Translator-computer interaction in action — an observational process study of computer-aided translation

The Journal of Specialised Translation Issue 25 – January 2016
106
Translator-computer interaction in action — an observational
process study of computer-aided translation
Kristine Bundgaard, Tina Paulsen Christensen and Anne Schjoldager, Aarhus
University
ABSTRACT
Though we lack empirically-based knowledge of the impact of computer-aided translation
(CAT) tools on translation processes, it is generally agreed that all professional
translators are now involved in some kind of translator-computer interaction (TCI), using
O’Brien’s (2012) term. Taking a TCI perspective, this paper investigates the relationship
between machines and humans in the field of translation, analysing a CAT process in
which machine-translation (MT) technology was integrated into a translation-memory
(TM) suite. After a review of empirical research into the impact of CAT tools on
translation processes, we report on an observational study of TCI processes in one
particular instance of MT-assisted TM translation in a major Danish translation service
provider (TSP). Results indicate that the CAT tool played a central role in the translation
process. In fact, the study demonstrates that the translator’s processes are both
restrained and aided by the tool. As to the restraining influence, the study shows, for
example, that the translator resists the influence of the tool by interrupting the usual
segment-by-segment method encouraged by translation technology. As to the aiding
influence, the study indicates that the tool helps the translator conform to project and
customer requirements.
KEYWORDS
Translator-computer interaction (TCI), computer-aided translation (CAT), translation
memory (TM), machine translation (MT), MT-assisted TM translation, professional
translation, translation processes, observational study.
1. Introduction
Currently, all professional translators tend to interact extensively with
computers in the course of their work (see Christensen and Schjoldager
2016), and yet we know very little about how computer-aided translation
(CAT) impacts on translation processes. In particular, we do not know
very much about the relationship between translators and machines in
actual working practices (Muñoz Martín 2014: 70). As Pym (2011: 2) puts
it, “when we ask what translators really do with translation memories and
machine translation, there is not an enormous amount of empirical data to
speak of”. However, as many practising translators will tell you, CAT tools
now play such a central role in professional translation processes that
translators can be assumed to be less in charge than they used to be,
which may mean that translators are being pushed towards the periphery
of the translation profession. Highlighting the business aspect of the
profession, Risku (2014: 336) actually goes so far as to characterise
today’s translation industry as a computer-assisted network economy.

107
In order to obtain a better understanding of how translations are produced
in the translation industry, more research on professional translators’ tool
usage is needed. In our paper, based on an observational study, we shall
look at the role of CAT tools in translation processes from a translator-
computer interaction (TCI) perspective, using a term that was coined and
introduced by O’Brien (2012) drawing on human-computer interaction
(HCI), a well-established discipline within computer science and social
sciences, especially applied psychology (e.g. Johnson 1992; Carroll 1997,
2013). Based on Olohan (2011), who uses Pickering’s “mangle of practice”
theory to analyse professional translators’ views on TM translation, TCI
may be seen as a “dance of agency”, in which a human agent (translator)
interacts with a non-human agent (the technology) in a process of
resistance and accommodation. Thus, for instance, while the CAT tool is
generally expected to aid and support the translation process, it may also
offer resistance and restrain the process in several ways. In order to
accommodate the restraining influence of the tool, translators may need
to carry out certain actions enabling the ongoing interaction between the
tool and the translator to progress.
The TCI processes that we have studied occurred in one particular
instance of CAT in which machine-translation (MT) technology was
integrated into a translation-memory (TM) suite (SDL Trados Studio
2011). The overall aim of the paper is to gain a deeper understanding of
the interaction between machines and humans during translation and to
investigate which types of TCI occurred in an actual translation process.
Furthermore, by providing illustrative examples from the data, the paper
attempts to document how the tool restrained and aided the translation
process. Thus, the study is an example of translation process research as
interaction research (cf. Risku 2014) investigating an actual CAT process
from a translator studies perspective (cf. Chesterman 2009). Inspired by
Olohan (2011) and based on the studies reviewed in section 2, below, we
assume that a translation tool may have a restraining influence on the
translation process when it seems to interfere with the translator’s
workflows and mental processes. When the tool seems to help the
translator — to meet customer requirements or improve translation
consistency, for instance — we regard this as an aiding influence.
When using a TM, the translator accesses and re-uses segmented and
paired source and target texts (text segments) stored in a database.
While translating, the translator is provided with translation proposals, so-
called matches. If the string of words of a segment from the database
matches the string of words of the source-text segment exactly, this is
referred to as a 100 % match. If the segment from the source text and
the segment from the TM also have the same “document context,” SDL
Trados Studio (the most used TM suite) refers to this as a “context
match,” meaning that the source segment and the TM segment are
preceded by the same segment. If a match is less than 100 %, it is
referred to as a “fuzzy match,” which — in principle — could be anything

