PETra (published at LREC 2026) is the first multilingual, human-validated corpus and detection framework for pragmatic explicitation — translator additions such as entity descriptions, measurement conversions, and translator remarks that make implicit cultural content explicit. Built exclusively from TED-Multi and Europarl, it covers twelve language pairs in formal registers, with a bias toward high-resource languages, and reports that entity- and system-level additions dominate and are detectable by null-alignment extraction plus an active-learning classifier. The authors themselves flag the central unresolved question: these patterns may not generalize to low-resource languages, oral registers, or literary genres, and the surface-based extraction may miss subtler, discourse-motivated additions. No later work has tested either claim. The problem is to extend the corpus and detection methodology to these regimes and establish whether pragmatic explicitation is a register- and resource-robust cross-linguistic phenomenon or a register-specific artifact of TED/Europarl data.

Background

Explicitation is the translation shift in which a translator spells out in the target text information that was only implicit in the source — expanding an acronym, converting a measurement unit, glossing an unfamiliar institution, or adding a clarifying remark. Since the 1980s, translation scholars have debated whether such shifts reflect a general "explicitation hypothesis", and more broadly whether any features of translated text are translation universals rather than artifacts of particular genres, language pairs, or corpora (Chesterman 2004; Becher 2011). Corpus-based studies have examined explicitation in particular registers and constructs: connective insertion (Becher 2011), literary translation (Vahedi Kia and Ouliaeinia 2016; Mahmood et al. 2025, using the classic obligatory/optional/pragmatic typology on two Urdu-English novel translations), machine versus human translation in a single scientific report (Krüger 2020), and referential explicitation by large language models, neural machine translation, and human translators across news, academic, and fiction genres (Liu and Yao 2026), which already finds genre-specific behavior. On the computational side, Han et al. (2023) built WikiExpl, a dataset of explicitations harvested from Wikipedia links and used to generate explanatory additions that improve multilingual question answering, and Lou and Niehues (2023) generated audience-specific explanations in machine translation. These computational efforts generate or exploit explicitations; they do not provide a validated multilingual corpus of translator-produced pragmatic explicitation.

PETra (Osmelak et al., LREC 2026; arXiv:2511.02721) fills that gap. It draws on two parallel corpora — Europarl (proceedings of the European Parliament) and TED-Multi (transcripts of TED conference talks with translations) — covering twelve English-centered language pairs (target languages German, Spanish, Italian, Portuguese, French, Dutch, Greek, Swedish, Hebrew, Arabic, Farsi, Thai). Candidate explicitations are extracted automatically through null alignments: tokens aligned by word-alignment tools (eflomal, SimAlign) that appear on only one side of a translation, filtered by named-entity recognition (NER, tagging of names of people, places, organizations) and part-of-speech (POS) heuristics to isolate additions likely tied to cultural entities or institutions. Human annotators then label candidates into four categories — Entities (ENT), System Transfers (SYS, e.g., metric or currency conversions), Linguistic Adjustments (LING), and Added Information (ADD, e.g., translator remarks) — with an active-learning loop (the classifier selects the most informative instances for the next annotation round) improving classification accuracy by 7–8 points, up to 0.88 accuracy and 0.82 F1. The two headline findings are that entity- and system-level explicitations are the most frequent categories, and that classifier transfer succeeds best for target languages typologically close to the seed languages (German, Spanish).

The authors' stated limitation is that this evidence base is biased toward high-resource languages and confined to two formal registers (prepared talks and parliamentary debate), so the observed patterns may not generalize to low-resource languages, oral registers, or literary genres; moreover, the null-alignment-plus-NER/POS extraction captures surface-level additions and may overlook implicit, discourse- or inference-motivated pragmatic adaptations. Adjacent evidence makes this a substantive question rather than a formality: Liu and Yao (2026) show fiction preserves implicit references where news and academic texts explicitate, and Wu, Li and Li (2026) show that in consecutive interpreting (an oral register) human explicitation draws on situational and sociopolitical context that a machine system misses — but neither study uses PETra's taxonomy, corpus, or detection pipeline, and no published work has extended PETra itself.

