mpirical human translation process research started in the 1980s (e.g. Lörscher 1991, Krings 1986) and since then has developed into a research direction which analyses, describes and models translators' behavior, using eye-tracker and keyboard logging devices (Jakobsen, 2003). While eye movements on the source text (ST) symbols give a picture of how meaning is constructed, typing behavior reflects how the target text (TT) is constructed.
This paper gives an overview of our current translation process research to describe types of translation styles, based on translators' activity data. We establish a repository of patterns of translation behavior. First we outline the experimental setting and the visualization tools used. We will then identify characteristic styles of translator behavior based on a qualitative and quantitative assessment of translators' activity data, and finally draw some conclusions that the findings might have on the technological development.
It has been reported that expert translators are less enthusiastic about automated translation assistance than student translators.
A better understanding of the human translation processes can inform the development of advanced translation assistance tools and provide a basis for more successful interaction between the human translator and the technology. It may also be instrumental for the development of tools for translator education. An empirically grounded cognitive model of translation will help not only understand translation activities, but also to design targeted translation assistance.
- Experimental setting
We analyze translation experiments which included 12 professional translators with at least two years' experience, and 12 MA students at the Copenhagen Business School (CBS), all of them specializing in translation between Danish (L1) and English (L2).
Three texts were presented to the participants in the keystroke logging program Translog User (Jakobsen and Schou 1999), which displays the source text (ST) in the top window of the screen, and enables the translator to produce the target text (TT) in the bottom window. The participants did not have Internet access and were not allowed to use dictionaries or other similar support, since this would produce a large amount of irrelevant data (for our purpose), and distort the timings and process data that we were interested in.
Process data consisted of eye-tracking and keystroke logging data from all the texts. In addition to the process data, the experiments generated product data in the form of translated output from all participants. All target texts were manually aligned with their source text at word or phrase level.
The combination of aligned source and target texts with eye movement and keystroke logging data enabled the presentation of translation progression graphs (Figures 1 to 4) showing relations between translation product and translation process data in time (Carl and Jakobsen, 2010). Thus, we were able to see, for instance, whether the translator was looking at a ST word while producing a translation of it, whether he/she was looking at another ST word somewhere else in the source text, and so on.
The translation progression graph in Figure 1 shows a fragment of 700 seconds in which an English ST of 160 words was translated into Danish. The graph shows the distribution of ST fixations on the 160 ST words and the keystrokes by which the TT was produced. Blue circles represent fixations on the ST, black dots TT insertions, and red dots TT deletions. Note that there are longer stretches of time with no gaze activities (i.e. no blue circles). These are likely to be times when the translator was looking at the keyboard or reading the target text (TT). The translation progression graphs only show reading behavior on the ST, since the software used in the experiment only registered and mapped gaze movements on the source screen. Due to the fact that some TT words could not be aligned to any ST word, there are a number of unaligned keystrokes. There are also unmatched keystrokes, which could not be associated with any of the ST words.
Figure 1 shows several translation phases: initial orientation, drafting and revision. In this paper we characterize translators according to (1) how they initially orient themselves in the ST, (2) how they plan translation drafting, and (3) whether they prefer online revision or end revision. The combination of a translator's behavioral characteristics in these three phases may serve to formulate different translator styles.
- Initial orientation
Initial orientation in the ST involves the translator's reading of the ST before starting to produce the translation. We knew from previous studies that the extent of initial orientation may vary. Some translators prefer to systematically read the whole source text before they start translating, some translators skim the text very briefly, and some translators just read the first couple of phrases or sentences before starting to type, or they simply go straight ahead with target text production with hardly any preliminary ST reading.
Initial orientation may be calculated as the amount of time spent before writing the first character1. On average, the 24 translators spent 3.3 per cent of total task time on initial orientation across the three texts, and most of them spent less than 10 seconds (in all tasks), which includes the reading required for producing the first translation segment of target text. In general, translators have a tendency for the initial orientation phase to be of approximately the same duration across all three tasks, but substantial individual differences were observed. In particular, three of the translators spent a long time on initial orientation before they started translating the text. The visualization provided by the translation progression graphs revealed that these three translators systematically read through the whole source text before starting to translate it. This type of behavior is exemplified by the translation progression graph in Figure 1, where the continuous string of blue circles starting at zero ms (bottom-left corner) and ending around 40,000 ms (top-left corner), just before the first key is pressed around 40,000ms, indicates systematic reading of the ST before typing was started.