108
between 99 % and 1 %, but the fuzzy-match threshold is usually set at 70
%. If no usable content is found in the TM, this is referred to as a “no
match” (SDL; see also Christensen and Schjoldager 2010). If an MT
engine is integrated into the TM suite, as in the current study, MT matches
are provided when the TM finds no matching segments, i.e. the “no
matches” are translated by the MT engine (see also section 3, below). We
shall refer to this mode of translation as MT-assisted TM translation, which
is a newly developed type of CAT (see also Flanagan and Christensen
2014; Christensen and Schjoldager 2016).
The data used in this paper to investigate one instance of MT-assisted TM
translation were generated as part of Bundgaard’s ongoing PhD project
(see Bundgaard 2013) at TextMinded Danmark A/S, the second-largest
translation service provider (TSP) in Denmark. The project is a field study
investigating how professional in-house translators revise TM and MT
matches in actual translation processes when carrying out MT-assisted TM
translation tasks. The present study investigates which types of
observable actions a single translator carried out in an experiment taking
place in the translator’s usual work environment. After a brief literature
review of relevant translation process research in section 2, we shall
present the set-up of the experiment and the research methods applied to
elicit the data analysed in the present study in section 3. In section 4, we
shall present and discuss some results of our analyses, followed by some
concluding remarks in section 5.
2. Empirical studies
In this section, we shall review translation process research on the impact
of CAT tools. Focusing on investigations into MT-assisted TM, we shall
confine ourselves to a presentation of results from selected studies which
(1) test the usefulness of CAT tools or (2) investigate the impact of CAT
on translators’ workflows and mental processes.
Quite a number of studies assess the usefulness of CAT by investigating
translation quality and productivity by means of parameters like
translation speed, the degree of post-editing and the number of errors in
the end product, that is, the impact of CAT on productivity and the
product. In particular, studies focusing on MT seem to evaluate translation
quality, typically comparing the quality of human translated texts with
post-edited MT output, or comparing the quality of TM and MT matches.
Fiederer and O’Brien (2009), for instance, compared the quality of post-
edited MT output with human translation and found that qualified
evaluators regard the quality of post-edited texts as on a par with, if not
greater than, the quality of human translations as far as clarity and
accuracy are concerned. However, as regards style, human translation is
preferred over the post-edited product by evaluators.

109
Garcia (2011), in an experiment comparing post-editing of MT-generated
text with human translation, found that productivity gains are only
marginal in connection with MT. With regard to quality, however, the
study showed that post-editing produces better results than human
translation. As for MT-assisted TM translation, Guerberof Arenas’ (2009)
experimental study suggested that professional translators have higher
productivity when using MT matches than when processing fuzzy matches
and that they make more errors in TM matches than in MT matches.
Interestingly, Guerberof Arenas’ 2012 study contradicted her former
findings when she found that there were no significant differences in
productivity or quality between TM and MT matches (Guerberof Arenas
2012). Yet the 2012 study also showed that translators worked
significantly faster when processing TM and MT matches compared to
translating from scratch and that they made significantly more errors in
segments translated from scratch than in TM and MT matches (also
reported in Guerberof Arenas 2014a, 2014b).
In an experiment with freelancers, Teixeira (2011) found that it does not
impact significantly on the overall speed and quality of the end product
whether or not translators are provided with information about the
provenance of a match during the MT-assisted TM process (i.e. whether
the match is an MT or a TM match, and at which match percentage).
Based on a field study, Federico et al. (2012) found that the post-editing
effort decreases when translators are supplied with TM matches as well as
with MT matches, and that most translators achieve substantial time
savings when they are offered MT matches in addition to TM matches.
Tatsumi (2010) investigated the editing speed and the degree of editing.
For instance, she found that it is faster to edit MT matches than to edit
75–79 % fuzzy matches.
Läubli et al. (2013) carried out an experiment with student translators
working in a realistic work environment with the aim of exploring quality
and translation speed under two conditions: In one task, translators had
access to a TM, a terminology database as well as any other translation
aid. In another task, translators were also offered MT matches. Läubli et
al. (2013) found that, by adding MT matches to other translation aids, the
translation time was reduced by 17.4 % and that the quality of
translations produced when MT is integrated equals, or in some cases, is
even better than the quality of translations produced without MT.
Relatively few studies have investigated how CAT tools impact on
workflows and mental processes (for overviews, see Christensen and
Schjoldager 2010 and Christensen 2011). Notwithstanding this lack of
empirical evidence, there is a consensus that the use of CAT tools is
changing the content and procedures by which professional translators
translate (Folaron 2010; Muñoz Martín 2014: 70). Also, the very role that
translators play in translation has changed: whereas most of their work

110
previously involved primarily linguistic transfer, professional translators
currently spend much of their time interacting with computers, besides
dealing with linguistic issues. It would be logical to expect this change to
influence translators’ mental processes too. Thus, for instance, due to the
inherent segment-by-segment method underlying most CAT tools, the
translation process tends to become more linear, as translators are invited
to work their way through the translation sequentially, rather than going
back and forth in a recursive process (Bowker and Fisher 2010: 4).
Jiménez-Crespo (2009: 233) found that this segment-based processing
leads to the replication of source-text structures (see also Dragsted
2006).
Applying processing speed as a measure of cognitive (i.e. mental) effort in
an experiment, O’Brien (2007) found that lower TM-match values require
more cognitive effort than higher match values. In an experiment carried
out by Alves and Liparini Campos (2009), it was found that using a TM
increases the number of pauses in which translators orientate themselves
during the translation process and forces translators to focus on TM
proposals as sources of information. This finding is supported by O’Brien
et al. (2010), who found that translators find the concordance feature
very useful, as it allows them to search for translation proposals, in the
form of strings of words or single words. Interestingly, this feature seems
to have a positive impact on translation quality, but a negative impact on
translators’ productivity.
Based on an experiment, Christensen and Schjoldager (2011) investigated
student-translators’ perceived impact of TM tools on translation processes.
All subjects found that TM-assisted translation differed from translating
without a TM, and most also said that they found TM translation easier,
more interesting, faster and more efficient. The subjects also argued that
TM translation is more consistent as the TM allows for easy access to
other people’s knowledge. A TM was, however, also considered a
deceptive tool as students thought they lost track of the text and the aim
of the translation, focusing too much on the source text, the sentence
level and terminology, and generally losing their critical sense.
Furthermore, subjects stressed that translating becomes more
mechanical, less personal, more fragmented and less creative.
To sum up, the above-mentioned studies seem to support our assumption,
which was also mentioned in the introduction, that translation tools have
an aiding as well as a restraining influence on the translation process.
Almost all studies reviewed in this section have been carried out in
experimental settings, and most studies of MT-assisted TM have focused
on factors such as productivity and quality in connection with MT
translation, rather than on what actually happens during the translation
process when TM and MT are integrated. Aiming at providing new insights
into CAT processes, the present study investigates which types of TCI
actions a professional translator carried out when working with an

