Cognitive processes in translation

The word “translation” in its broadest sense is to change something written or spoken into another language (Christoffels & de Groot, 2009). Cognitively, translation involves “comprehension” of the source language (SL), “code-switching” between the source and the target language (TL), and “output production” in the target language, each stage including several processes (Macizo and Bajo, 2004). Based on theories and discoveries by previous researchers, Macizo and Bajo (2004) proposed two possible models for the three sets of processes. They argued that Seleskovitch’s

“deverbalization” theory (1976) basically looked at the process of translation from a

“vertical perspective”. In the vertical model, the recoding stage takes place only after comprehension is finished, and the linguistic form in the SL is lost. Only the message of the original discourse is retained and restructured according to the grammar of the TL. In other words, there’s no direct links between the two languages “at the lexica/

syntactic levels” (p.182). (See Figure 2-3.) The other model takes the “horizontal view”. Translators have partially started reformulating some of the lexicons and syntactical structures for semantic matches in the TL when comprehending the SL.

That is to say, there is “direct processes of recoding from one linguistic code to another” (p.182). (See Figure 2-4.)

25

Figure 2-3. Sequence of processes involved in translation: Vertical approach (Macizo & Bajo,

2004) Abbreviation: SL (Source Language); TL (Target Language)

Figure 2-4. Sequence of processes involved in translation: Horizontal approach (Macizo & Bajo,

2004) Abbreviation: SL (Source Language); TL (Target Language)

26

Macizo and Bajo (2004) believed that comprehension in both normal reading and in translation included the same components, from speech processing, lexical access, sentential process and discourse processing. In the vertical model, since the

deverbalized message is only reconstructed after the comprehension stage, there should be no difference between normal reading and reading for translation. However, in the horizontal model, where part of the formulation takes place during SL reading, additional cognitive effort would need to be put into normal comprehension.

Eventually, the horizontal view was supported by their experiment, where

professional interpreters had been asked to read and repeat sentences and then to read and translation sentences. It was found that more time was spent on sentence reading before utterance during the task of reading and translation.

Some researchers referred to vertical and horizontal view as “a meaning-based strategy” and “a transcoding strategy” (Christoffels & de Groot, 2009). The view was also adopted by Huang (2011) and Chen (2013) when analyzing their data. However, transcoding is viewed negatively by some interpreters and/or researchers, believing that it should be limited to items with fixed correspondences such as proper names, numbers and specialized terms (Pöchhacker, 2004). In Macizo and Bajo’s model (2004), vertical translation stays at the lexical and syntactical level. However, if transcoding is merely replacing words with their translation equivalents, commonly called word-for-word translation, the strategy would render unintelligible

interpretation (Christoffels & de Groot, 2009). The proposal by Huang (2011) and Chen (2013) that novices’ adopted a vertical approach to interpreting, and that experienced interpreters a horizontal approach would be in conflict with the quality assessment of their performance (Chen, 2013).

The view on transcoding, in fact, should be re-examined. Paradis (1994) proposed that transcoding can take place at different levels of the language system,

27

not only at phonological, morphological and syntactical level but also at semantic level. Paradis (1994) argued that rules under transcoding has to be learned, so supposedly more experienced interpreters’ translation could better manage it while novices’ translation is usually meaning-based. Moreover, transcoding and

meaning-based interpreting strategies are not mutually exclusive; “both strategies can be available to the experienced interpreters” (Christoffels & de Groot, 2009, p.459).

Based on Paradis (1994), Christoffels & de Groot (2009) created a new graph for the two alternative but co-existing strategies in the process of translation. (See Figure 2-5.)

Figure 2-5. Two alternative but co-existing interpreting strategies (Christoffels & de Groot,

2009).

28

2.5.2 Cognitive processes in interpreting

In a narrower sense, “translation” only refers to text-to-text translation, while verbally rephrasing something original in another language is called “interpreting” (de Groot, 2000; Christoffels & de Groot, 2009). According to Moser-Mercer (2010), Interpreting also involves language comprehension, reformulation (termed “transfer mechanism”), and production. However, in interpreting, the auditory input is transient and irretrievable, with its rate determined by the speaker. In translation, the source text is always available so the translator could read and reread at his or her preferred rate. The differences in input mean differences in the cognitive process of written translation and interpreting (Christoffels & de Groot, 2009). In Gile’s effort model (1995), the listening and analysis effort in interpreting consists of all

“comprehension-oriented operations” (p. 162), while the production effort includes both lexical and syntactical reformulation and the final oral output production.