Figure 1: The translation progression graph shows time in ms on the X-axis, and the Y-axis shows the source text (from 0 to approx. 160 words). The small dots in the graph indicate fixations, and the lines between them are eye movements (saccades). The black and red characters are the TT keystrokes.
Only three translators carried out systematic reading during initial orientation. In the majority of cases, the translator either started translating right away (head-start) or read the first couple of words or sentences, and then pressed the first key (quick planning). Some translators skimmed the text rapidly (skimming). Generally, the initial orientation phase seems to be oriented towards limited context and not the whole text.
- Reading during drafting
Translators also differ with respect to ST reading and planning while translating. We looked at where in the ST the translators' eyes were fixated while translating a given word or phrase. First of all, a distinction can be made between fixations on words which were about to be translated (looking ahead), and fixations on ST words that had already been translated (looking back).
- Looking ahead
When people read aloud, there is a lag, termed the eye-voice-span by analogy with the celebrated eye-mind span (Just and Carpenter 1980). The eye-voice span is a measure of how far the eyes are ahead of what is articulated at any point in time. If the lights in a room are turned off and the eyes can no longer see what is being read, we are still able to produce two or three more words (Staub and Rayner 2007: 329). Similarly, in a simple copying task, where subjects were asked to rewrite a text in the same language, the copyist typically looked two to three words to the right2 of the word being retyped (John 1996).
Figure 2: Large context planning: source text fixations ahead of the word currently being translated
The average look-ahead for our translators was around four words (mean: 4.28, median: 3.90), suggesting that translators generally look for slightly more context than typists performing a simple copying task. Most ST fixations were in the area between 2 and 6 words to the right of the word being translated. Thus all translators in the study shared a tendency to look beyond the current word. A certain amount of forward planning is a general feature of the translation process. This does not imply that the translators never look at a ST word simultaneously with producing an equivalent in the TL, but all translators had most fixations to the right of the word being translated.
We found some variance in individual planning behavior across the three tasks, but standard deviations were generally below or just around 1. The level of text difficulty did not appear to have an effect on planning behavior, signified by the lack of a clear correlation between text and the number of forward words viewed. Translators with a tendency to look far ahead in the text did so fairly consistently across all tasks, and translators who focused on the near environment of the word or phrase being translated similarly exhibited this behavior regularly. Online planning behavior can thus be divided into two different categories: large-context planning and small-context planning.
In large-context planning, translators are oriented in a broad context, with fixations sometimes far to the right of the word or phrase which is currently being translated. For instance, the large context planners may attend to a sentence further down in the text or read long stretches of text, sometimes several sentences, immediately following the current position, as in Figure 2.
Some of the fixations far ahead in the source text may be random. In the experimental setup we used, translators frequently had to move their eyes from the bottom window of the screen, where the translation was typed, to the top window of the screen, where the source text was displayed. Therefore the eyes may sometimes incidentally have 'stumbled' when moving up or down on the screen, leaving a single or a couple of fixations at a random place in the text.
There turned out to be a strong correlation between the qualitative analyses of the graphs and the quantitative measure of the number of words read ahead. Translators identified qualitatively as large context planners had an average read-ahead of 5 words or more (on average for all three texts), with the exception of two translators who looked only 4.7 and 4.2 words ahead on average. Ten translators were categorized as large-context planners.
Figure 3: Small context planner, looking only a few words ahead (with some random regressive/progressive fixations.
It is characteristic of small-context planning that the translators focus their gaze on a small context of no more than a few words ahead of the word being translated. An example is given in Figure 3. Even if small context planners may seem to have some leaps ahead, these are not very systematic, and the translator does not read longer sequences further ahead in the text. Rather, a major part of the fixations are on or very close to the word currently being translated. Here the fixations (blue circles) frequently overlap with the keystrokes, which indicates that ST fixation and typing of TT equivalent occur simultaneously.
Again, we found a strong correlation between translators classified (qualitatively) as small context planners and the quantitative measure of read-ahead in that all of the 12 translators identified as small context planners had read-ahead measures below 5.
We will now turn to fixations on ST words which have been translated, i.e. the translator is in the process of translating a word or phrase at one place in the text, but is looking at words which have been translated previously.
Looking back may be triggered by TT monitoring where the translator reads his or her translated output and compares it with the ST input to check that the translation was correct, or it may be associated with the translator searching for clues in previous ST words or sentences in order to be able to continue production, i.e. they attempt to understand and reproduce words currently being translated by revisiting previously translated text. Our classification of translator behavior as backtracking or non-backtracking is concerned mainly with the latter behavioral feature of consistently rereading sequences of previously translated ST, because it signals a broader orientation perspective on the translation.