111
authentic MT-assisted TM translation in his/her usual work environment at
Denmark’s second largest TSP, TextMinded Danmark A/S.
3. Data
At the time of data collection (2013), TextMinded Danmark A/S employed
11 in-house translators, and the company was in the process of
integrating MT into its usual TM tool. As mentioned above, the data used
for this paper were generated as part of Bundgaard’s ongoing PhD project,
in which professional translators’ revision of MT-assisted TM translation is
investigated, using a combined approach to study micro-level translation
processes (based on textual data) and macro-level translation processes
(based on contextual data). Micro-level translation processes are
investigated through an experiment with eight in-house translators, and
the macro-level translation processes are studied primarily through
ethnographic methods, namely participant observation, semi-structured
interviews and document collection. In the experiment, which we shall
focus on in this paper, data were collected using screen capture (BB
FlashBack Express), keystroke logging (Inputlog) (see Leijten and Van
Waes 2013), observation, retrospective interviews and a post-
experimental questionnaire.
In the experiment, eight translators worked at their usual desks and
computers and had access to their usual sources of information. They
were asked to translate a technical text and a marketing text, which were
authentic translation assignments from a major Danish production
company, a regular customer of TextMinded. The TM suite used was the
TM software SDL Trados Studio 2011 integrated with the MT engine SDL
BeGlobal Enterprise. The TM applied was a client-specific one, and the MT
engine had been trained with TM data on the language pair of English-
Danish as well as a client-specific termbase. In addition, the client-specific
termbase was added as a terminology management system in the TM
suite.
In this paper, we shall analyse the process regarding the technical text
translated by a state-authorised translator with more than 10 years of
professional experience. The translator had previously translated texts for
the client in question and thus had some knowledge of the client’s
preferences. The translator had extensive experience with technical
translation and CAT, but reported no experience with MT-assisted TM
translation. At the TSP, s/he was known as a very good and highly
productive translator.
The source text, comprising Frequently Asked Questions for a surround-
sound speaker system, contained 625 words divided into 76 segments,
which were pretranslated, using the TM and the MT engine. When the TM
contained matches with match values of 70 % and above, these matches
were inserted by the CAT tool into the target-text segments and displayed

112
on the screen. Segments with a match value below 70 % were treated as
no matches and were pretranslated using the MT engine. Thus, for every
source-text segment, the translator was automatically presented either
with a TM match — that is, a context match, a 100 % match or a fuzzy
match above 70 % — or an MT match. As a reference text, the translator
received a PDF file with the fully formatted source text.
Figure 1 shows the interface of the TM suite and some of the matches that
were provided. In this tool, MT matches are marked by the abbreviation of
AT for Automated Translation.
Figure 1: Screen shot of the MT-assisted TM environment.
The translator in question spent approximately 25 minutes on the
translation, which made him/her the fastest translator in the experiment.
After the translator had translated the text, we compared the final
translation with the pretranslated version by means of the software
SDLXLIFF Compare, in order to highlight all changes made by the
translator. Approximately one hour after the experiment, the translator
participated in a cued retrospective interview in which s/he was asked to
verbalise his/her processes regarding specific segments, while watching
the corresponding parts of the translation process on the screen-capture
video. For this, Bundgaard had selected segments in which the translator
had made either many, very few or no changes. At the end of the
interview, the translator filled in a post-experimental questionnaire about
his/her background and experience.

113
In order to investigate which types of TCI actions can be observed in the
process analysed for this paper, the entire process was first analysed to
quantify the number of the various match types offered to the translator.
As a translator interacting with an MT-assisted TM suite can either choose
to accept a match as it is, i.e. without any changes, to revise it or to
reject it, i.e. delete the match and translate the segment from scratch, the
translator’s observable actions were categorised for each match as
belonging to one of three choices: Accept, Revise or Reject. Based on an
analysis of the screen-capture data, Table 1, below, gives an overview of
match types and the translator’s choices. In this case, the MT-assisted TM
tool offered the following match types: two context matches, 28 100 %
matches, 16 fuzzy matches (from the 70-99 % match range) and 30 MT
matches.
Match types Number of
matches in
the text
Translator’s choices
Accept Revise Reject
Context
matches
2 2 0 0
100 %
matches
28 26 2 0
Fuzzy matches 16 2 14 0
MT matches 30 2 28 0
In total 76 32 44 0
Table 1: Overview of match types and the translator’s choices.
As shown by Table 1, regarding the 76 matches analysed, the translator
chose to accept 32 matches, to revise 44 matches and to reject none. This
means that at no point did the translator choose to delete a proposed
match and translate it from scratch. Consequently, in the following we
shall not deal further with the reject category. However, as one might
expect the MT engine to generate at least some MT matches of poor
quality, we find it very interesting that the translator did not reject any of
them and translate the segment him/herself. In the course of the
retrospective interview, the translator explained that sometimes s/he
found MT matches hard to handle (in our translation):
[…] in some cases it is actually confusing, you think it is rubbish, and I would
actually have been better off just thinking this from the beginning because now I
have been trapped by this and this word, which I feel obliged to use, and it can be
difficult to set this aside and say it is me who makes this translation, it is me who
decides what to write […].
Subsequently, the translator mentioned that his/her reason for feeling
“trapped” was that s/he did not know the origin of the MT matches: The
translator did not know whether the MT engine drew on material from the

