measurement guidelines. However, even this can be inaccurate at times. Consider the following: "It rained on Monday" vs. "On Monday it rained", or "The speedo was maxxed into the red line" vs. "The speedometer needle was hard against the red line". Individual preferences for linguistic style would score these variations differently.

Using humans to measure quality is usually slow and expensive. There are also many different human assessment metrics, such as scoring each sentence on a defined quality scale of 1 to 5 or similar approaches. But what is usually more important than measuring some absolute quality is understanding quality in the SMT translation engine process. It is useful to measure the improvement in quality over time, as data configurations are varied or measuring the difference between automated translations using different translation platforms and tools.

To facilitate the rapid evolution and evaluation of potential translation engines, some automated quality measurement tools have been developed. There are a number of different metrics. BLEU, F-Measure and Translation Error Rate (TER) are all useful indicators of quality. Asia Online has developed a set of free downloadable tools and utilities to help you to measure quality of automated translation engines. Each of the metrics uses a simple basic concept - have an original document translated by a human as a "gold standard" or reference for comparison. Then translate the same original document using the software and compare the automated translation output to the human reference. The Asia Online tools and the BLEU, F-Measure and TER measurements can be used to compare multiple different sources of translated output, which can help in understanding the best translation platform or software for your documents.

To build a custom domain, typically the more high quality data you have the higher the quality of translated output. Traditionally this meant you needed 400,000 or more multilingul segments. Because of the extensive foundation work that Asia Online has performed on our baseline engines and our data optimization and generation capabilities, we have been able to build high quality custom domains on as little as 20,000 multilingual segments. We have also developed a number of custom domains when there was no data available from customers by applying our data generation technologies.

Contact Asia Online  for a more detailed discussion on your custom domain requirments, custom domain case studies and a demonstration of some of the existing custom domains.

Asia Online is always interested in talking to content partners about creative ways to monetize their content in a variety of Asian markets.

Contact Asia Online  to discuss your unique content and requirements with an Asia Online representative.

Asia Online encourages growth and education. We work with a number of educational and non-profit organizations. Contact Asia Online  to discuss your requirements with an Asia Online representative. In most cases, Asia Online will be happy to contribute and help.

The latest jobs are always listed on our careers  page.

Asia Online is currently funded by its founders, angels and series A and B institutional investors Japan Asia Investment Corporation (JAIC) and T-One Capital. For further investment information, please contact Asia Online  and an Asia Online representative will discuss your interest in more detail.