Although the data showed that all translators had some regressions to previously translated ST words, there was no indication of this behavior generally characterizing the translation process. Six of the 24 translators consistently, i.e. across all three tasks, looked back in the ST whereas the remainder only sporadically and fairly randomly, it seems, reread previously translated ST words.
Figure 4: Source text fixations looking back from the word being translated.
There is a correlation between the qualitative data on which the classification was based, and the quantitative measure of fixations looking back from the word being translated. For translators who exhibit backtracking behavior, the mean number of words backtracked is consistently above 1.5, whereas for the rest of the translators, it is consistently below.
Revision/editing activities can be divided into two broad categories: online revisions, i.e. revisions carried out during the drafting phase; and end revisions, carried out after the drafting phase.
Online revisions were calculated as the number of text elimination keystrokes during the drafting phase, irrespective of whether these were associated with the correction of typos, rephrasing of words, phrases and sentences, or with change of word order, etc.3 All translators made corrections while drafting, but the amount of online editing varied from an average number of 27 text elimination keystrokes to an average of 134.
Again, for many of the translators there was a tendency for the number of revisions to remain fairly consistent across the three tasks. For instance the translator with the lowest average had 41, 27 and 14 deletion keystrokes respectively in the three texts, whereas the translator with the highest average made 108, 164 and 129 eliminations respectively. However, we also found some variation across the tasks.
As for end revision, only some of the translators made corrections after the drafting phase. Eight translators had no text elimination keystrokes in any of the three tasks, and most translators consistently made few deletions, i.e. in the range from 5-10, after the drafting phase. Even so, some of the translators made an effort after the drafting phase, signaled by time and gaze activity, to review their own target text. In some cases no or few corrections were in fact made, resulting in the low number of text elimination keystrokes after the drafting phase in some translators. Thus, it seemed that a better measure of end revision would be to measure the time spent after the drafting phase (end revision) relative to the time spent during drafting as described below.
We find a clear preference among the translators for allocating more time to the drafting phase than to the end revision phase. Yet, eight translators spent a considerable 20 per cent or more of the overall translation time on end revision. Not surprisingly, the same translators had a fair amount of text elimination keystrokes during the end revision phase (range 4-34, mean 12.7). Somewhat unexpectedly, however, four of the eight end revisers also had a large number of online text elimination keystrokes, i.e. higher than the average of 68, and their behavior could thus better be classified as constant revision behavior, whereas the other translators in this group exhibited end revision behavior. Behavior in translators who spent less than 20 per cent of their time on end revision could be classified as online revision behavior.
Based on empirical data, we have observed differences and similarities in the translators' working styles, and classified translation types according to the observable behavior. We have identified three translation phases: initial orientation, translation drafting and final revision.
- Initial orientation: functions as an initial text planning phase. Several types of behavior can be distinguished:
- systematic initial orientation: the translator systematically reads through the ST before translation.
- skimming: the translator skims the ST rapidly before translation
- quick planning: the translator reads the first couple of words or sentences, and then presses the first key.
- head start: the translator starts translating right away
- Translation drafting: the phase in which the actual translation is drafted. We distinguish several translation styles:
- large-context planning: the translator reads text sequences, sometimes whole sentences, far ahead in the source text.
- small-context planning: the translator frequently fixates the word being typed or a couple of words, but rarely a whole sentence ahead.
- backtracking: the translator has a tendency to re-fixate ST words which have already been translated
- non-backtracking: the translator does not systematically re-fixate ST words which have already been translated
The translators may show traces of different kinds of behavior during drafting, but the data provide evidence for an overall preference for one of the two kinds of planning ahead (small context or large context planning) as well as a preference with respect to looking back at previously translated ST words. The two types of planning behavior may or may not be combined with backtracking.
- Revision: this phase serves to review the text and refine translation choices. Three types of revision can be distinguished:
- online revision: the translator revises the text during the drafting phase
- end revision: the translator spends 20 per cent or more of his/her time on end revision
- constant revision: the translator spends more than 20 per cent of translation time on end revision, but at the same time makes a large number (above average) of online revisions.
Development of translation tools could benefit from incorporating knowledge of human translation behavior and translator styles.
Given the separation into translation phases, and their different function in the translation process, it might be helpful to design separate tools that support translators in their specific need during these phases. Repeated expressions, term translation, or passages that are difficult to translate may be of interest in the initial orientation phase and tools could be designed to highlight these occurrences during this phase.