114
Internet, which s/he was allowed to reject, or whether it drew on material
from the customer in question, which s/he should probably accept, in
accordance with a general norm of giving priority to customer preferences.
The fact that the translator felt trapped in this way suggests that the tool
had a restraining influence on the translator’s processes, and that his/her
decision not to delete any MT matches, but to retain (parts of) them,
might be regarded as his/her way of accommodating resistance from the
tool, to use Olohan’s (2011) terms (see section 1, above).
4. Results and discussion
To gain a deeper understanding of how this particular translator interacted
with the CAT tool, to describe what happened in the translation process
and, if possible, to determine the types of TCI actions carried out, the
matches mentioned in Table 1 belonging to the Accept and Revise
categories were analysed adopting a qualitative and inductive approach
(cf. Saldanha and O’Brien 2013: 190). In section 4.1, we shall look at the
matches that were analysed as belonging to the Accept category and, in
section 4.2, we shall look at the Revise category (4.2). To illustrate types
of TCI actions, examples will be provided with a description of the
unfolding translation process. If relevant, the translator’s retrospective
interview will be used as supplementary data. For each example, we shall
consider if the tool seemed to have a restraining or aiding influence on the
translation process.
4.1. The Accept category
As shown by Table 1, above, of the 32 matches which the translator
accepted without making any changes, 26 were 100 % matches, two were
context matches, two were fuzzy matches (from the 95-99 % match
range) and two were MT matches. Hence, it seems that the translator
tended to accept 100 % matches and content matches, whereas only very
few fuzzy matches and MT matches were accepted. At TextMinded, when
project managers prepare the source files for translation, SDL Trados
Studio is configured automatically to mark all 100 % matches and context
matches as confirmed translations. This means that when a translator
opens a file, these segments already have the status of Translated. As the
translator in question almost always used the shortcut Ctrl+Enter to jump
to the next unconfirmed segment, this meant that the system
automatically skipped these segments and therefore also that the
translator only entered these segments when actively choosing to do so.
In other words, the TM suite encouraged the translator simply to accept
these matches. It can be argued that, when the translator is encouraged
to act like this, s/he might more easily lose track of the text as a cohesive
entity (see also Pym 2011, 2014; Federico et al. 2012), which may be
seen as a restraining influence. As for fuzzy matches from the 95-99 %
match range, the translator probably accepted these without changes as
these matches are characterised only by minor differences between the

115
wording of the current source-text segment and the wording of the
segment retrieved from the TM.
What might be surprising is that the translator decided to accept two MT
matches without changes. One of these is illustrated by Example 1, below,
indicating segment number, match type, the wording of the source-text
segment, the wording of the provided match and the translator’s final
translation. Back-translations (BT) are provided for the match and the
final translation. Lines are broken in the BTs to make them more
comparable with the Danish segments.
Segment
number
Match
type
Source-
text
segment
Wording of
provided
match
Final
translation of
the segment
66 MT Placement
of the
indicator
light.
Placering
af
indikatorlampe.
BT: Placement
of
indicator light.
Placering
af
indikatorlampe.
BT: Placement
of
indicator light.
Example 1: The Accept category.
As can be seen in Example 1, the source-text segment is very short,
which might be the reason why the MT engine was able to produce a
translation that was acceptable to the translator. According to our data,
what the translator did was to run a concordance search on the term
indicator light before accepting the match. Interestingly, the term was
offered by the termbase and, thus, the tool may have had an aiding
influence on the translation process by providing the translator with a
suggestion which conformed to customer requirements. That the
translator chose to run the concordance search anyway indicates that s/he
either did not notice the termbase entry or for some reason felt a need to
check whether the proposed term reflected the typical language use in the
client’s texts.
4.2. The Revise category
As shown by Table 1, above, the translator chose to revise 44 matches,
which included two 100 % matches, 14 fuzzy matches and 28 MT
matches. Hence, our data indicate that fuzzy matches and MT matches
were typically revised by this translator, whereas content matches were
never revised and 100 % matches only very rarely so. As the coding of
the segments belonging to the Revise category revealed two overall types
of TCI, the category was divided into subcategories: match-internal
revision and match-external revision. Match-internal revision refers to a

116
match in which the translator revises without making use of any other
sources or functionalities than the proposed match itself. Match-external
revision refers to a match in which the translator makes use of one or
more sources or functionalities that are external to the match. As we
analysed the translator’s observable actions, not conjectures, we could
only categorise a revision as match-external when an observable action
indicated that the translator had looked at information outside the match,
e.g. when s/he ran a Concordance search. It should also be noted that the
use of the so-called Autosuggest function integrated in the TM suite to
speed up translation, offering automatic and predictive suggestions after
the translator has typed only a few of the first characters of a word, was
categorised as a match-internal activity as the suggestions were provided
automatically. Of the 44 revised matches, 18 matches were match-
internal revisions, while 26 were match-external revisions.
4.2.1. Match-internal revision
Example 2, below, illustrates a match-internal revision, indicating segment
number, match type and the wording of the source-text segment. Steps in
the process are numbered and registered in chronological order, including
the wording of the match provided (step 1). The example also shows the
interim translation of the segment in question (Interim target segment).
In this way, so-called micro-actions below the segment level are also
included. Translations are provided for Danish words (italicised) in the
explanations of each step, and a BT is provided for the first interim target
segment. Again, lines are broken in the BT to make it more easily
comparable with its Danish source.
Segment
number
Match
type
Source-text segment
39 MT Why does my BeoLab 14 not switch to
standby after 3 minutes with no
sound, when the MODE switch is set to
LINE or AMP ?
Steps Interim target segment
1 Wording of provided
match
Hvorfor
bliver min BeoLab 14 ikke omskifter
til standby efter 3 minutter uden
lyd, når tilstanden omskifter
er indstillet til Line eller AMP ?
BT: Why
does my BeoLab 14 not become switch
to standby after 3 minutes without
sound, when the mode switch
is set at Line or AMP ?

