TECHNOLOGIES IN TRANSLATION PRACTISE

3.2 Catford – the Sentence-Bound Linguistic Approach to Translation

This sentence-focused principle of translation assessment reminds us of Catford’s early

linguistic approach to translation (1965). Because translation, according to Catford (1965, p.

1) is ‘an operation performed on languages’, a theory of translation needs to draw on a

linguistic theory, and Catford selects the General Linguistic Theory presented by Halliday

(1961). The highest rank of this grammar is (as in most grammars) the sentence, and Catford

considers that it is at this grammatical level that translation equivalence can most often be

established:

SL [source language] and TL [target language] texts or items are translation equivalents

when they are interchangeable in a given situation. This is why translation equivalence

can nearly always be established at sentence-rank—the sentence is the grammatical unit

most directly related to speech-function within a situation. (Catford, 1965, p. 49)

(italics in original). That said, his model allowed for shifts, ‘departures from formal correspondence’ between levels and between categories (Catford 1965, p. 73). For example, a sentence can be

translated into a phrase or a word, and vice-versa, which constitutes a level shift. A category

shift occurs when structures, classes and units do not correspond formally between a ST and a

TT. For example, a clause structure shift takes place when the English clause ‘John loves

Mary’, which has the structure Subject – Predicator – Complement, is translated into Gaelic

as ‘Tha gradh aig Iain air Mairi’, which has the structure Predicator – Subject – Complement

– Adjunct (Catford 1965, pp. 76-77).

3.3 The Textual Approach to Translation

The notion of translation equivalence at the sentence level and the observation of its

manifestation in the target text has limitations. For example, it does not explain what the

translator really does when translating (Fawcett, 1997, p. 56). Since the late twentieth century,

Translation Studies scholars have drawn on other branches of linguistics which had

developed by then to explain the real-world experiences of producers, readers and translators

of texts. These include, among others, text linguistics. ‘Text’ is understood in different ways

by different schools of linguistics, often being interchangeably used with ‘discourse’. The

concepts were famously imported into translation by Hatim and Mason (1990) from De

Beaugrande and Dressler (1981).

De Beaugrande and Dressler (1981, p. 3) maintain that a text is a communicative occurrence

which meets seven standards of textuality: cohesion, coherence, intentionality, acceptability,

informativity, situationality, and intertextuality. These seven standards are all necessary to

make the text ‘communicative’, but the standards that are particularly relevant in the current

discussion are cohesion and coherence. Cohesion in text is produced by certain linguistic

devices, which are language specific. For instance, English has five such devices: reference,

substitution, ellipsis, conjunction and lexical cohesion (Halliday and Hasan, 1976). These

devices link different parts of text together, which enables the reader to recognise one part of

a text as connected to another part of the text. However, textuality involves another important

concept, coherence, because a text must hang together, ‘both linguistically and conceptually’

(Hatim and Mason, 1990, p. 192). For it to hang together conceptually, it needs coherence as

well as cohesion. Coherence is not an observable textural feature, but, instead, is produced by

assumptions held by the reader. To recognise a set of linguistic devices as a vehicle for

creating coherence, the reader needs to have world knowledge along with knowledge of

coherence relations such as cause-effect, problem-solution, and temporal sequence. Both

cohesion and coherence are standards of text (Hatim and Mason, 1990, p. 195).

Application of these concepts are not straightforward in translation, though. This is because

while the knowledge of the sequence of coherence relations is (most likely) shared by users

of the source and target languages, cohesive devices in those languages are language specific.

The translator’s job is to negotiate between the two systems and produce translation which

satisfies both systems and achieves equivalence at text level (Baker, 2018/1992, pp. 134-234).

As Shreve puts it, paying attention to textual properties can lead to a ‘quantum leap ... in the

“textual quality” of the translation’ (Shreve, 2017, p.175); but text production through

negotiating between two systems (i.e. translation) requires considerable linguistic knowledge

and skills and major cognitive effort. This is because the unit of translation is considered to

be a sentence (Huang and Wu, 2009) and, as a body of cognitive research indicates, the way

that translation proceeds (i.e., sequential and step-by-step) brings the translator’s focus,

necessarily, down to the sentential level (Shreve, 2017, p.175). This leads us to deduce that

producing a cohesive and coherent text on the text level in translation requires additional

effort and skills (such as cross-sentential revision) by the translator compared to mono-

lingual text production.

4 The Influence of Technologies on Translation Practice

4.1 The Influence of TM on Text Production

Considering the cognitive processes involved in translational text production, as well as the

mechanism of the TM function in a CAT tool, it is easy to understand that TM-assisted

translation exacerbates the difficulty of producing translation with appropriate textuality. The

translator is required to minimise the adverse effects caused by the sentential restrictions as

well as the influence from the CAT tool interface and the nature of TM. Under these

circumstances, the translator must pay sufficient attention to translation suggestions from the

TM and apply necessary edits so that the final translation achieves maximum textuality.

Bowker (2006) offers detailed examples of the difficulties involved in this. One obvious

example is related to the sequence of coherence relations. A sentence in a text stands in

certain logical relations to other parts of the text, but these relationships obtain only in that

particular text-context configuration. When a sentence is stored in a TM, the context is

stripped away. And when the sentence is suggested as a recyclable segment in a new

translation, the suggestion is not only made without regard to context, but may conflict with

the new context in the target text (TT). Therefore, to make the suggested sentence work in the

new translation, the translator needs to ‘work outside the artificial boundaries of sentences, so

the sentence-by-sentence approach imposed by TMs may not be conducive to effective

translation of the text's message as a whole’ (Bowker, 2006, p. 180). Another example

concerns polysemy. Bowker uses the example of the French translation of the English

expression ‘empty the pipe’, which will have different meanings in a text about plumbing

(pipe=tuyau) and in a text about smoking (pipe=pipe) (Bowker, 2006, p. 179). Since a

suggestion from the TM database is presented out of context, the translator will need to check

the context, and if necessary, consult relevant reference materials.

