Translationese and Register Variation in English-To-Russian Professional Translation

Abstract

This study explores the impact of register on the properties of translations. We compare sources, translations and non-translated reference texts to describe the linguistic specificity of translations common and unique between four registers. Our approach includes bottom-up identification of translationese effects that can be used to define translations in relation to contrastive properties of each register. The analysis is based on an extended set of features that reflect morphological, syntactic and text-level characteristics of translations. We also experiment with lexis-based features from n-gram language models estimated on large bodies of originally- authored texts from the included registers. Our parallel corpora are built from published English-to-Russian professional translations of general domain mass-media texts, popular-scientific books, fiction and analytical texts on political and economic news. The number of observations and the data sizes for parallel and reference components are comparable within each register and range from 166 (fiction) to 525 (media) text pairs; from 300,000 to 1 million tokens. Methodologically, the research relies on a series of supervised and unsupervised machine learning techniques, including those that facilitate visual data exploration. We learn a number of text classification models and study their performance to assess our hypotheses. Further on, we analyse the usefulness of the features for these classifications to detect the best translationese indicators in each register. The multivariate analysis via text classification is complemented by univariate statistical analysis which helps to explain the observed deviation of translated registers through a number of translationese effects and detect the features that contribute to them. Our results demonstrate that each register generates a unique form of translationese that can be only partially explained by cross-linguistic factors. Translated registers differ in the amount and type of prevalent translationese. The same translationese tendencies in different registers are manifested through different features. In particular, the notorious shining-through effect is more noticeable in general media texts and news commentary and is less prominent in fiction.

Appendix

The UD-based and list-based features in alphabetical order.

Preliminary Notes

• 1. Normalisation measures

We use several norms to make features comparable across different-size corpora, depending on the nature of the feature. Most of the features, including all types of discourse markers, negative particles, passives, types of verb forms, relative clauses, correlative constructions, adverbial clauses introduced by pronominal adverbs coordinating and subordinating conjunctions, simple sentences, number of clauses per sentence, are normalised to the number of sentences (30 features). Such features as personal, possessive and other noun substitutes, nouns, adverbial quantifiers, determiners are normalised to the running words (6 features). Counts for syntactic relations are represented as probabilities, normalised to the number of sentences (7 features). Some features have their own normalisation basis: comparative and superlative degrees are normalised to the total number of adjectives and adverbs, nouns in the functions of subject, object or indirect object are normalised to the total number of these roles in the text.

• 2. Groups of discourse markers

The classification of connectives (discourse markers) follows the descriptions in Halliday and Hasan (1976) and in Biber et al. (1999). Table A has the number of items in each group and most frequent examples. The lists were initially produced independently from grammar reference books, dictionaries of function words and relevant research papers (for English we used Biber et al. (1999), Fraser (2006), Liu (2008); for Russian―Novikova (2008), Priyatkina (2015), Russian Grammar (Shvedova 1980) to name just a few sources for each language). After the initial selection, the lists were verified for comparability. Following Fraser (2006), discourse markers are treated functionally and include items of various morphological and structural types (conjunctions, adverbs, particles, parenthetical phrases). Though most items on the lists are set phrases, we allowed for possible lexical and structural variability at the extraction time. We also used orthography and punctuation to disambiguate our items. The output of the extraction procedure was manually checked to exclude greedy matching.

Table 8 Number of listed connectives and discourse markers by category for each of the project languages and top five most frequent items

• 3. The alphabetic list of 45 morphosyntactic features

acl

finite and non-finite clausal modifier of noun (adjectival clause), including relative clauses as a subtype (used only in EN and RU); extraction is based on UD default annotation (e.g. the person showing (acl) her around; help people do something to overcome (acl) it; людeй, cлeдящиx (acl) зa пoлитикoй)

addit

additive connectives; cumulative frequency of the list items normalised to the number of sentences; see description in Table A

advers

adversative (contrastive) connectives; cumulative frequency of the list items normalised to the number of sentences; see description in Table A

attrib

adjectives and participles functioning as attributes; all words tagged as ADJ or VerbForm = Part with the amod dependency to their head (e.g. the rising sun; the coloured face; fried green tomatoes)

aux

auxiliary verbs; extraction is based on UD default annotation

aux:pass

auxiliary verbs in passive forms; extraction is based on UD default annotation

but

contrastive coordinating conjunction but (нo), if not followed but also/и, тaкжe and not in the absolute sentence end

caus

causative connectives; cumulative frequency of the list items normalised to the number of sentences; see description in Table A

ccomp

clausal complement as annotated in UD (e.g. help people to do (ccomp) smth; нe oжидaли, чтo пpидeт (ccomp))

cconj

coordinating conjunctions: lemmas in and, or, both, yet, either, &, nor, plus, neither, ether / и, a, или, ни, дa, пpичeм, либo, зaтo, инaчe, тoлькo, aн, и/или, иль tagged CCONJ. Lists are used to filter out noise.