117
2 Moves the text
cursor so it is
placed after Hvorfor
(‘Why’) and writes
skifter (‘switch’).
Deletes bliver min
(‘does my …
become’).
Hvorfor skifter BeoLab 14 ikke
omskifter til standby efter 3 minutter
uden lyd, når tilstanden omskifter er
indstillet til Line eller AMP?
3 Moves the text
cursor so it is
placed between
Hvorfor (‘Why’) and
skifter (‘switch’),
moves it back
between skifter
(‘switch’) and
BeoLab 14 and
deletes an
unnecessary space.
Hvorfor skifter BeoLab 14 ikke
omskifter til standby efter 3 minutter
uden lyd, når tilstanden omskifter er
indstillet til Line eller AMP ?
4 Replaces the space
between BeoLab
and 14 with a non-
breaking space.
Deletes omskifter
(‘switch’).
Hvorfor skifter BeoLab 14 ikke til
standby efter 3 minutter uden lyd, når
tilstanden omskifter er indstillet til Line
eller AMP ?
5 Replaces the space
between 3 and
minutter (‘minutes’)
with a non-breaking
space.
Moves the text
cursor so it is
placed between når
(‘when’) and
tilstanden (‘the
mode’).
Writes MODE-om.
Deletes tilstanden
(‘mode’).
Deletes the space
between om and
omskifter (‘switch’)
and deletes om in
omskifter (‘switch’).
standby efter 3 minutter uden lyd,
når MODE-omskifter er indstillet til
Line eller AMP ?
6 Writes the definite
article -en at the
end of omskifter
standby efter 3 minutter uden lyd,
når MODE-omskifteren er indstillet til

118
(‘the switch’).
Selects Line (‘Line’)
and changes it to
capital letters using
a shortcut.
Deletes an
unnecessary space
after AMP (‘AMP’).
LINE eller AMP?
Example 2: Match-internal revision
As is visible from the wording of the match (step 1), the MT engine
translated the construction Why does my BeoLab 14 not switch to […] into
Hvorfor bliver min BeoLab 14 ikke omskifter til […] (’Why does my BeoLab
14 not become switch to’), which is not a correct translation since the
verb switch has been translated into the Danish noun omskifter (‘switch’).
Furthermore, the noun omskifter (‘switch’) cannot be combined with bliver
(‘become’). The reason for this translation is possibly that the termbase
contains the English noun switch with the Danish noun omskifter (‘switch’)
as its translation, and that the termbase is set to overrule the MT
translation if the source-text segment contains a word included in the
termbase (here: switch). Hence, it seems that the integration of MT and
the termbase required the translator to revise this part of the segment. In
Olohan’s (2011) terms, the combination of MT and the termbase might
have constituted resistance offered by the system, and therefore the
translator ended up accommodating this resistance by revising the match
and by using the Danish verb skifter (‘switch’), which was the appropriate
translation (steps 3 and 4). Interestingly, however, in step 5, the tool in
fact provided a correct translation of the noun switch into the noun
omskifter (‘switch’), thereby aiding the translator to meet the customer
requirements.
Prior to the experiment, the translator was told not to translate words
written in red in the source text. Thus, for instance, s/he was not to
translate MODE in Example 2. However, no techniques had been applied in
the MT process to prevent words written in red from being translated and
the MT engine thus translated MODE into tilstanden (‘the mode’). As
shown, the translator changed this back to MODE (step 5). Hence, here
the translator again accommodated resistance from the tool.
The integration with the termbase is probably also the reason for another
problem in the match (step 1), namely the translation of LINE into Line
(‘Line’). In the termbase, LINE is translated into Line. As a result, the
translator has to change the lower-case letters of Line to capital letters:
LINE (step 6). By contrast, AMP is not included in the termbase and the
MT engine therefore transferred this unchanged to the target text.

119
The rest of the match seems to have been acceptable to the translator,
except that s/he chose to delete min (‘my’) (step 2) and to replace the
space between BeoLab and 14 with a so-called non-breaking space (a
and 6. These activities were explained by the translator in the interview
like this: The target-text item of min (‘my’) was deleted because the
customer in question prefers to keep texts more impersonal in Danish,
and the insertion of a non-breaking space between BeoLab and 14 was
due to the fact that the customer prefers to write product names without
line breaks. Both these preferences are also mentioned in the customer
style guide, which, however, was not consulted by the translator during
the translation process, most probably due to his/her prior experience
with this particular customer (cf. section 3, above). That translators are
expected to give priority to customer preferences was also mentioned by
other translators participating in the experiment.
4.2.2. Match-external revision
For the purposes of this paper, match-external actions (defined in section
4.2, above) are operationalised as follows:
(1) Use of the Copy Source to Target function (where the proposed
match is replaced by the source text segment)
(2) Use of the Concordance Search
(3) Insertion of an MT match instead of a proposed TM match
(4) Moving back to a previous segment
(5) Use of the reference text (the fully formatted source text)
In the following, due to space restrictions, we shall confine ourselves to
discussing examples of match-external actions 1 and 3-5. Example 3
includes action 1 and action 5. Again, BTs are provided, in lines that
correspond to those of the source.
Segment
number
Match type Source-text segment
3 MT Which MODE setting to use
depends on the type of your Bang
& Olufsen television.
match
Hvilken TILSTAND indstilling
afhænger af dit Bang & Olufsen
fjernsyn.
BT: Which MODE setting
depends on your Bang & Olufsen
television.
2 Looks in the pdf with Which MODE setting to use

