nucleation

THEORY-1153: In the PETra corpus of English-source TED and Europarl translations, the additions for the target audience that a named-entity and null-alignment filter, and a classifier trained on its finds, turned up, as labelled by one annotator with no agreement reported, are mostly system transfers, chiefly unit conversions, and entity changes, and their mix differs by target language

2 min read

Properties4
StatusActive
Version2
Tagstranslation, pragmatics
SourceLIT-1798 — PETra: A Multilingual Corpus of Pragmatic Explicitation in Translation

Graph View

Source

Osmelak, Dutta Chowdhury, Sentsova, España-Bonet and van Genabith (2026), LIT-1798, §§3–6 and Table 2, pp. 8758–8764 of the LREC 2026 proceedings; read in NOTE-1344.

What was actually shown

English sentences from TED-Multi and Europarl were aligned to their translations in twelve languages with eflomal and SimAlign. Pairs with a named entity and an unaligned target noun, pronoun or proper noun were annotated by hand (p. 8760). A BERT classifier trained by active learning on German and Spanish found further pairs (§5). Each explicitation was typed under a four-family schema (Table 1, p. 8760).

Table 2 (p. 8761) counts 992 entity, 1,722 system, 744 linguistic and 296 added-information instances (the ADD column as printed sums to 307). In the release, measurement conversion alone accounts for 1,093 of about 3,880 labels. The mix differs: German and Spanish TED have mostly system transfers (708 of 1,072; 352 of 570), Persian mostly entity changes and added information (99 and 63 of 200), Europarl German mostly entity changes (100 of 172).

What could have come out otherwise: the filter requires a named entity, so entity-linked additions were bound to be common. That measurement conversion, which needs no entity of its own, leads is less forced. It may still follow from the sentence carrying some entity, or from the classifier learning units as an easy cue.

What this does not say

  • Not a rate of explicitation in translation. The counts are within a filtered candidate pool. The paper’s own rates (2.6% of TED candidates extracted, 2.5% classified positive; p. 8761) are rates within that pool.
  • Not a reliable labelling. No inter-annotator agreement is reported, and the text points to one annotator (fn. 5, p. 8763). Three quarters of the pairs were found by a classifier (2,250 of 2,982).
  • Not a cross-linguistic law. Persian, Hebrew, Thai and Arabic pairs are classifier finds only, with no test set in those languages. The family-level statements (p. 8764) are not backed by a statistic.
  • Nothing about receivers. Whether an addition changes what target readers take the speaker to be doing, or what source readers would say about the added content, is not measured.
  • Not pragmatics in the sense of act or footing. “Pragmatic” here is Klaudy’s audience-oriented explicitation of background knowledge.

Backlinks