NOTE-1344: Osmelak et al. — PETra

Contribution

A released corpus of pragmatic explicitations in human translation, in the translation-studies sense: “fragments of translated text that make implicit cultural or background knowledge explicit to the reader” (§3, p. 8758). It holds 2,982 sentence pairs by Table 2’s counts (2,989 rows in the release), English source and twelve targets, each addition bracketed in the target and labelled with one of 19 subtypes in four families (Table 1, p. 8760) and a style, replace or add (p. 8761). It also gives a pipeline that finds such pairs: null-alignment extraction, then a multilingual BERT classifier trained by active learning. Before it there was no multilingual, labelled corpus of this phenomenon; Han et al.’s WikiExpl (2023) is described as a set of generated explicative additions, not a validated corpus (p. 8757).

Key insight

What a translation adds for its new audience can be found mechanically, as target tokens with no source alignment near a named entity, and it is mostly mundane: “1 mile” → “1.6km”, “Tucson” → “Tucson (Arizona)”, “Myanmar” → “Myanmar ( former Birma )” (pp. 8758–8761). The paper’s framing example is “Angela Merkel, former German Chancellor”: “This added phrase does not alter the literal meaning but enriches the translation by making implicit, tacit knowledge in the source explicit in the target” (p. 8756). The resource is a catalogue of additions to content, not of changes to act, stance or footing.

Assumptions

• English source, existing human translations. Every pair is English → X, taken from TED-Multi subtitles and Europarl (§4.1, p. 8759). Nothing is known of the translators’ briefs. There is one route per item.
• A candidate needs a named entity and an unaligned content word. The extraction keeps “sentence pairs containing both a named entity (PERSON, NORP, FAC, ORG, GPE, LOC, PRODUCT, EVENT, WORK_OF_ART, LAW, or LANGUAGE) and an unaligned content word (NOUN, PRON, or PROPN) in the target text” (p. 8760). Additions with no entity nearby, or with no unaligned noun, are outside the net. The authors say the methods “may overlook more implicit pragmatic or cultural adaptations” (Limitations, p. 8764).
• Exclusions by rule. Pure translations (“Germany” → “Deutschland”), acronym changes where both forms are equally familiar, simplifications and idiomatic rephrasing are excluded (pp. 8758–8759), so the category boundaries are the authors’ judgement of what counts as cultural.
• Binary detection, multi-type annotation. The classifier is “bert-base-multilingual-cased for binary classification” (p. 8762), though §2 says the authors “adapt a multi-label active learning framework” (p. 8758). Types and styles are assigned by hand.