120
the reference text.
Selects afhænger af dit
Bang & Olufsen
fjernsyn (‘depends on
your Bang & Olufsen
television’) and copies
it.
Uses ”Copy Source to
Target”.
3 Writes Den (‘The’).
Deletes Which.
Den MODE setting to use depends
on the type of your Bang & Olufsen
television.
4 Writes -indstil (‘set’)
after MODE (‘MODE’)
and accepts indstilling
(‘setting’) suggested by
the AutoSuggest
function.
Deletes an unnecessary
space after MODE
(‘MODE’).
Den MODE-indstilling setting to use
5 Places the text cursor
after MODE-indstilling
and writes , du skal
brug, (‘, you should
us,’).
Deletes the second
comma.
Adds –e, to brug (‘us,’)
Den MODE-indstilling, du skal
bruge, setting to use depends on
the type of your Bang & Olufsen
television.
6 Inserts afhænger af dit
Bang & Olufsen
fjernsyn (‘depends on
your Bang & Olufsen
television’).
bruge, afhænger af dit Bang &
Olufsen fjernsyn setting to use
7 Deletes setting to use
depends on the type of
your Bang & Olufsen
television.
bruge, afhænger af dit Bang &
Olufsen fjernsyn.
8 Inserts a non-breaking
space between Bang
and &.
bruge,
Olufsen fjernsyn.
Example 3: Copy Source to Target (action 1) and reference text (action 5).
As it appears, the translator first decided to take a look at the reference
text (step 2). This indicates that the translator wished to see the text in

121
context before determining how to translate the segment. Next, the
translator copied afhænger af dit Bang & Olufsen fjernsyn (‘depends on
your Bang & Olufsen television’) and used the Copy Source to Target
function. Interestingly, our data show that the translator typically uses
this strategy to transfer visual elements or different types of formatting to
the target text, the red colour of MODE, for instance. In the interview,
when commenting on another segment, the translator stated that s/he
usually uses this function to make sure that the formatting is the same in
the translation as in the source text and in order to save him/herself from
unnecessary typing. The translator’s action of using the Copy Source to
Target function illustrates both an aiding and a restraining influence from
the CAT tool: The translator was aided when s/he reduced the typing
effort by inserting the source segment into the target segment, but was
restrained when the tool could not transfer formatting from the source
text. After having used the Copy Source to Target function, the translator
typed the translations of which and setting (steps 3 and 4).
The MT engine left out to use in the translation and the translator
therefore typed the Danish translation du skal bruge (‘you should use’)
(step 5). At the end of the interview, the translator touched upon cases
like this, saying that a difficult thing when working with MT matches was
that small parts of the source-text segment were sometimes omitted and
that the MT engine does not highlight these omissions. As the translator
reported no experience with MT prior to the experiment, we assume that
this perception mainly stems from experience gained during this
experiment, though it may also stem from a webinar held at the TSP
about MT prior to the experiment, during which some very general
guidelines about working with MT were discussed, or from some
colleagues with MT experience. In any case, the fact that parts of the
source-text segment were omitted by the MT engine can be regarded as
resistance from the CAT tool. The translator subsequently inserted the
sentence copied and deleted the rest of the source-text segment (namely
“setting to use depends on the type of your Bang & Olufsen television”, cf.
steps 6 and 7). Finally, the translator inserted a non-breaking space
between Bang and & (step 8), to meet the customer preference already
discussed above.
In Example 4, below, the translator carried out match-external action 3
and action 1.
Segment
number
Match
type
Source-text segment
51 70 PICTURE A software update in
progress:
match
En optagelse er i gang

122
BT: A recording is in progress
2 The translator
inserts an MT match
from the Translation
Results window
using the shortcut
alt+2.
En softwareopdatering i gang:
3 Selects En
softwareopdatering
i gang (‘A software
update in progress’)
and copies it using
ctrl+c.
Uses Copy Source
to Target.
PICTURE A software update in
progress:
4 Selects A software
update in progress
and replaces it with
En
softwareopdatering
i gang (‘A software
update in progress’)
using ctrl+v.
PICTUREEn softwareopdatering i
gang:
Example 4: Insertion of MT match instead of proposed TM match (action 3) and
Copy Source to Target (action 1)
As illustrated by Example 4, when offered an alternative MT match, the
translator chose to insert this into the target-text segment instead of the
proposed TM match (step 2). As the MT match does not contain the
picture included in the source text, the translator then copied En
softwareopdatering i gang (‘A software update in progress’) in order to be
able to re-use it before s/he carried out a match-external action 1 (Copy
Source to Target function) (step 3). Then s/he selected A software update
in progress and replaced it with En softwareopdatering i gang by using
ctrl+v (step 4). Consequently, the final translation of the segment is a
combination of the MT match and the source-text segment.
Examples 5a and 5b, below, illustrate match-external action 4. As shown
by example 5a, the actions were carried out in segments 28 and 31 in
combination with segment 34:
Segment
number
Match
type
Source-text segment
28 100 Shown by light grey areas in the
illustration.

123
31 100 Shown by medium grey areas in
the illustration.
34 99 Shown by the dark grey areas in
the illustration.
Example 5a: Moving back to a previous segment (action 4)
What is interesting about Example 5a is that, at first, the translator
accepted segments 28 and 31 without any changes. Segment 28 was
automatically skipped because of the use of Ctrl+Enter, as described
above, and the translator briefly entered segment 31, but then left it
again. However, as illustrated by Example 5b, below, after entering
segment 34, which is a 99 % TM match with a character string very
similar to those of segments 28 and 31, albeit somewhat differently
structured, the translator chose to copy part of segment 34, returning first
to segment 28 and then to segment 31 to insert the wording and the
same sentence structure provided by the TM for segment 34. The
translator then returned to segment 34 and revised something here as
well, thus obtaining parallel structures in all three segments.
Segment Steps Interim target segment
match
Det er vist med
de mørkegrå områder i
illustrationen ovenfor.
BT: This is shown by
the dark grey areas in
the illustration above.
Selects er vist med (‘is
shown by’) and copies it
using ctrl+c
Det er vist med de mørkegrå
områder i illustrationen
ovenfor.
match
Dette
viser de lysegrå områder i
BT: This
the light grey areas show
in the illustration above.
Selects viser (‘show’)
and replaces it with er
vist med (‘is shown by’)
using ctrl+v
Dette er vist med de lysegrå
områder i illustrationen
ovenfor.
match
Dette
viser de mellemgrå områder
i illustrationen ovenfor.