However, in addition to comprehension, reformulation and output production, which exists in written translation as well, interpreting also involves short-term memory effort. Since the auditory information is not constantly available, in the case of both consecutive interpreting (CI) and simultaneous interpreting (SI), interpreters often have to store the heard information for later use.

In CI, the interpreter first listens to a section of speech and does not start oral output the translation until the speaker pauses for interpreting. In Gile’s model, CI consists of two phases: the listening and note-taking phase, and the speech production phase. Memory is required in both phase one and phase two. In phase one, since note-taking takes more time than speech production, there is a time lag between the speakers’ speech and the notes jotted down by the interpreter. The lag creates demand on short-term memory. In phase two, the interpreter needs to recall what was said in the original speech, with the help the notes, to produce oral translation.

29

SI also requires memory effort. Despite its name, simultaneous interpreting is not really simultaneous in that there is a time lag or ear-voice span (EVS) between a segment uttered in the original speech and its oral translation by the interpreter. The lag is considered to reflect the time involved in processing the incoming information, which is stored in the interpreter’s short-term memory before uttered (Anderson, 1979).

Since many of the above discussed efforts are taken at the same time, they need to be coordinated. The first phase of CI calls for coordination effort as the interpreter has to take-notes while listening and analyzing the original speech. In SI, there’s also the coordination effort (Gile, 1995), as an interpreters not only has to comprehend and store input segment in the source langue, but also transform an earlier segment from source to target language, while producing an even earlier segment in the target language at the same time (Christoffels & de Groot, 2004). During almost 70% of the time, SI interpreters are talking and processing the input at the same time (Chernov, 1994). That is to say, during the time, all the above mentioned tasks have to be

coordinated to be performed together, which required significant cognitive control and attention (Moser-Mercer, 2010). Christoffels & de Groot (2004) concluded from their experiments that it was the combination of both the simultaneity (of language

comprehension and production) and the language transformation that caused a significant drop in performance, suggesting high cognitive demand from the combination.

2.5.3 Studying interpreting process

Citing Klausy (2004), Bakti (2009) stated that SI might be viewed as “a

psycholinguistic experiment designed to test how the processes of speech production, speech perception, translation and monitoring operate simultaneously” (p. 1). To cope

30

with the distinct nature and the high cognitive demand of SI, interpreters gradually develop strategies, or procedural knowledge, that may be employed consciously or unconsciously (Riccardi, 2005; Shlesinger, 2000). Previous findings supported the argument that supplementary skills have been developed by interpreters so that they could gain more control over their interpreting (Ericsson, 2000). What are these

“mechanisms and mental representations” (Ericsson, 2000, p.246) involved in the process of interpreting? Many researchers attempted to find the answers, but there aren’t many tools available (Tiselus & Jenset, 2011).

The Think Aloud Protocol (TAP) is one commonly-used methodology for

studying someone’s cognitive process when performing a task. It has been established that representative tasks in skilled performance can be reproduced in a lab setting (Ericsson & Lehmann, 1996). Participants are instructed to verbalize their thoughts while focusing on their tasks (Ericsson & Simon, 1998). However, TAP is not a viable tool for interpreting study (Ivanova, 2000; Shlesinger, 2000) since it is apparently not possible for interpreters to verbalize their mental processes while producing oral output.

Another approach following the laboratory tradition of cognitive psychology is to break down the complex processes involved in interpreting into small components of comprehension or production, such as word recognition and categorization

(Ericsson, 2000). However, Shelesinger (2000) challenged the approach by asking “is the process decomposable”. Individually studying independent subcomponents doesn’t necessarily tell us how they work in coordination in a complex task like SI. A case in point would be the research of Köpke & Nespoulous (2006), where novice and experienced interpreters were asked to perform memory tasks such as digit or word recall. It was discovered that in nearly all the tasks, the novices performed better than the experienced interpreters. While the researchers concluded that “working memory

31

capacity is more likely to develop in novices who are struggling with a new task”

(p.16), the finding raises doubt as to whether it is appropriate to isolate memory from SI and to reduce it to digit and word storage and recall, when SI is a complex task of processing information (Anderson, 1979) from meaningful, contextualized materials (Shelesinger, 2000).

One of the earliest approaches to interpreting processes is introspective analysis.

