One of the core arguments against the use of machine translation (MT) by translators, especially those working in creative fields, is the machine’s so-called ‘inability’ to handle the metaphorical use of language to evoke images and feelings. Another is the stifling of creativity on the part of the translators when asked to post-edit machine translated text as opposed to giving life to a new version of an original that creates an equivalent effect to the audience in another target language. I was naturally intrigued when asked to contribute to a workshop on the topic of “Literary Machine Translation as a Human-Machine Dialectic”, organized by the University of Liège.
The workshop featured three panels in which MT researchers, practitioners, translators, experts in translation studies, and students discussed how MT technology could be best used to assist in the creative process of literary translation. The panelists talked about technical and creative aspects of literary MT, as well as presented reflections on use cases such as post-editing and MT systems adapted not only to the domain of literary translations, but to the style of a particular translator – personalized MT. One of the talks also described legal implications with respect to authorship rights when MT is used for the translation of literature.
It was very interesting to me to see not just panelists, but all workshop participants actively discussing the progress MT technology has achieved in recent years, which made it possible to even consider MT for the highly creative task of literary translation. At the same time, it was pointed out that MT still makes errors and that blatant post-editing of its output may lead to formally correct but quite dull translations that do not reflect the quality of the original literary creative work. It was reported that many MT systems still fail to address the needs of translator end users, such as consistency in the use of approved glossary-based translations across an entire book, for example.
While it is true that many MT implementations are not always done with the end user in mind, I wanted to highlight with my contribution the ways in which it is indeed possible for the MT to become a translator’s trusted colleague. My talk was entitled “Giving translator full control: customizable MT for literary translation”. I presented AppTek’s metadata-aware MT systems and showed how a sample from an English novel could be translated (into German or French, for example) using machine-translated suggestions presented to translators in an imaginary computer-assisted interface.
One of the most common issues with the use of neural MT is its lack of consistency with translating proper nouns, as was discussed by workshop participants. With AppTek’s customized MT, a translator is able to define glossary overrides in order to enforce their creative translations of character names when translating a novel, choose their preferred translations of terms, or even transfer word play into the target language, as in the example below (Image 1).
In Image 1, we see examples in the right column where AppTek’s generic multi-domain MT system output is provided, not adapted to literary content. Nevertheless, the system already provides a ‘genre’ parameter which can be set at a click of a button (prose in this example), so that the user obtains translations that are usually representative of the genre. The formality level or style of the translation can also be set as another parameter, for the entire document or chapter and can be adjusted for individual paragraphs or even sentences during the translation process if so needed (Image 3).
As with any text that deals with the translation of direct speech, the speaker gender makes a difference in the translation from a grammatical point of view as word-form endings will differ depending on gender. In a novel, dialog turns are provided with the use of paragraph breaks, so that the speaker identity of each direct speech sentence can be determined automatically and/or with the help of the translator (Image 2). By assigning gender labels to each speaker, the MT system is able to provide automatic translations which respect the speaker gender. This way translators do not need to worry about fixing such grammatical errors and the MT output can be more readily usable, as shown in the example below (Image 3).
Participants at the workshop seemed particularly interested in the ability of systems like AppTek’s MT to provide multiple translation alternatives for a given sentence or phrase, which could be useful to a translator when looking for a synonym or more creative alternative to be used in a particular context. It is important that alternatives are provided in such a way so that the ergonomics of using the MT are effortless, not requiring many mouse clicks.
As expected, there will be also instances in which the MT output is not useful to the translator. In such cases, it is preferable not to make it available at all, so that translators do not waste time going through the MT, or that the problematic words or phrases are highlighted so that they can make a more informed decision about using it. AppTek’s systems provide sentence-level and word-level confidence measures which can be used for this purpose by setting individualized thresholds that determine the confidence level below which an MT option should be hidden from view or highlighted (Image 4). Translators can then benefit by working only with the most reliable translations the system can offer. It is worth noting that document-level translation MT systems, such as AppTek’s, can be configured to use the extended context of neighboring sentences for the better translation of pronouns, references, and for better context disambiguation in general.
All the examples shown in the images above were generated directly via AppTek’s cloud-based API using the ‘extended context’ feature. All other metadata-based controls were also provided through specific API parameters. If you would also like to experiment with metadata for your MT use case, please contact AppTek for a demo account.
I left Liège having enjoyed a full day of fruitful discussions that helped spark new ideas. For me, it was important to get direct feedback from translators on how they think they can use MT effectively and creatively, as well as to get their impression of the current levels of MT quality and its role in their work. Some of this feedback will be incorporated into our R&D process at AppTek, so look out for new features in future editions of our MT models. I really think we need more opportunities which enable the exchange of ideas between representatives of humanities like translators and translation teachers on one side, and natural sciences – computer scientists and computational linguists like me – on the other.
I am now looking forward to the Languages and the Media conference in Berlin (7-9 November), where I hope to have more interesting discussions with translators from another creative field, that of media localization. AppTek is proud to sponsor and exhibit at the conference, so look out for our booth and let us show you what we have been working on!
AppTek is a global leader in artificial intelligence (AI) and machine learning (ML) technologies for automatic speech recognition (ASR), neural machine translation (NMT), natural language processing/understanding (NLP/U) and text-to-speech (TTS) technologies. The AppTek platform delivers industry-leading solutions for organizations across a breadth of global markets such as media and entertainment, call centers, government, enterprise business, and more. Built by scientists and research engineers who are recognized among the best in the world, AppTek’s solutions cover a wide array of languages/ dialects, channels, domains and demographics.