124
BT: This
the medium grey areas show
in the illustration above.
Selects viser (‘show’)
and replaces it with er
vist med (‘is shown by’)
using ctrl+v
Dette er vist med de
mellemgrå områder i
34 Adds -te to det (‘this’) Dette er vist med de
mørkegrå områder i
Example 5b: Match-external revision
As illustrated by Examples 5a and 5b, the translator first accepted the 100
% matches in segments 28 and 31 and then returned to revise them after
having been provided with relevant matches later in the translation
process. The translator’s attempt to ensure consistency in the target text
by breaking with the sequential, segment-by-segment method encouraged
by the CAT tool may be seen as his/her way of accommodating resistance
from the tool.
5. Concluding remarks
The observational study of TCI processes reported on in this paper gives
us some indication of the impact of CAT tools on translation processes
from the translator’s perspective and, in particular, provides us with a
deeper understanding of how CAT in the form of MT-assisted TM is carried
out as an interaction between translator and computer. Of course, it
should be noted that our data are limited to one particular instance of MT-
assisted TM translation involving a single translator, which means that our
results cannot be generalised.
According to our results, out of the 76 TM and MT matches that were
offered to the translator during the process analysed for this paper, the
translator chose to reject none, to accept 32 and to revise 44. We were
surprised that no MT matches were rejected as these, unlike TM matches,
are machine-generated and therefore could be expected to be of poorer
quality. However, the behaviour of the translator in the experiment
indicates that s/he generally valued all kinds of matches and tried to re-
use as much of the provided matches as possible, regardless of whether it
was an MT match or not.
As for the matches accepted by the translator, these were content
matches and 100 % matches, 95–99 % fuzzy matches and very short MT
matches, as might have been expected. As regards matches revised by
the translator, the majority were MT matches and fuzzy matches.

125
It was found that, when revising matches, the translator used two overall
strategies, which we categorised as match-internal revisions and match-
external revisions. The revisions were mainly categorised as match-
external, indicating that the translator felt a need to carry out various
kinds of research outside the match offered. Hence, it seems that the
translator was aware that the suggestions provided by the tool were not
always adequate for a professional translation. So, while the translator
seemed to have been aided by the CAT tool, in several cases, the
translator also appeared to accommodate resistance from the tool, to use
Olohan’s (2011) terms. Such accommodation was necessary, for instance,
when the translator wished to conform to a general norm of giving priority
to customer preferences — a norm that is also mentioned by LeBlanc
(2013).
It was interesting to see how the translator broke with the sequential,
segment-by-segment method that is encouraged by the CAT tool,
indicating that s/he was (still) focused on the text as a cohesive entity, an
approach that is generally assumed to be discouraged by the CAT tool
(Pym 2011, 2014; Federico et al. 2012). Thus, even though the translator
was invited by the tool to work on one segment after the other checking
for terminological and phraseological consistency, which Pym (2011: 3)
describes as the prototypical behaviour when translators are asked to use
TM in combination with MT, the translator never seems to have lost track
of the syntagmatic cohesion. Hence, our results do not confirm the
assumption of Bowker and Fisher (2010), for instance, that using CAT
tools necessarily imply working sequentially with the text instead of going
back and forth in a recursive process. The translator in our experiment
seems to have worked recursively sometimes, and — contrary to the
students in the experiment of Christensen and Schjoldager (2011) — s/he
did not seem to lose track of the text and of the aim of the translation.
However, our study also suggests that translators need to be better
informed about the origin of MT matches. Thus, for instance, the
translator in question might have felt less “trapped” by the MT matches
(see the quote from the retrospective interview in section 3, above) and
interacted better with the CAT tool had s/he known more about the
technicalities of MT translation, how the MT engine had been trained, for
instance.
To sum up, the observational study reported on in this paper sheds light
on the interaction between the translator and the computer and
exemplifies how the CAT tool restrains and aids the translation process.
The study indicates that the CAT tool played a central role in the process
and that the translator was able to juggle various kinds of challenges,
such as customer preferences, complications caused by the integration of
the MT engine and the termbase, a feeling of being “trapped” by the
suggestions provided by the MT engine and a wish to create a cohesive
text within the CAT environment. Consequently, though the CAT tool
seems to have offered resistance in different ways, the translator seemed

126
to be able to accommodate this resistance, to use Olohan’s (2011) terms.
Hence, in this particular instance of MT-assisted TM translation, the
translator appeared to remain in charge of the translation process and,
consequently, though in constant interaction with the technology, s/he
remains at the centre stage of the translation profession.
Bibliography
 Alves, Fabio and Tânia Liparini Campos (2009). “Translation Technology in Time:
Investigating the Impact of Translation Memory Systems and Time Pressure on Types
of Internal and External Support.” Susanne Göpferich, Arnt Lykke Jakobsen and Inger
M. Mees (eds) (2009). Behind the Mind. Methods, Models and Results in Translation
Process Research. Copenhagen: Samfundslitteratur, 191–218.
 Bowker, Lynne and Des Fisher (2010). “Computer-aided translation.” Yves Gambier
and Luc van Doorslaer (eds) (2010). Handbook of Translation Studies, Volume 1.
Amsterdam/Philadelphia: John Benjamins, 60–65.
 Bundgaard, Kristine (2013). Computer-assisted Translation with a Focus on Revision
at TextMinded Danmark A/S – a Case Study. Thesis proposal. Aarhus University.
 Carroll, John M. (1997). “Human-Computer Interaction: Psychology as a Science of
Design.” Annual Review of Psychology 48, 61–83.
 — (2013). “Human Computer Interaction – brief intro.” Mads Soegaard and Rikke Friis
Dam (eds) (2013). The Encyclopedia of Human-Computer Interaction, 2nd
Ed. Aarhus:
The Interaction Design Foundation. http://www.interaction-
design.org/encyclopedia/human_computer_interaction_hci.html (consulted
14.04.2014).
 Chesterman, Andrew (2009). “The name and nature of translator studies.” Hermes
42, 13–22.
 Christensen, Tina Paulsen (2011). “Studies on the Mental Processes in Translation
Memory-assisted Translation – the State of the Art.” trans-kom 4(2), 137-160.
 Christensen, Tina Paulsen and Anne Schjoldager (2010). “Translation-Memory
(TM) Research: What Do We Know and How Do We Know It?” Hermes 44, 89–101.
 — (2011). “The Impact of Translation-Memory (TM) Technology on Cognitive
Processes: Student-Translators' Retrospective Comments in an Online Questionnaire.”
Bernadette Sharp, Michael Zock, Michael Carl and Arnt Lykke Jakobsen (eds) (2011).
Proceedings of the 8th International NLPCS Workshop: Human-Machine Interaction in
Translation. Copenhagen: Samfundslitteratur, 119–130.
 — (2016). “Computer-aided translation tools – the uptake and use by Danish
translation service providers.” The Journal of Specialised Translation 25, 89–105.
 Dragsted, Barbara (2006). “Computer-aided translation as a distributed cognitive
task.” Pragmatics & Cognition 14(2), 443–464.
 Federico, Marcello, Alessandro Cattelan and Marco Trombetti (2012).
“Measuring user productivity in machine translation enhanced computer assisted
translation.” Proceedings of the Tenth Conference of the Association for Machine
Translation in the Americas (AMTA).