Key results

• Corpus size. The abstract says “3,000 sentence pairs” (p. 8756); the text, “Approximately 3,000 instances were manually labeled across the full dataset” (p. 8761). Table 2 (p. 8761) gives, per target language, pairs by source (P = found by the final classifier, E = null-alignment extraction, T = active-learning training) and type instances. I summed it: P 2,250, E 507, T 225, so 2,982 pairs; three quarters were found by the classifier. The release has 2,989 rows (POOL 2,257, EXTR 507, TRAIN 225); the only difference is the Europarl German pool, 59 rows against Table 2’s 52.
• Type counts. Table 2’s totals are ENT 992, SYS 1,722, LING 744, ADD 296, Σ 3,754 type instances. The ADD column as printed sums to 307 and Σ to 3,765, so the printed ADD and Σ totals are 11 short; the other three columns and every row sum correctly. The largest single subtype in the release is MEAS_CONV, 1,093 of about 3,880 labels.
• Language coverage is uneven. German (TED) has 801 pairs and 1,072 type instances; Greek has 72 pairs, Arabic 65. Persian, Hebrew, Thai and Arabic have no extracted or training pairs at all, only classifier finds (Table 2). Europarl appears for German only in Table 2, though footnote 4 lists Europarl rates “on DE, ES, IT, PT, FR” (p. 8761).
• Rates. “Of all candidate sentence pairs, 6% (2.6% for TED, 42% for Europarl) were extracted by EXTR. 3% of the sentence pairs (2.5% for TED, 7.7% for Europarl) were classified as containing cultural explicitation by CLF” (p. 8761). These are rates within the filtered candidate set, not within the corpora.
• Classifier. Seed 100 (one third positive), fixed test 300 (§5.1, p. 8762); 13 cycles, 100 then 50 queries each, then 1,000 more annotated German and Spanish pairs (L14). “Active learning improves performance by roughly 7–8 percentage points” (p. 8763). In-language at L13: “German reaching 0.88 accuracy and 0.57 F1, and Spanish 0.81 accuracy and 0.59 F1” (p. 8763). Transfer at L14: Portuguese and Dutch “0.82–0.83 in accuracy and 0.80–0.82 in F1”; Greek “accuracy peaking at 0.74 and F1 at 0.76” (p. 8763). Each transfer test set is 100 null-alignment pairs, and “These datasets only contain very clear cases of explicitations, while the German and Spanish test sets contain also more subtle cases, due to the proficiency of the annotator” (footnote 5, p. 8763). Results are averaged over five seeds (Fig. 2 caption).
• The abstract’s headline mixes two settings. “up to 0.88 accuracy and 0.82 F1 for the best transfer languages” (p. 8756) joins German’s in-language accuracy at L13 (F1 0.57) with the best transfer F1 at L14 on easier test sets.
• Typological patterns, descriptive only. MEAS-DIM is “particularly common in Germanic languages”; “European languages usually convert customary units to metric ones, while Semitic languages often keep the original measurement systems”; “German translations tend to assume more cultural background knowledge, using fewer explicitation strategies overall” (p. 8764). No test, rate or count backs these sentences, and the last sits oddly with German having the most pairs and instances in Table 2, which reflects annotation effort as much as translation.

Claims

id	claim	strength	support
C1	A labelled corpus of about 3,000 EN→X pairs with bracketed explicitations in 12 target languages exists and is released	strong	Table 2; release checked (2,989 rows)
C2	Entity and system-level (measurement, institution) explicitations are the most frequent types	moderate	Table 2 counts; but the extraction requires a named entity and most pairs were classifier-found, so the ordering is partly built in
C3	Active learning raises the detector’s accuracy by 7–8 points	moderate	Fig. 2a, five seeds, one fixed 300-item test set per language
C4	A German–Spanish classifier transfers to Portuguese, Dutch, Italian, French (best F1 0.80–0.82), less to Greek	weak	100-item test sets of “very clear cases” (fn. 5)
C5	Explicitation strategies differ systematically by language family	weak	qualitative statements, p. 8764; no rates normalized by candidates
C6	”pragmatic explicitation is a consistent and measurable phenomenon across languages” (p. 8763)	weak	no agreement statistic; one annotator implied (fn. 5)

Concepts

• Pragmatic (cultural) explicitation — an addition or replacement in a translation that supplies background the target audience is assumed to lack (§3). Klaudy’s sense, set against obligatory or structural explicitation (p. 8757). Not pragmatics in the sense of act, force or footing.
• Null alignment — a target token that eflomal and SimAlign leave unaligned, read as an addition (p. 8760).
• Style — R, the source term is replaced (“1 mile” → “1.6km”), or A, material is added beside it (“1 mile (1.6km)”) (p. 8761).
• CLEAR and DEIX — the two subtypes nearest discourse pragmatics: replacing a pronoun by its referent, and resolving “here” to “in the US”. CLEAR covers information “known to the source audience, for example due to visual input, or prior discourse input” (p. 8761). 93 and 27 labels in the release.

Connections

• Klaudy 1993, Blum-Kulka 1986: the explicitation typology and the explicitation hypothesis the corpus operationalizes (p. 8757). Not held.
• Han, Boyd-Graber and Carpuat 2023 (WikiExpl) and Lou and Niehues 2023: automatic explicitation and audience-specific explanations, cited as generation work (pp. 8757–8758). Not held.
• The sibling resources in this batch: VakyArth (LIT-1817), TyDiP (LIT-1799) and Ward and Marco’s similarity judgments (LIT-1812), all read the same day for CASE-185. PETra is the only one of the four built from translations people actually made.