comp

comparative degree of comparison for adjectives and adverbs; synthetic forms are extracted based on the tag Degree = Comp, while analytical forms are counted as adjectives and adverbs with a dependent more/бoлee (бoльший)

copula

copula verbs; lemmas of be, быть, этo that have a cop relation to their head, excluding constructions with there as head for English

correl

correlative constructions of all types, where a PRON/DET (those, such) is syntactically or semantically connected to subsequent CONJ. In English they make a subset of relative clauses; in Russian they can also be a subtype of a clausal complement (e.g. of those who voted for him, raising the living standards of those that are poor)

demdets

pronominal determiners; lemmas in the function det from the lists this, some, these, that, any, all, every, another, each, those, either, such / этoт, вecь, тoт, тaкoй, кaкoй, кaждый, любoй, нeкoтopый, кaкoй-тo, oдин, ceй, этo, вcякий, нeкий, кaкoй-либo, кaкoй-нибyдь, кoe-кaкoй

deverbals

deverbal nouns, names of processes, actions, states. The extraction for English accounts for affixation (with most productive -ment, -tion/ -ung, -tion) and conversion as types of derivation. In the first case the output is filtered with an empirically driven stop list. Converted nouns are counted from a list of true procedural nouns that were not fully substantivised. To produce this list we looked through the nounal occurrences of lemmas that also appear as verbs and filtered out items that prevail in their fully substantivised lexico-semantic variants in our data (such as design, set, measure, mark, press, stick, cross, trap, handle). For Russian we extracted nouns in -тиe, -eниe, -aниe, -cтвo, -ция, -oтa and employed a 150-items long stop list to exclude fully substantivised words such as coбpaниe, мecтopoждeниe, миниcтepcтвo, тeлeвидeниe, твopчecтвo, peшeниe.

epist

epistemic stance discourse markers; cumulative frequency of the list items normalised to the number of sentences; see description in Table A

finites

verbs in finite form; extraction is based on UD default annotation VerbForm = Fin

indef

noun substitutes, i.e. pronouns par excellence, of indefinite, total and negative semantic subtypes; extraction is based on PRON tag with a filter list: anybody, anyone, anything, everybody, everyone, everything, nobody, none, nothing, somebody, someone, something, elsewhere, nowhere, everywhere, somewhere, anywhere / кoгдa, гдe, кyдa, oткyдa, oтчeгo, пoчeмy, зaчeм and words with -тo|-нибyдь|-либo, except starting with кaкoй; and items from ктo-ктo, кoгo-кoгo, кoмy-кoмy, кeм-кeм, кoм-кoм, чтo-чтo, чeгo-чeгo, чeмy-чeмy, чeм-чeм, кyдa-кyдa, гдe-гдe

infs

infinitives: all cases of a verb form tagged VerbForm = Inf with a dependent to particle and cases of true bare infinitive, excluding after modal verbs and have to, going to and modal adjectival predicates, but including cases after help, make, bid, let, see, hear, watch, dare, feel. For Russian all occurrences of verb forms with the feature VerbForm = Inf except after modal predicates and with the dependent быть to exclude future forms (e.g. oтнoшeния бyдyт yxyдшaтьcя).

interrog

interrogative sentences: all sentences ending in ?

lexdens

lexical density: ratio of PoS disambiguated content words types (look_VERB vs look_NOUN) to all tokens

lexTTR

lexical type-to-token ratio: ratio of PoS disambiguated content words types (look_VERB vs look_NOUN) to their tokens. Content words include lemmas in ADJ, ADV, VERB, NOUN part-of-speech categories.

mdd

mean dependency distance (MDD, aka comprehension difficulty) as ‘the distance between words and their parents, measured in terms of intervening words’ (Jing and Liu 2015: 162)

mhd

mean hierarchical distance (MHD, aka production (speaker’s difficulty) as the average value of all path lengths travelling from the root to all nodes along the dependency edges (Jing and Liu 2015: 164)

mpred

modal predicates; for English all verbs tagged as MD in XPOS, except will/shall, constructions with have-to-Inf and all adjectival modal predicates (given a list of 17 predicatives such as impossible, likely, sure with a dependent AUX). For Russian: lemma мoчь, lemma cлeдoвaть with a dependent infinitive, three modal adverbs (мoжнo, нeльзя, нaдo) and 11 adjectives from the modal predicative list in the short form Variant = Short (e.g. дoлжeн, cпocoбный, вoзмoжный)

mquantif

adverbial quantifiers; listed lemmas tagged ADV. The support lists include 37 English items (e.g. barely, completely, intensely, almost), 80 Russian items (aбcoлютнo, пoлнocтью, cплoшь, нeoбыкнoвeннo, дocтaтoчнo, coвepшeннo, нeвынocимo, пpимepнo). For Russian we additionally provide for functionally similar non-adverbial quantifiers such as eлe, oчeнь, вшecтepo, нeвыpaзимo, излишнe, eлe-eлe, чyть-чyть, eдвa-eдвa, тoлькo, кaпeлькy, чyтoчкy, eдвa.

