Terminological resources increasingly use machine translation as a method to speed up time and reduce costs. With a view to enhancing the multilingual representation of multiword terms (e.g. passive stall-regulated wind turbine) in terminological resources, we describe an analysis of English-Spanish multiword term translation in various machine translation systems, paying special attention to the errors encountered. A comparison of machine translation output with the equivalents found in a comparable corpus is also presented. Even though machine translation often shows errors, it can serve as a basis for human post-editing, thus saving time and costs in terminological work. Comparable corpora, on the other hand, offer better results, but searches are more time-consuming.
With a view to expanding markets and disseminating knowledge, specialized texts generate a large
volume of translations. Terminological resources should assist in this respect by means of the inclusion
of multilingual information. In this sense, machine translation is increasingly being used as a method
to speed up time and reduce costs [
This paper focuses on the translation of distinctive units of scientific texts, i.e. multiword terms (e.g.
passive stall-regulated wind turbine), which pose problems both to human translators and natural
language processing systems. However, multiword term machine translation has not been the focus of
attention with some exceptions, such as [
In order to enhance the multilingual representation of multiword terms in terminological resources,
we carried out the following tasks: (i) we analyzed English-Spanish multiword term translation in
various machine translation systems; (ii) developed a proposal of the causes that may generate errors in
multiword term machine translation; and (iii) compared machine translation output with the equivalents
that may be manually found in corpora. For this purpose, a set of three-, four-, and five-term English
multiword terms related to environmental science were extracted from a specialized corpus on this field
(10,228,919 words, [
Far from the classic challenging view of machine translation, according to which it would replace human translators, machine translation also presents opportunities not only to human translators, as evidenced in the great demand for machine translation post-editing (i.e. reviewing and enhancing a machine translation), but also to terminologists. Evidently, including post-editing in the workflow brings added value to machine translation, minimizing possible mistakes and providing quality equivalents to be included in terminological resources.
Even though training a neural machine translation system by means of carefully selected corpora from the specialized subject field could provide better results than using generic machine translation engines, the truth is that translators usually do not have user friendly tools to train their own domainspecific engines. For this reason, the selected English multiword terms were provided without context to different generic machine translation engines: Google Translate and DeepL (neural systems), and Apertium (rule-based system).
To compare machine translations with equivalents found in corpora, parallel or comparable corpora
can be used. Parallel corpora are sets of original texts aligned with their translations, thus facilitating
the identification of equivalents. However, such corpora are scarce, especially in languages other than
English, and generally show a marked influence of the source text on the translation. In contrast,
comparable corpora are more useful. Since they are two sets of original texts of the same type and
subject, they can be used to analyze native expressions in each language [
Therefore, a Spanish comparable corpus was used, which includes environmental texts originally
written in this language (10,667,434 words). Techniques for identifying multiword term equivalents in
corpora [
Multiword terms pose problems both to human translators and natural language processing systems
since their adequate translation must consider aspects such as their internal dependencies, the semantic
relation between constituents, the specialization of elements, etc. [
Not surprisingly, our results revealed that machine translation systems’ output varies in the different engines. They often show errors of different nature and magnitude, which were used to establish the different causes behind them and could be used to enhance machine translation systems. These errors include: (i) the wrong identification of internal dependencies ([doubly fed] [induction generator] > inducción alimentada doblemente generador, lit. *generator doubly fed induction); (ii) the wrong translation of constituents (wave turbulence interaction parameterization > interacción de turbulencia ondulatoria parameterization); and (iii) the wrong identification of the internal semantic relation (windgenerated electricity > viento-electricidad generada, lit. *generated wind-electricity). However, machine translation can serve as a basis for human post-editing, thus saving time and costs in terminological work.
Comparable corpora, on the other hand, offer better results, but searches are more time-consuming. Ideally, these different techniques should be integrated into translators’ and terminologists workflow, something that language service providers in the 2020s are bound to do. Furthermore, these results can be integrated into training for future translators and terminologists, who will have to work in this everchanging reality.
This research was carried out as part of projects PID2020-118369GB-I00, Transversal integration of culture into an environmental terminological knowledge base (TRANSCULTURE), funded by the Spanish Ministry of Science and Innovation; and project A-HUM-600-UGR20, Culture as a transversal module in an environmental terminological knowledge base (CULTURAMA), funded by the ERDF Operational Programme for Andalucía 2014-2020.