 4.2 The Influence of MT on Text Production

The rift between the concepts of translation quality at the text level and at the sentence level

is evident in text production involving MT, too.

Lumeras and Way (2017), drawing on Kay (2014), illustrate the rift using the dichotomous

concepts of ‘syntactic translation’ and ‘pragmatic translation’. Lumeras and Way (2017, p.

30) provide the following example.

(i) ‘Est-ce que c’est ta cousine?’ ‘Non, je n’ai pas de cousine.’

The cousin who is talked about in this conversation is female, which is clearly marked by a

female form of the word ‘cousine’ in French. However, as the English language does not

have lexical items to distinguish male and female cousins, an MT engine is most likely to

render the French sentences like (ii).

(ii) “Is that your cousin?” “No, I don’t have a cousin.”

On the other hand, if the conversation is to be translated by a human translator, the translator

will infer the gender of the cousin from the other parts of the text, or from their world

knowledge about the text, and may produce a text like (iii).

(iii) “Is that girl your cousin?” “No, she’s not my cousin.”

A sentence like (ii) is called a ‘syntactic translation’ and a sentence like (iii) is called a

‘pragmatic translation’. On the sentence level, (ii) is correct, but is not the optimal translation

on the text level.

MT translates a text which consists of more than one sentence through a production model

that Lumeras and Way (2017, p. 30), following Kay (2014), call the Syntactic Model of

Translation. The model bases itself on the idea that ‘a long translation is a sequence of

short(er) translations, we memorize short translations (lexical items), and these short

translations can be reordered’ (p. 30). This model is obviously incongruent with the notion of

text we saw above.

The adherence to sentence-level text production is evident in MTPE training, too. A piece of

training material of one of the major localisation companies (SDL, 2017, p. 32) advises:

1. Read the source segment first then the MT output.

2. Determine the usable elements (single words and phrases) and make them the basis for

your translation.

3. Build from the MT output and use every part of the MT output that can speed up your

work.

The guidelines then suggest that all grammatical and terminological errors be corrected

whereas corrections of stylistic errors are optional. After editing MT outputs to the end of the

text in this way, the post-editor is instructed to run the automatic quality check function

available in the CAT tool, which detects spelling, grammar and terminological errors. There

are, however, no instructions in these guidelines to check text-level errors (cross-sentential

checks). This indicates that quality assessment on the text level is not expected in MTPE. In

the same vein, the post-editing guidelines published by the Translation Automation User

Society (TAUS) (Massardo et al., 2016, p. 17) encourage post-editors, while verifying edits,

to ‘[e]nsure that no information has been accidentally added or omitted [to the source

segment]’. The reason behind these instructions is easy to deduce: trying to assess the quality at text

level, which may necessitate additional cross-sentential translation strategies, will make the

exercise complex and time-consuming, which defeats the purpose of MTPE, i.e. fast

turnaround. This sentence-focused process in MTPE tends to cause text-level problems just like

translation using TMs does. Čulo, Gutermuth, Hansen-Schirra and Nitzke, (2014, p. 208)

report a case of English-German post-editing. Killer nurse receives four life sentences. Hospital nurse C.N. was imprisoned for life today for the killing of four of his patients. (Source Text)

Killer-Krankenschwester zu viermal lebenslanger Haft verurteilt. Der Krankenpfleger

C.N. wurde heute auf Lebenszeit eingesperrt für die Tötung von vier seiner Patienten.

(Post-edited text). ‘Killer woman-nurse to four times life-long imprisonment sentenced. The man-nurse C.N. was today for lifetime imprisoned for the killing of four of his patients.’ (Back

Translation). In the first sentence, which is a newspaper headline, the word ‘nurse’ in the English source text was machine translated as ‘Krankenschwester’ (female nurse), which the post-editor

accepted. In the following sentence, however, the post-editor edited the same word to

‘Krankenpfleger’ (male nurse) as the gender of the nurse was clarified by the pronoun ‘seiner’

(=his) in the same sentence. The inconsistency across the two sentences is not fixed in this

example, leaving the text incoherent at the text level. Čulo et al. (2014, p. 212) conclude that

the post-editing task required the post-editors to use the translation strategy of ‘explicitation’.

4.3 Justification from the Industry

We have seen that the translator’s ability to produce high-quality translation text tends to be

affected by the use of both TM and MT. These tools turn the process of translation into a

mechanical task of producing and assessing translation sentence by sentence, which can be

considered to be side effects of the technologies. When the phenomena are considered in

sociological frameworks, the side effects of TM and MT can be understood as a burden

imposed on translators by tool makers and employers/commissioners of translators who

prioritise the operationality of technology over and above the importance of the concept of

‘text’. As a result, the translator’s pursuit of quality translation as text is subjected to ‘the

industrialization and globalization forces that demand higher productivity and speed’

(Jiménez-Crespo, 2017, p. 161). Despite the importance of the concept of text as the basis of

many new translation research and training projects, the concept is undermined in practice by

technologies such as TM and MT (Jiménez-Crespo, 2017, pp. 158-9).