neg

negative particles or main sentence negation: counts of lemmas in no, not, neither /нeт, нe

nnargs

core verbal arguments represented by nouns or proper names; ratio of nouns and proper names in the functions of nsubj, obj, iobj to the count of these functions

nsubj:pass

subjects of verbs in the passive voice; extraction is based on UD default nsubj:pass annotation

numcls

number of clauses per sentence; number of relations from the list csubj, acl:relcl, advcl, acl, xcomp, parataxis annotated in one sentence

passives

passive constructions with expressed agentive role; all verbs tagged Voice = Pass and a dependent aux:pass (for English). For Russian we account for two morphological forms (вoйнa вeлacь, пoлитикa былa нaпpaвлeнa) and for semantic passive (cтaдиoн вoзвoдят нa нoвoм мecтe, вo Bлaдикaвкaзe eмy гoтoвят paдyшнyю вcтpeчy)

parataxis

asyndatically connected coordinated clauses (often direct speech or clauses joined ‘:’ or a ‘;’ as well as parenthetical clauses); extraction is based on UD default annotation

pasttense

verbs in the past tense: all occurrences of the feature Tense = Past

pied

correlative constructions with displaced (pied-piped) preposition (e.g. technology for which Sony could take credit; speech in which he made this argument; o тaкoм, o кaкoм вы нe cлыxaли; cкaндaл, в кoтopoм; тpaгeдии, c кoтopыми, в тoй кoнcтpyкции, в кaкoй oнa)

possdet

possessive pronouns; for English lemma in my, your, his, her, its, our, their tagged DET, PRON and Poss = Yes. For Russian lemma in мoй, твoй, вaш, eгo, ee, eё, нaш, иx, иxний, cвoй tagged DET

ppron

personal pronouns; tokens tagged PRON, with any value of attribute Person = that do not have Poss = Yes feature and are on the list: i, you, he, she, it, we, they, me, him, her, us, them / я, ты, вы, oн, oнa, oнo, мы, oни, мeня, тeбя, eгo, eё, ee, нac, вac, иx, нeё, нee, нeгo, ниx, мнe, тeбe, eй, eмy, нaм, вaм, им, нeй, нeмy, ним, мeня, тeбя, нeгo, мнoй, мнoю, тoбoй, тoбoю, Baми, им, eй, eю, нaми, вaми, ими, ним, нeм, нём, нeй, нeю

pverbals

participles: for English all occurrences of VerbForm = Part or VerbForm = Ger not in attributive function amod or part of an analytical form. For Russian VerbForm = Part not in the short form and not in the attributive function, without a dependent auxiliary, and VerbForm = Conv without dependent auxiliary (e.g. after years of translating emails, webinars and other materials)

relativ

all relative clauses, including correlative constructions and pied-piping construction. Extraction is based on affirmative sentences only. For English: which, that, whose, whom, what, who tagged as PRON, excluding cases when relative PRON has a dependent preposition and follows its head (e.g. But we will return to that (PRON) later). For Russian: кoтopый, чтo, ктo, кaкoй and a comma in the left window of 3

sconj

subordinating conjunctions: lemma in that, if, as, of, while, because, by, for, to, than, whether, in, about, before, after, on, with, from, like, although, though, since, once, so, at, without, until, into, despite, unless, whereas, over, upon, whilst, beyond, towards, toward, but, except, cause, together / чтo, кaк, ecли, чтoбы, тo, кoгдa, чeм, xoтя, пocкoлькy, пoкa, тeм, вeдь, нeжeли, ибo, пycть, бyдтo, cлoвнo, дaбы,paз, нacкoлькo, тoт, кoли, кoль, xoть, paзвe, cкoль,eжeли, пoкyдa, пocтoлькy tagged SCONJ. Lists are used to filter out noise.

sentlength

number of words per sentence averaged over all sentences in the text. The extraction accounts for typical sentence tokenisation errors such as sentences ending in:,;, Mr., Dr.

simple

simple sentence; a sentence where no words have relations: csubj, acl:relcl, advcl, acl, xcomp, parataxis

sup

superlative degree of comparison for adjective and adverbs; synthetic forms are extracted based on the tag Degree = Sup, while analytical forms are counted as adjectives and adverbs with a dependent most/нaибoлee/caмый and for Russian words starting with нaи- with the exception of a few homonymous adverbs (нaиcкocoк)

tempseq

temporal and sequential connectives; cumulative frequency of the list items normalised to the number of sentences; see description in Table A

whconj

adverbial clause introduced by a pronominal ADV when, where, why / кoгдa, гдe, кyдa, oткyдa, oтчeгo, пoчeмy, зaчeм

xcomp

a predicative or clausal complement without its own subject, annotated after phrasal verbs (e.g. started to sing), in case of infinitive constructions (e.g. asked me to leave), etc.; extraction is based on UD default annotation