Researchers, who are often practicing interpreters, analyze cognitive processes in interpreting based on their observation and insights, giving a more general, holistic description (Ericsson, 2000; Shlesinger, 2011). Running the risk of being subjective and intuitive, the approach could only be a starting point for formulating hypothesis rather than testing them (Shlesinger, 2011). A similar approach is often taken in product-based research. For example, Agrifoglio (2004) asked 6 professional

interpreters to perform the task of ST, SI and CI. Based on the observable errors and difficulties in the oral output, she suggested possible reasons for the errors and

difficulties and possible cognitive efforts required during the interpreting process. The conclusions were solely inferred from the researcher’s observation of the end-product and her insights into the three tasks.

How could researchers really probe into interpreters’ mind rather than explaining observed phenomenon based on their own understanding of interpreting processes?

Although verbalization of thoughts is not possible during interpreting, doing it in retrospection is a possibility. There are two approaches to retrospection. In Bartlomiejczyk (2006) and Ivanova (1999), interpreters were asked to recall their thoughts during a task after the entire task was finished. A transcript of their output was provided to the interpreter to facilitate the recall. In other retrospection studies, behaviors noted during the task and/or errors occurred in the end-product were used by researchers as prompts (Ivanova, 2000). Alternatively, immediate retrospection

32

may be employed to create “frozen time” (Tiselius & Jenset, 2011). The approach could only be taken in an experimental setting as the researcher would interrupt the interpreters during the course of interpreting to ask them about the reasoning behind what they just did before the interruption (Shlesinger, 2000). In Mead (2000), immediately after each consecutive interpreting, recordings were played back to the subjects, who’d be asked why they paused at a certain time in their interpreting delivery.

In immediate retrospection, the interval between the actual processing and the retrospection is shorter. However, even in immediate retrospection, where lack of memory could be argued to play an insignificant role, awareness of the processing in the first place is still an issue. Verbal descriptions and explanations given by expert and other subjects are often inconsistent with careful observation of their actual behavior. Automaticity of basic skills is a key feature in expertise (Feltovich et al., 2006; Leighton, 2009; Moser-Merser, 2010). Similarly, Moser-Merser (2010) called the final stage in interpreting expertise development “the autonomous stage,” where many of the cognitive processes required have been internalized (Shelesinger, 2000) and occur “to a great extent beyond awareness” (Riccardi, 2005). If the processes are so automated that the interpreters are not aware of them, it’s not likely they would point them out in retrospection. A relevant methodological issue is that instead of recalling, some interpreters may attempt to reconstruct the processes based on

previous experience (Ivanova, 2000). In other words, what they report in retrospection is not “what happened” but “what should have happened.”

The key issue with either introspective or retrospective research of interpreting is that the only available data is the recorded output, the end-product. The intangible processing that produced the output is lost once the task is finished, so researchers have no choice but to rely on the subjects’ memory and account and the researchers’

33

own judgment and experience for inference. If researchers would like to avoid being subjective and take a further step in understanding interpreters’ mind, apparently they need more than output data. Agrifoglio (2004) for example, judged from the false starts in two of her subjects’ Spanish output in an ST task that they “had trouble locating the main noun (flows) within the long English noun clause” (p.54). It’s a possible explanation, but cannot by fully justified without knowing whether the interpreters had already read the main noun at the end when they started orally outputting the clause. Another set of data, one that could at least reflect cognitive process, is needed. This is where eye-movement data comes in.

2.6 Eye movement

When light from external objects enters human eyes, it travels through the structure of the eyes to fall on the retina, a light-sensitive layer of tissue. The visual stimuli are eventually sent from the retina to our brain to be processed so that we “see”

the objects. The brain recognizes the perceived items, symbols or words before making further responses. The retina consists of several areas, among which the central area, or fovea, is the most sensitive to light. Since visual acuity is the highest in the fovea, our eyes have to move constantly so that external light from objects could fall onto the area. (蔡介立、顏妙璇、汪勁安, 2005).

Eye-tracking measures the motion of the eyes. It enjoys wide application in psychology studies on cognitive processes during reading, scene perception, and visual search. In the studies, eye movement is studied as a measure of overt attention (Duchowski, 2002; Rayner, 2009). It has been established that eye movements reflect cognitive processes (Rayner, 1977).

34

2.6.1 Eye movement in reading

According to Radach & Kennedy (2004), “many basic facts about eye

movements in general have been established in the context of reading research” (p.3).

