Topics — What a work is about. Nucleation holds whatever reading the owner files here, whether or not an Anthology of the SOTA topic could also hold it; instruction for ML practice belongs to the anthology (ADR-048). The words are seeded from the clusters the reading-time triage of 2026-09-25 found, and grow by decision.
Translation — translation and its theory — equivalence and its kinds (formal, dynamic, functional), interpretive resemblance, register and quality assessment, translation as skopos-driven action, intersemiotic translation between media, and what a translation preserves or changes (ADR-033).
11 of 1352 THEORY documents. Back to the full index.
| # | Title | Summary | Status |
|---|---|---|---|
| THEORY-170 v2 | Languages differ in what their grammar obliges a speaker to convey rather than in what they can convey, so interlingual translation must add some information and leave some unexpressed, and repeated translation can erode a message | Jakobson (1959), LIT-786, after Boas: every cognitive content can be expressed in any language, but each grammar obliges its speakers to answer certain questions (aspect, number, gender, tense, definiteness). Translating between grammars that ask different questions forces the translator to supply answers the source did not give and to leave unexpressed answers it did, and a chain of translations of an isolated sentence can lose its content, less so the richer the context. Argued from examples, not tested; it concerns cognitive content, not pragmatic force. | Proposed |
| THEORY-226 | LLM revisions of an essay move it further in sentence-embedding space than its author’s own revision does, and in a direction shared across essays, models and prompts | Abdulhai et al. (2026), LIT-997, read in NOTE-750, §4.2 and Figs. 1, 5, 13–25. Three LLMs revising 86 ArgRewrite-v2 drafts under five prompts moved them further than the students’ own 2021 revisions did, and in one common direction, under two embeddings. Shown in plots, with no distance, angle or test. Meaning is measured only as embedding position: the paper does not show that the writer’s intended meaning changed, and the “grammar” prompt came with content feedback. | Proposed |
| THEORY-270 v3 | In Goldwasser, Gruber, Kalai and Paradise’s common-nonsense model, where the source language and the plausibility prior are uniform over sets from which shared random nonsense is removed and translators are one-to-one, maximum likelihood with the prior alone translates without parallel data with error O(log|Θ| / (α·min(m, |T|))) with high probability, falling in the shared nonsense α, in the size of the source language and in the number of samples | Goldwasser, Gruber, Kalai and Paradise (NeurIPS 2023), LIT-941, read in NOTE-804. Proved for maximum likelihood in stylised random models with uniform languages and one-to-one translators; in one model the bound O(log|Θ| / (α·min(m,|T|))) is tight up to a constant. Information- theoretic only. The prior substitutes for parallel data in learning the decoder, not for information in a message. | Active |
| THEORY-806 | Translation without parallel data, as of an unknown animal communication system into a human language, is learnable from a plausibility prior over the target language alone, with error falling as the common ground between the languages grows and as the source language grows richer | The general reading that THEORY-270 held before the promotion audit narrowed it to the common-nonsense model its bound is proved in (batch 044). Goldwasser, Gruber, Kalai and Paradise (NeurIPS 2023), LIT-941, read in NOTE-804, motivate their framework by translating animal communication and state its lesson for translation in general, while saying their stylised models are not intended to capture natural language. Filed under ADR-047 v4 so the wide form stays a thesis evidence can confirm or reject. | Proposed |
| THEORY-1120 v2 | On 416 translations into English by four LLMs of 104 culturally loaded utterances in Hindi, Punjabi, Tamil and Malayalam, each scored for pragmatic adequacy by one native rater per language, the reference-based COMET metric correlates with the ratings at Pearson r = 0.813 and Spearman ρ = 0.860, against r 0.80 and ρ 0.79 for XCOMET and r 0.73 and ρ 0.79 for MetricX, with no test of the differences; four outputs picked after the fact that invert or misplace the source’s meaning get COMET scores of 0.65–0.86 against the rater’s 1 or 2 | Singh et al. (2026), LIT-1817, read in NOTE-1345. On 416 outputs (104 items × 4 models), wmt22 COMET correlates with 1–5 pragmatic-adequacy scores at Pearson 0.813, Spearman 0.860; XCOMET ρ 0.792, MetricX ρ 0.788 (Table 3). Four hand-picked outputs scored 0.65–0.86 by COMET and 1–2 by the rater. Each language had one rater, so there is no reliability against which to read 0.81, and the claim that metrics fail “especially for implicature and deixis” rests on per-phenomenon differences of 0.1 or less. | Active |
