Many of us, involved with Machine Translation are familiar with the importance of using high quality parallel data to build and customize good quality MT engines. Building high quality MT engines with sparse data is a challenge faced not only by Language Service Providers (LSPs), but any company with limited bilingual resources. A more economical alternative to creating large quantities of high quality bilingual data can be found by adding monolingual data in the target language to an MT engine.
Statistical Machine Translation systems use algorithms to find the most probable translations, based on how often patterns occur in the training data, so it makes sense to use large volumes of bilingual training data. The best data to use for training MT engines is usually high quality bilingual data and glossaries, so it’s great if you have access to these language assets.
But what happens when access to high quality parallel data is limited?
Bilingual data is costly and time-consuming to produce in large volumes, so the smart option is to come up with more economical language assets, and monolingual data is one of those economical assets. MT output fluency improves dramatically, by using monolingual data to train an engine, especially in cases where good quality bilingual data is a sparse language resource.
More economical…
Many companies lack the necessary resources to develop their own high quality in domain parallel data. But, monolingual data – is readily available in large volumes across different domains. This target language content can be found anywhere; websites, blogs, customers and even company specific documents created for internal use.
Companies with sparse parallel data can really leverage their available language assets with monolingual data to produce better quality engines, producing more fluent output. Even those with access to large volumes of bilingual data can still take advantage of using monolingual data to improve target language fluency.
Target language monolingual data is introduced during the engine training process so the engine learns how to generate fluent output. The positive effects of including monolingual data in the training process have been proven both academically and commercially. In a study for TAUS, Natalia Korchagina confirmed that using monolingual data when training SMT engines considerably improved the BLEU score for a Russian-French translation system.
Natalia’s study not only “proved the rule” that in domain monolingual data improves engine quality, she also identified that out of domain monolingual data also improves quality, but to a lesser extent.
Monolingual data can be particularly useful for improving scores in morphologically rich languages like; Czech, Finnish, German and Slovak, as these languages are often syntactically more complicated for Machine Translation.
Success with Monolingual Data…
KantanMT has had considerable success with its clients using monolingual data to improve their engines quality. An engine trained with sparse bilingual data (the sparse bilingual data was still greater than the amount of data in Korchagina’s study) in the financial domain showed a significant improvement in the engine’s overall quality metrics when financial monolingual data was added to the engine:
- BLEU score showed approx. 40% improvement
- F-Measure score showed approx. 12% improvement
- TER (Total Error Rate), where a lower score is better saw a reduction of approx. 50%
The support team at KantanMT showed the client how to use monolingual data to their advantage, getting the most out of their engine, and empowering the client to improve and control the accuracy and fluency of their engines.
How will this Benefit LSPs…
Online shopping by users of what can be considered ‘lower density languages’ or languages with limited bilingual resources is driving demand for multilingual website localization. Online shoppers prefer to make purchases in their own language, and more people are going online to shop as global internet capabilities improve. Companies with an online presence and limited language resources are turning to LSPs to produce this multilingual content.
Most LSPs with access to vast amounts of high quality parallel data can still take advantage of monolingual data to help improve target language fluency. But LSPs building and training MT engines for uncommon language pairs or any language pair with sparse bilingual data will benefit the most by using monolingual data.
To learn more about leveraging monolingual data to train your KantanMT engine; send the KantanMT Team an email and we can talk you through the process (info@kantanmt.com), alternatively, check out our whitepaper on improving MT engine quality available from our resources page.