In reading, as in various other tasks, eyes move constantly and fast in a coordinated fashion, and the movement is called “saccade”. Saccades are interrupted by

“fixations”, when the eyes remain relatively stable to receive visual information (Radach & Kennedy, 2004; Rayner, 2009). No new information is acquired during saccades, but cognitive processing does continue (Rayner, 2009).

Studies that use eye-tracking techniques to study cognitive process during reading are done on the basis of two assumptions proposed by Just and Carpenter (1980): the immediacy assumption and the eye-mind assumption. That is to say, a reader starts to process a word as soon as it is fixated, and the eyes would stay on the word for as long as the time it takes to be processed. In other words, eye-tracking produces moment-to-moment data that reflects the readers’ cognitive activity during the time of reading.

The average eye fixation duration when reading English is about 200 to 250 ms, and the mean saccade size is 7 to 9 letter spaces (Rayner, 1998). In Chinese reading, the average fixation duration is 220-230 ms, with an average saccade length of 2.5 to 3.5 characters (蔡介立、顏妙璇、汪勁安, 2005). Previous researchers have found that in English reading, fixation time on a word is influenced by lexical and linguistic variables that affect the difficulty of processing the word, such as word frequency, predictability, number of meaning, age of acquisition, semantic relations between the fixated words and prior word, and word familiarity (Rayner, 2009). It has been discovered that when reading conceptually more difficult text (Rayner, 1998) or syntactically ambiguous sentences, readers’ fixation duration increases and saccade length decreases (Rayner, 2009). Typographical factors such as quality of print, line

35

length, and letter spacing also influence eye movements (Rayner, 1998). Regressions (backward saccades to a previous position in the text) are more frequent when the text is more difficult or when typographical issues, such as font difficulty, exist (Rayner, 2009). Reading skills also influence eye movement. Less skillful readers tend to have longer fixations, shorter saccades and more regressions (Rayner, 2009). In addition, compared with silent reading, mean fixation durations are longer when a piece of text is read out (Rayner, 1998).

2.6.2 Eye movement in interpreting

McDonald and Carpenter (1981) were among the first researchers to use eye-tracking to study interpreting. They invited two expert and two amateur translators for a sight-translation task from English to German. The texts to be translated were manipulated so that the translators would be prompted to understand and translate English idioms in their literal sense. The aim of the experiment was to learn more about the translators’ eye-movement when the text was read, translated and when the errors were discovered and corrected. Form the different fixation patterns in literal and idiomatic translation, it was discovered that when reading a piece of text the first time, the translators looked for a meaningful unit in English (“parsing” or

“chunking” the words) to be translated into German. The reformulation and oral output only began when the translator went back (or regressed) to read the unit the second time (or in its second reading pass). When there was discrepancy found between the newly read information and the old material, they regressed to the ambiguous word to reparse the phrases and retranslate it. Based on their findings, the researchers proposed a model with three components: (1) a reading and parsing process; (2) an error recovery mechanism for reparsing; (3) a translation process where the reparsed unit is reformulated and ouputed.

36

Thirty years later, from 2011 to 2013, three related eye movement studies were conducted to better understand interpreters’ cognitive processing during sight translation. Huang (2011) conducted an experiment to collect and compare novice interpreters’ eye movement data in silent reading, reading aloud and sight-translating.

The 18 novice participants were students of graduate interpreting and translating programs in Taiwan, who had finished sight translation courses in the first year of their program. She found that the first-pass reading in reading aloud took longer than it did in silent reading and sight translation, and no difference was found between the first-pass reading time in silent reading and in reading aloud. She argued that since reading time reflects the cognitive demand required, “the processing efforts of silent and sight translation proved to be similar in the first pass” (p.55). She further

suggested the results supported the vertical view of translation process, as opposed to the horizontal view backed by the findings of Macizo and Bajo (2004).

Chen (2013) reproduced Huang’s experiment (2011), this time targeting

experienced interpreters, who had at least 150 days of work experience. By comparing the eye movement of the novices (Huang, 2011) and the experienced interpreters, she found that compared with silent reading, the experienced interpreters spent more time reading in the first pass, indicating that they were not merely reading but took

additional effort for something else. Significant difference was also found in the first pass reading time between the novice and the experienced interpreters: it was longer

additional effort for something else. Significant difference was also found in the first pass reading time between the novice and the experienced interpreters: it was longer