| THEORY-1134 v2 | A translation metric that predicts a sentence score and token-level error spans from one shared encoder reaches segment-level agreement with expert MQM annotation of Kendall τ about 0.4 and span F1 under 0.33, the best among open metrics and GPT-4 prompting at the time, while the score computed from its own spans correlates with its regression scores at Pearson 0.73 to 0.82, and at 0.90 to 0.92 with its final score, which includes the span score by construction | Guerreiro, Rei, van Stigt, Coheur, Colombo and Martins (TACL 2024), LIT-1816, Tables 2, 4 and 5, read in NOTE-1341. The numbers are the paper’s, on WMT22 News MQM (zh-en, en-de, en-ru). The paper reads the heads’ agreement as support for the spans as explanation; the record reads it as coherence. The top of the agreement, 0.90–0.92, is partly arithmetic, since the final score includes the span score with weight 2/9. Active as a report of one test set (ADR-065); the record’s reading of the gap, across different statistics, is CLAIM-507. | Active |
| THEORY-1142 | A translation between two logics presented as institutions is correct, in the Satisfaction Condition’s sense, when a translated sentence holds in a model exactly when the original holds in the model’s translation; given the model translation this fixes the sentence translation up to semantic equivalence, and when every model of the target is reached it preserves and reflects entailment, so one logic’s prover can decide the other’s; but the condition alone admits collapsing translations and is relative to a model translation it does not supply | Goguen and Burstall (1992), LIT-1661, read in NOTE-1348 from the published JACM article: Definition 32 (p. 125), Proposition 33 and Proposition 35 (p. 126), Theorem 39 (pp. 127–129). Active: proved, and checked on toy morphisms between propositional institutions. The collapsing morphism and the atom-negating morphism are my constructions, not the paper’s. It is a criterion relative to a chosen model translation, and says nothing about translation beyond truth conditions. | Active |
| THEORY-1143 | When one domain carries more information than the other, a translation trained only to make the round trip close cannot meet that requirement through a correspondence in the intermediate’s visible code, so the round trip closes through a channel the intermediate’s audience does not read; agreement of the round trip with its source therefore says nothing about whether the intermediate renders the source’s content for that audience | Chu, Zhmoginov and Sandler (2017), LIT-1800, §2 and §5, read in NOTE-1346. The authors argue it from the many photographs a single map stands for, and support it with THEORY-1127’s evidence on one model; they propose, and do not run, the test of adding capacity to the smaller domain. Proposed. It is the converse of the record’s collapse case for CLAIM-496: there routes agree because nothing travels, here because everything does, out of sight. | Proposed |
| THEORY-1145 | Unsupervised alignment of two monolingual word-embedding spaces recovers translations only where the spaces are close to isometric, and the distance from isometry depends on the corpora and the embedding algorithm as well as on the languages: adversarial alignment fails for English with Estonian or Finnish, across corpus domains, and between skip-gram and CBOW spaces of the same language pair, GW alignment is weakest on the same Finnish pair, and a spectral dissimilarity of small nearest-neighbour graphs tracks success | Søgaard, Ruder and Vulić (2018), LIT-1819, read in NOTE-1337, with the GW results of Alvarez-Melis and Jaakkola (2018), LIT-1818, read in NOTE-1336. Proposed: the evidence is single runs of one adversarial method, plus one GW table, and the authors themselves report that some seeds succeed where the method otherwise fails. What it does not support is a typological law about dependent marking, or the reading that a good fit establishes that the matched words mean the same. | Proposed |
| THEORY-1151 v2 | A learned translation metric trained with synthetic hallucinations and diagnostic perturbations alongside human MQM flags most injected localized critical errors (all but 0 to 17 percent, by type and model size) and ranks fully detached hallucinations below other translations with AUROC up to 0.96, and without a reference it ranks hallucinations better than with one; how far this reflects the synthetic training data has not been separated from how far it generalises | Guerreiro et al. (TACL 2024), LIT-1816, §7 (Tables 6 and 8, Fig. 4), read in NOTE-1341. Localized errors are SMAUG perturbations of English references; hallucinations are one de-en benchmark of real system outputs. The training data includes synthetic detached and oscillatory hallucinations and DEMETR perturbations. The localized-error table’s “All” row is inconsistent with its rows. The larger model is worse on oscillatory hallucinations. Active as a report (ADR-065); whether detection generalises is QUESTION-345. | Active |
| THEORY-1153 v2 | 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 | Osmelak et al. (2026), LIT-1798, read in NOTE-1344 from the LREC 2026 PDF. Of 3,754 type instances in Table 2, 1,722 are system transfers (mostly measurement conversion) and 992 entity changes; the mix varies by language. It is a count over a pool selected by a filter that requires a named entity, mostly found by a classifier, labelled with no reported agreement. It is not a base rate of explicitation in translation, and it says nothing about how readers take the additions. | Active |