According to behavioral preferences in the translation phases, different online assistance tools might help translators who work, for instance, as small-, or large-context planning. Translation completion tools might be more helpful for small-context planning, whereas translation memories or machine translation post-editing might be more acceptable for large-context planning.
Previous research has shown that professional translators and novices generally exhibit different translation behavior (Jäaskelainen 1988, Jensen 2001, Jakobsen 2003, Tirkkonen-Condit 2005). Preliminary investigations suggest that professional translators tend to be characterized by head-starting, small-context planning, and end-revising, while student translators more frequently perform systematic initial orientation, large-context planning and online revising (Carl and Buch-Kromann, 2010). Future studies will investigate this preliminary finding in more detail.
Also, it has been reported that expert translators are less enthusiastic about automated translation assistance than student translators (Koehn, 2009). It is unclear why this might be the case, and how translation assistance could be designed to help all translators alike, but developing customized translation tools taking account of different translation styles seems could be a direction to pursue.
Better cognitive models of translation have to be developed, since experiments to design and evaluate such CAT tools are extremely labor intensive to develop and to test. As Knight et al. (2007) point out, "the combination of small usability studies and cognitive modeling [may help to] make an informed decision about critical aspects of a User Interface".
Carl, Michael, and Jakobsen, Arnt Lykke, 2010. Towards statistical modeling of translators' activity data, International Journal of Speech Technology 12(4), 124-146
Carl, M & Buch-Kromann, M 2010, Correlating Translation Product and Translation Process Data of Professional and Student Translators, In Proceedings of EAMT, Saint-Raphaël, France
Dragsted, B. 2010. Coordination of reading and writing processes in translation: an eye on uncharted territory. In G. Shreve and E. Angelone (eds.) Translation and Cognition. Amsterdam/Philadelphia: John Benjamins. 41-62.
Knight, Andreas, Guy Pyzark and Collin Green 2007. When Two Methods are Better Than One: Combining user study with cognitive modeling. Proceedings of CHI, San Jose, 1783-1788
Koehn, Philipp, and Haddow, Barry, 2009. Interactive Assistance to Human Translators using Statistical Machine Translation Methods, In Proceedings of MT Summit
Loescher, W. 1991. Translation performance, translation process and translation strategies. Tuebingen: Guten Narr.
Jäaskelainen, R. 1988. Automatised processes in professional vs. non-professional translation: a think-aloud protocol study. In Künzli, A (ed) Empirical research into translation and interpreting: processes and products. Neuchâtel: Université de Neuchâtel. 89-109.
Jakobsen, A.L. 2003. Effects of think aloud on translation speed, revision and segmentation. In Triangulating translation. John Benjamins, Alves (ed) 69-95.
Jakobsen, A. L. and Schou, L. 1999. Logging target text production with Translog, In Copenhagen Studies in Language, volume 24, Samfundslitteratur, Copenhagen, 9-20
Jensen, Astrid. 2001. The Effects of Time on Cognitive Processes and Strategies in Translation [Copenhagen Working Papers 2]. Copenhagen: Copenhagen Working Papers in LSP.
John, B. E. 1996. TYPIST: a theory of performance in skilled typing. Human-Computer Interaction. 11(4).
Just, M. A. and Carpenter, P. 1980. A Theory of Reading: From Eye Fixations to Comprehension. Psychological Review, 85. 109-130.
Krings, H.P. 1986. Translation problems and translation strategies of advanced German learners of French. In J. House, & S. Blum-Kulka (Eds.), Interlingual and intercultural communication (pp. 263-75). Tubingen: Gunter Narr.
Staub, A. and Rayner, K. 2007. Eye movements and on-line comprehension processes. In Gaskell, G. (ed.), The Oxford Handbook of Psycholinguistics. Oxford: Oxford University Press. 327-342.
Tirkkonen-Condit, S. 2005. The monitor model revisited: evidence from process research. In Lee-Jahnke, H. (ed) Processus et cheminements en traduction et interprétation. Montreal: Presses de l'Université de Montréal. 405-414.
1 This is not a perfectly accurate calculation. The first keystroke is likely to be preceded by a short period of 'local' preparation that should ideally count as part of the drafting phase.
2 This applies to languages which are read from left to right. In languages such as Hebrew that are read right to left, fixation spans are to the left of the word (Staub and Rayner, 2007)
3 See Carl et al. (2010) for a discussion on short and long distance revisions of the same text material.