Unlocking the potential of machine translation

Harnessing the power of machine translation (MT) is now a priority for many functions in many organizations around the world. For product and marketing departments, MT can accelerate entry into overseas markets, reach more international users without a linear increase in costs, and help accelerate the return on translation investments.

In terms of translation and localization functions, MT as a technology is rapidly evolving, and we expect productivity gains that will allow us to do more with fewer resources.

Our vision is to use MT solutions to solve your next business challenges:

  • Translate content that is not being translated due to lack of funds or time
  • Reduce time to market
  • Achieve higher productivity
  • Reduce translation costs
  • Identify best practices and apply MT as needed

Untapped potential

Publishers are turning to machine translation to achieve their goal of delivering high-quality content to customers around the world. Traditional, purely manual translation methods, even when used in conjunction with translation memory, often fail to achieve their goals. They are too expensive and time-consuming. If your goal is to effectively support customers in their local language, machine translation may be the answer. For many large, global companies, MT offers a very attractive return on investment.

ABC Translations offers its customers two main machine translation solutions that cover a wide range of potential manufacturing applications for MT today: an MT post-processing package and a comprehensive MT implementation solution designed with a flexible partner strategy in mind.

ABC Translations' MT partner strategy

ABC Translations has a neutral position among the emerging MT engine suppliers, which allows us to choose and recommend the most suitable engines according to our customers' requirements. We are in contact with several service providers. However, our preference is for statistical machine translation (SMT).

While many rule-based machine translation (RBMT) systems have a long history and excellent performance, we have made the strategic technological decision that SMT promises the fastest improvement in translation quality, lower post-editing costs and, in turn, lower costs for our clients. The engine can then be trained for a specific domain (subject area) or even a specific customer or product area. For this reason, we work primarily with SMT service providers.

MT post-processing.

ABC Translations has the most advanced post-processing solutions in the industry. Our entire global partner network is tailored to your post-processing needs. We've developed courses, guides and processes to help you get started. Our post-processing solutions can be applied to results we've created from our trained and managed engines, or to results from your existing in-house MT solution.

MT Implementation Solutions

Our MT implementation solutions typically start with a pilot project. This approach is based on the current maturity of MT and the various scenarios that define a successful implementation. Therefore, a typical first step in assessing the potential of MT for a particular customer is to conduct a pilot project.

The purpose of the pilot project is to assess how MT can be used in real production, what expectations can be met in terms of quality, cost and productivity, and what the best production processes are.

The following key factors determine the success of MT implementation:

  • Is MT expected to improve productivity, reduce time to market, or localize untranslated content? The anticipated use of MT will determine the specific types of services that ABC Translations will provide and the content that the pilot will cover. Of primary importance will be the quality of the end result and whether and to what extent MT will be post-edited (so-called "light" post-editing or "full" post-editing).
  • Domains and content types. Some content types are particularly suitable for machine translation, such as customer service content and user documentation. In some cases, the use of MT for software user interfaces is becoming more common. Marketing communications materials and other more creative content are generally not suitable for MT unless MT techniques are used to achieve a "raw" result that reflects the overall intent.
  • Language pairs. Some target languages or language combinations are more productive, while others, especially Asian languages, are less suitable for MT.
  • Quality of source language. A better understanding of the source language and the use of appropriate authoring tools usually greatly increases MT's potential. This includes the use of established grammar and style rules as well as terminology.
  • Legacy content and availability of client and domains specific corpora. For statistical machine translation methods recommended by ABC Translations, the more high-quality legacy content available to train the selected machine translation engine, the better for you from the start.
  • The amount of bilingual content (translation memory) required for production-level adaptation can be as high as 2 million words for European languages and 10 million words for Asian languages, but ABC Translations can reduce these requirements through its ability to clean up existing corpora.
  • Providing tailored custom dictionaries. Similar to providing legacy content and corpora, well-prepared and maintained dictionaries can significantly improve the quality of MT results and increase the cost-effectiveness of MT use.
  • Quality of translations. In addition to statistical training of the MT engine with bilingual content, additional content in the target language is added to train the MT engine and put the finishing touches on the translation.

Expectations and Process

To assess whether machine translation can help you achieve your business goals, we'll go through the following typical process with you

  • Customized MT Engine Configuration. ABC Translations selects partners based on knowledge of their strengths and trains them to work with the specific dataset the client has. The client provides bilingual content to customize the MT engine, and ABC Translations analyzes and prepares the TM/language content for the initial configuration of the engine.
  • Select sample content for the MT
  • Agree on how to measure quality and performance.
  • Determine the quality of the raw MT output.
  • Identify recurring problems in the MT output
  • Calibrate the motor if necessary.
  • Perform rework and determine the type of rework and achievable performance

Finally, we provide the customer with a report on the overall quality of the custom engine and a detailed plan for the actual MT production rollout.

Human factors

Human factors are an important part of quality control. A certain percentage of carcasses must be pre-inspected by humans. The higher this percentage and the more effort put into inspection and cleaning, the higher the quality of the final MT product.

Data cleaning is an important but often overlooked aspect. The data in existing and selected MT databases is not always in perfect condition. It needs to be normalized, standardized, and cleaned to achieve optimal results. This requires an effective combination of linguistic and data processing capabilities that ABC Translations can provide.

Measuring translation quality with MT

Custom MT engine quality is measured using a tool that compares MT engine performance for examples not included in training with human translation of the same content. This content is referred to as the "test set," and the human translation of the test set is often referred to as the "reference translation."

To ensure objectivity in measuring the quality of the relevant MT engine, the test suite content must be separated from the data used for the relevant engine. If the test suite data is not separated out and used for engine adaptation, the system will use the same (or a very similar) translation for the test suite, which will affect the results (for such a test suite, the quality can be much better).

To evaluate the quality of the translation resulting from the MT system, we use three metrics

  • NSS
  • BLEU

The results measure the similarity of MT results to a reference translation and are expressed as a percentage.

All metrics compare human translation (the reference set) with a set of machine-translated fragments (the test set). It is important to select the right test suite for each machine translation engine, according to the following rules

  • The test suite should contain a representative set of data to be translated in production.
  • The data included in the test suite should be representative of the data used in the training corpus (training set) of the MT engine.
  • Ideally, the test set should contain complete sentences, not just one or two words.
  • The test set and the reference set should match exactly.

In order to properly assess the suitability of machine translation for a particular purpose, it is important to prepare and manage expectations in advance. With these considerations in mind, we hope to make good use of the potential of machine translation.

More To Explore

Do You Want To Boost Your Business?

drop us a line and keep in touch

ABC Translations