Journal of Computational and Applied Linguistics

LANGUAGE CORPORA AS A RESOURCE FOR MACHINE TRANSLATION

2026-09-29T08:47:54+00:00

The article discusses the importance of linguistic corpora in both statistical and neural machine translation. It explains that these models require large, representative parallel corpora to capture language phenomena and determine translation probabilities between source and target languages. Corpus linguistics, the field dedicated to creating and structuring such corpora, is essential for developing linguistic resources that can support language learning, lexicography, and linguistic analysis.

MULTIDIMENSIONAL EVALUATION OF A DOMAIN-SPECIFIC MACHINE TRANSLATION ENGINE

2026-10-08T14:51:34+00:00

The study presented in this article aims to evaluate the performance of WIPO Translate, a domain-specific machine translation (MT) tool, against three commercially available MT systems: Google Translate, DeepL and OpenAI’s ChatGPT 3.5. The evaluation focuses on assessing the translation outputs of patent excerpts across five dimensions, based on the Multidimensional Quality Metrics framework (MQM), namely Accuracy, Audience Appropriateness, Linguistic Conventions, Style and Terminology, from English to Brazilian Portuguese, deriving an overall quality score from the error annotations. The findings suggest that WIPO Translate, despite being specifically trained on patent corpora and performing reasonably well, does not outperform the commercially available MT systems in terms of translation quality. Generic systems demonstrate higher OQS averages in most cases, with DeepL generally achieving the highest scores among them. Automatic quality metrics were also implemented with BLEU, BERTScore, COMET and COMET-QE. WIPO Translate also ranked lowest in the automatic evaluation metrics. These results indicate that the broad training datasets and advanced algorithms used by general-purpose MT systems provide a competitive edge, even in specialized domains such as patent translation. Future improvements to WIPO Translate could focus on refining its algorithms
and expanding its training data to bridge this performance gap. Google and ChatGPT also presented superior performances, the latter being an interesting resource for users to customise the translation through detailed prompts. WIPO Translate still performs well according to the automatic metrics, and can also be very useful as a corpora tool, assisting users in gauging patent style and drafting better patents on their own.

EMOTIONALTURK, A NEW STANDARD IN EVALUATING AUTOMATIC TURKISH EMOTION DETECTION: A CASE STUDY IN A CROSSLINGUAL TRANSFER SETTING

2026-09-29T09:29:16+00:00

Current Turkish emotion detection from text (EDFT) research, and EDFT research in general, does not justify the rationale behind the emotion framework used in their EDFT datasets. Adapting emotion frameworks directly from psychology research, the most popular ones based on physiology, creates a gap between what is latently available in textual data and what the model output is, regardless of its performance. Therefore, the article introduces a dataset, EmotionalTurk, built with an emotion framework empirically grounded in textual data. While EmotionalTurk is currently limited to a test set in size, the article seizes the opportunity to explore cross-lingual transfer capabilities of multilingual transformers – an area that has not been previously investigated – using mBERT-based and XLM-RoBERTaBASE-based models, and establishes the SOTA score, which also serves as the baseline in this study in cross-lingual Turkish EDFT.

CAN THE USE OF TEXT-TO-SPEECH TECHNOLOGY SUPPORT POST-EDITORS WORKING INTO THEIR L2?

2026-10-08T15:12:40+00:00

Speech technologies have benefited greatly from the rise of artificial intelligence (AI) and they have developed over the past few years. Many industries have harnessed their power with great gains. In our research, we investigated to what extent professionals in the localisation industry accept machine translation post-editing (MTPE) with the help of speech technologies, and more particularly before and after a relevant experiment. We also examined their attitudes towards this approach in MTPE depending on their academic and professional background. Participants who achieved significant gains in terms of error corrections and relative gain in terms of time spent post-editing, no matter how negative or positive they were before the experiment, became open to TTS-enabled MTPE. The academic background of the participants did not seem to influence their attitude heavily. However, their professional background created some patterns not only in their attitude before the experiment, but also after its completion.

CAN CAT TOOLS AND LARGE LANGUAGE MODELS TRANSLATE FROM UNRECOGNISED LANGUAGES?

2026-09-29T09:13:58+00:00

First designed as a case study for high-school students competing in the field of linguistics, this experiment is based on the hypothesis that CAT tools and large language models have arrived at a stage where they can translate texts from languages that have not been explicitly listed as source languages. We took an excerpt in old French ancien français from Chrétien de Troyes’s novel Le Chevalier de la Charrette, versified into octosyllables. We machine translated it into Bulgarian – first from Latin and French with Phrase, and then without specifying the source language with ChatGPT. Phrase yielded better results with Latin than with modern French. Still, ChatGPT’s translation was much better and could be considered as a gisting translation. We had a reference human translation – into modern French, non versified, and into Bulgarian, versified. When comparing the results, we found that Phrase couldn’t provide a gisting translation, while ChatGPT rendered the general events and characters present in the text, even though on a contextual level there were still a lot of discrepancies; detail was missing; the text was rendered in prose. The experiment was conclusive that LLMs manage to render the general meaning of a text from an unknown dead language. It was illustrative for the students who were introduced to the specificities of medieval French and were given some linguistics problems to solve, based on the particularities of this language and its versification. The experiment shows the extent to which cultural heritage such as medieval knights’ novels can be rendered accessible to modern-day public, without having to resort to
professional translation.