A pivot language is a third or intermediate language that can bridge the gap between language pairs. For example, if there are translations between English to French and the same English to Spanish available, through the pivot language English, translations between French and Spanish can be generated.
The major drawback and concern of generated translations through pivoting is the translation quality, as it is possible to produce erroneous translations by transferring errors or ambiguities from a language pair to another through the pivot language. However, when language resources in specific language pairs do not exist or are scarce, the use of pivot languages as data bridges can prove to be a convenient linguistic shortcut for offering language services or building and enhancing language resources.
This is the case, for example, for most resources in Asian, Arabic and other “exotic” languages between them or in connection to European languages. Moreover, as referred to the TAUS article: “Machine Translation and Asian Languages”(2012), the majority of translation scenarios that involve Asian languages and, in particular, translation from and into Chinese or Japanese, require much more language data than translation between European languages, for instance.
Selection of the pivot language
English is, in general, the pivot language of choice due to the richness of the available language resources or the available language competency, in combination with other languages. However, factors like language relatedness in terms of language family, language structure and perplexity, vocabulary and so on, can affect the choice of the pivot language for a given language pair with limited language coverage and yield better translation quality results. The two papers below refer to selection of pivot language for the area of Statistical Machine Translation, however the discussed insights can also be relevant to other areas of application.
In the paper “On the Importance of Pivot Language Selection for Statistical Machine Translation” (2009), M.Paul et al. investigate the appropriateness of languages other than English as pivot languages. The experimental results revealed that English was indeed more frequently (45.5% out of 110 language pairs) selected as the best pivot language over any other examined language. However, its usage is limited to translations between Indo-European languages and some Asian languages like Thai or Vietnamese. Experimental results using state-of the-art statistical machine translation techniques to translate between twelve languages revealed that the translation quality of 61 out of 110 language pairs improved when a non-English pivot language was chosen, in this case Asian languages and especially Japanese, Malay, Indonesian, or Korean.
In the more recent paper “How to Choose the Best Pivot Language for Automatic Translation of Low-Resource Languages” (2013), M.Paul et al. provide new insights into factors such as language family, vocabulary, sentence length, language perplexity and so on, that make a pivot language effective and investigate the impact of these factors on the overall pivot translation performance for translation between 22 Indo-European and Asian languages. Experimental results using state-of-the-art statistical machine translation techniques revealed that the translation quality of 54.8% of the language pairs improved when a non-English pivot language was chosen.
Areas where a pivot language can be deployed
Pivot language approaches can be used in simultaneous or consecutive interpretation (oral) and in human translation (written), including translations from audio recordings with or without script. Pivot languages can also be deployed in building and enhancing bilingual lexicons, translation memories and machine translation systems.
In part I of this blog article we will discuss the first two areas and in part II, which will be published in the near future, we will discuss the other three areas.
Relay Interpreting i.e. when the interpreter works via a pivot language, is necessary when more than two languages are involved in an interpreted event and there is no available interpreter able to cover all of the languages, or when there is no available interpreter for a given language combination.
Relay interpreting has been employed throughout history when explorers, traders or conquerors came into contact with previously unknown groups. For example, Karttunen in “Between Worlds: Interpreters, Guides, and Survivors”, 1994) mentions the use of relay interpreting between Totonac, Nahuatl, Maya, and Spanish during Cortes’s conquest of Mexico.
Today, this approach is used extensively in large international organizations such as the European Union: in a conference interpretation, when a delegate speaks a language that is not covered by an active language booth, this booth connects via an audio link to another booth that covers this language and “takes the relay”: the interpreter works through a pivot language. The quality of the target language version depends on the language competence of both interpreters’ and the choice of the pivot language.
In multilingual institutions as the European Commission or European Parliament, where currently documents have to be translated into 23 languages rapidly, these can sometimes be translated into the pivot languages (usually English, French or German) and then translated to the other languages. This practical and cost-effective approach requires further quality control, mainly due to possible invoked ambiguities when translating between languages with different constructions in this two-step translation approach.
An example from Game Localization
In the game localization industry, English language is often used as a pivot for localizing Japanese games into other European languages. Sometimes the games are dubbed into English and the dialogues are used as a basis for producing the subtitles into other languages. The translation decisions made into English (in this case) have to be followed for the other languages as well. To address issues that may undermine the original message and translation quality, translation strategies for the Japanese (in this case) games are required to determine what to domesticate and what to foreignize.
An interesting example (which can also be considered as a cultural filtering) is given by O’Hagan’s article: “Multidimensional Translation: A Game Plan for Audiovisual Translation in the Age of GILT” (2005) is in a scene of the game Final Fantasy X, where the last words spoken by the principal character Yuna to her lover Tidus were translated into English as “I love you" in the North American version from the original Japanese words “ありがとう(arigato) [thank you]”. Despite of the fact that this translation version caused a negative reaction among some fans that considered it at odds with Yuna’s introverted character, the European translators had, however, no option but to base their subtitles on the American version, since this was the one that players would listen to in the European versions of the game.
The pivot language approach for the above discussed areas can be seen as a practical, time and cost-effective option for producing translations for low-resourced and rare language pairs despite the further quality control or translation strategies that may be needed to complete the translation process. The right selection of the pivot language for each translation scenario may play an important role in the quality of the generated translations.