Bearing on the record

• CLAIM-257 (change of cover): a natural source of the stimuli it needs, and nothing more yet. CLAIM-257 asks that a target’s new question be put to source readers too, so that an addition costs. PETra’s add-style items (1,958 of the release’s labels; the other 1,902 are replacements, R, which change rather than add) are naturally occurring cases of a target saying something the source did not: the Dutch target glosses Yellowstone as a national park in the US known for its geysers, and the Spanish one places Monterey in California and adds that it hosted the conference (Table 3, p. 8762). Each comes with its source and a bracket marking what was added, which is what a study would need to ask source readers the added question. The corpus carries no receiver judgements, so it says nothing about whether such additions change what readers take the utterance to do, which is CLAIM-257’s defeat condition. Most items add referential knowledge, so they test CLAIM-257 on its easiest ground: an addition that every source reader would affirm (“Tucson” is in Arizona). Items where source readers would disagree or not know are the informative ones, and PETra does not mark them.
• CASE-185: usable as a pool of candidate items, not as labelled items. Three limits. (a) Reliability. No inter-annotator agreement is given anywhere; footnote 5 speaks of “the annotator”; most pairs were found by a classifier, with “manual spot checks” for the large-scale annotation (p. 8763), while the text also says about 3,000 instances were “manually labeled” (p. 8761), so how many corpus labels a human assigned is unclear. The case’s Stage 0 rule (targets fixed in advance, two passes, judges’ agreement) cannot be met from PETra’s labels; items would have to be re-annotated. (b) Routes. Every pair is a single English → X translation. TED-Multi is multi-parallel, so an English sentence may have several targets, but no X → Y translation is given, so no relay route can be read off. (c) Kind. The items are additions to propositional background, not requests, promises or honorific choices. The case’s list of items is right to name them as one kind among several.
• CASE-055: no bearing on Stage 0 beyond the lack of a reliability figure. PETra measures no stance or footing.
• CLAIM-202: none directly. Nothing here concerns response to later framings.
• CLAIM-226: a minor instance. Whether “Ivy League” → “universities of excellence” is an explicitation is decided for the target audience by one annotator. The paper’s own caution is that results are “diagnostic indicators rather than exhaustive measures of cross-cultural representation” (p. 8764).
• Should produce THEORY-1153 (Proposed): the distribution of explicitation types, with what it does not establish.
• Machine-learning practice. The active-learning recipe (query strategies, 100-sample cycles) is a method, but the paper states it as its procedure, not as a recommendation, and its main product is a corpus. Not flagged for the anthology.
• Correction to the LIT’s description. The summary’s “2,900 sentence pairs” is 3,000 in both the arXiv v2 and the LREC abstract; Table 2 sums to 2,982 and the release has 2,989. The twelve “language pairs” are all English → X.

Limitations

• No reliability statistic. No kappa, no agreement, no second annotator is reported for detection, type or style.
• The sample is defined by the extractor. Additions without a named entity and an unaligned noun are missed, so category proportions and language differences are not base rates of explicitation in translation.
• Classifier-found majority. 2,250 of 2,982 pairs come from the final classifier (Table 2). Whatever the classifier misses, the corpus misses.
• Evaluation sets are small and differ in difficulty. 300 items for German and Spanish, 100 “very clear” items for six transfer languages, and none for Persian, Hebrew, Thai or Arabic, which are in the corpus anyway.
• Arithmetic slip. Table 2’s ADD and Σ totals are 11 below the column sums.
• Domain. The authors: the corpus “remains biased toward high-resource languages and formal domains such as TED talks and parliamentary proceedings” and the patterns “may not generalize to low-resource, oral, or literary contexts” (p. 8764).

Open questions

• How reliable are the type labels? A second annotator on a random sample of the release, with kappa per family, would say.
• Do target readers of an add-style item read the utterance differently from target readers of the same item with the bracket removed? That is the receiver study the corpus makes possible and does not run.
• What is the rate of explicitation per sentence when candidates are not filtered by named entity? A random sample of TED-Multi pairs, annotated blind to the extractor, would give it.