Problem Statement

Determine whether the pragmatic-explicitation patterns established by PETra generalize beyond the high-resource, formal-register parallel corpora (TED-Multi and Europarl) on which it was built, and whether the phenomenon can still be captured automatically there. Concretely: (1) extend the PETra corpus construction and annotation scheme — null-alignment candidate extraction refined by NER/POS heuristics, human annotation into the four categories Entities, System Transfers, Linguistic Adjustments, and Added Information (or a documented extension of it) — to at least one register dimension — (a) literary genres (e.g., translated fiction) or (b) oral or spoken registers (e.g., interpreted or conversational material) — and to (c) low-resource language pairs, producing a human-validated annotated corpus; (2) measure, under this extension, the category-frequency distribution and the performance of the detection framework, and compare them against the original PETra regimes, so as to establish or refute the generalization of the two headline findings — the dominance of entity- and system-level explicitations, and the detectability and cross-lingual transfer of pragmatic additions; and (3) assess whether subtler, discourse- or inference-motivated pragmatic additions that surface alignment and NER/POS heuristics miss can be systematically identified under an extended detection or annotation procedure. These three components serve one scientific objective: establishing whether pragmatic explicitation is a register- and language-resource-robust cross-linguistic phenomenon or a register-specific artifact of TED/Europarl-style data. Whether the extension confirms or refutes generalization is itself the empirical question, to be decided by the extended corpus, the measured distributions and detection performance, and an analysis of register-, language- and corpus-convention confounds (including translation direction and genre provenance); no particular outcome is prescribed, and any generalization conclusion must be scoped to the regimes actually tested.

References

  1. Doreen Osmelak, Koel Dutta Chowdhury, Uliana Sentsova, Cristina España-Bonet, Josef van Genabith. PETra: A Multilingual Corpus of Pragmatic Explicitation in Translation. Proceedings of the Fifteenth Language Resources and Evaluation Conference (LREC 2026), pages 8756-8766, 2026. DOI: 10.63317/56tberz7nmwy. https://doi.org/10.63317/56tberz7nmwy (arXiv:2511.02721, https://arxiv.org/abs/2511.02721)
  2. HyoJung Han, Jordan Boyd-Graber, Marine Carpuat. Bridging Background Knowledge Gaps in Translation with Automatic Explicitation. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 9718-9735, 2023. DOI: 10.18653/v1/2023.emnlp-main.603. https://aclanthology.org/2023.emnlp-main.603/
  3. Ralph Krüger. Explicitation in Neural Machine Translation. Across Languages and Cultures 21(2):195-216, 2020. DOI: 10.1556/084.2020.00012. https://doi.org/10.1556/084.2020.00012
  4. Zhaoxia Liu, Guangyuan Yao. Referential explicitation in translations by large language models, neural machine translation system, and human translators across genres. Across Languages and Cultures 27(1):153-174, 2026. DOI: 10.1556/084.2026.01199. https://doi.org/10.1556/084.2026.01199
  5. Xinyue Wu, Guanji Li, Dechao Li. Machine Versus Human Explicitation in Consecutive Interpreting: Evidence From ChatGPT. Interpreting and Society: An Interdisciplinary Journal, 2026. DOI: 10.1177/27523810261471996. https://doi.org/10.1177/27523810261471996
  6. Mehdi Vahedi Kia, Helen Ouliaeinia. Explicitation across literary genres: Evidence of a strategic device? The International Journal of Translation and Interpreting Research, 2016. DOI: 10.12807/ti.108202.2016.a06. https://doi.org/10.12807/ti.108202.2016.a06
  7. Muhammad Asim Mahmood, Ansa Sattar, Muhammad Imran, Norah Almusharraf. Bridging cultures through explicitation: a corpus-based analysis of bilingual literary translations. Cogent Arts & Humanities 12(1), 2025. DOI: 10.1080/23311983.2025.2508210. https://doi.org/10.1080/23311983.2025.2508210
  8. Viktor Becher. When and why do translators add connectives? A corpus-based study. Target 23(1):26-47, 2011. DOI: 10.1075/target.23.1.02bec. https://doi.org/10.1075/target.23.1.02bec
  9. Andrew Chesterman. Hypotheses about translation universals. In Claims, Changes and Challenges in Translation Studies (Benjamins Translation Library 50), pages 1-13, 2004. DOI: 10.1075/btl.50.02che. https://doi.org/10.1075/btl.50.02che
  10. Miguel A. Jiménez-Crespo. Testing explicitation in translation: Triangulating corpus and experimental studies. Across Languages and Cultures 16(2):257-283, 2015. DOI: 10.1556/084.2015.16.2.6. https://doi.org/10.1556/084.2015.16.2.6
  11. Renhan Lou, Jan Niehues. Audience-specific Explanations for Machine Translation. arXiv:2309.12998, 2023. https://arxiv.org/abs/2309.12998