127
http://amta2012.amtaweb.org/AMTA2012Files/papers/123.pdf (consulted
01.08.2014).
 Fiederer, Rebecca and Sharon O’Brien (2009). “Quality and Machine Translation: A
realistic objective?” The Journal of Specialised Translation 11, 52–74.
 Flanagan, Marian and Tina Paulsen Christensen (2014). “Testing post-editing
guidelines: how translation trainees interpret them and how to tailor them for
translator training purposes.” The Interpreter and Translator Trainer 8(2), 257–275.
 Folaron, Deborah (2010). “Translation tools.” Yves Gambier and Luc van Doorslaer
(eds) (2010). Handbook of Translation Studies, Volume 1. Amsterdam/Philadelphia:
John Benjamins, 429–436.
 Garcia, Ignacio (2011). “Translating by post-editing: is it the way forward?” Machine
Translation 25, 217–237.
 Guerberof Arenas, Ana (2009). “Productivity and quality in the post-editing of
outputs from translation memories and machine translation.” Localisation Focus 7(1),
11–21.
 — (2012). Productivity and quality in the post-editing of outputs from translation
memories and machine translation. PhD thesis. Universitat Rovira i Virgili.
 — (2014a). “The Role of Professional Experience in Post-Editing from a Quality and
Productivity Perspective.” Sharon O’Brien, Laura Winther Balling, Michael Carl, Michel
Simard and Lucia Specia (eds) (2014). Post-Editing of Machine Translation: Processes
and Applications. Newcastle upon Tyne: Cambridge Scholars Publishing, 51–76.
 — (2014b). “Correlations between productivity and quality when post-editing in a
professional context.” Machine Translation 28(3-4), 165–186.
 Jiménez-Crespo, Miguel A. (2009). “The effect of Translation Memory tools in
translated Web texts: Evidence from a comparative product-based study.” Linguistica
Antverpiensia 8, 213–234.
 Johnson, Peter (1992). Human Computer Interaction: Psychology, Task Analysis and
Software Engineering. London: McGraw Hill.
 Läubli, Samuel, Mark Fishel, Gary Massey, Maureen Ehrensberger-Dow and
Martin Volk (2013). “Assessing Post-Editing Efficiency in a Realistic Translation
Environment.” Sharon O’Brien, Michael Simard and Lucia Specia (eds) (2013).
Proceedings of MT Summit XIV Workshop on Post-editing Technology and Practice,
83–91. http://www.mt-archive.info/10/MTS-2013-W4-Laubli.pdf (consulted
20.11.2014).
 LeBlanc, Matthieu (2013). “Translators on translation memory (TM). Results of an
ethnographic study in three translation services and agencies.” Translation and
Interpreting 5(2), 1–13.
 Leijten, Mariëlle and Luuk Van Waes (2013). “Keystroke Logging in Writing
Research: Using Inputlog to Analyze and Visualize Writing Processes.” Written
Communication 30(3), 358–392.
 Muñoz Martín, Ricardo (2014). “A blurred snapshot of advances in translation
process research.” Ricardo Muñoz Martín (ed.) (2014). Minding Translation. Special
Issue 1 of MonTI, 49–84.

128
 O’Brien, Sharon (2007). “Eye-Tracking and Translation Memory Matches.”
Perspectives: Studies in Translatology 14(3), 185-205.
 — (2012). “Translation as human-computer interaction.” Translation Spaces 1(1),
101–122.
 O’Brien, Sharon, Minako O’Hagan and Marian Flanagan (2010). “Keeping an eye
on the UI design of Translation Memory: How do translators use the “Concordance”
feature?” European Conference on Cognitive Ergonomics. Delft, 1–4.
http://doras.dcu.ie/16693/1/ECCE_2010.pdf (consulted 20.11.2014).
 Olohan, Maeve (2011). “Translators and translation technology: The dance of
agency.” Translation Studies 4(3), 345–357.
 Pym, Anthony (2011). “What technology does to translating.” Translation and
Interpreting 3(1), 1–9.
 — (2014). Exploring Translation Theories. Second Edition. London/New York:
Routledge.
 Risku, Hanna (2014). “Translation process research as interaction research: From
mental to socio-cognitive processes.” Ricardo Muñoz Martín (ed.) (2014). Minding
Translation. Special Issue 1 of MonTI, 331–353.
 Saldanha, Gabriela and Sharon O’Brien (2013). Research Methodologies in
Translation Studies. Manchester: St. Jerome Publishing.
 SDL (n.d.). About Translation Memory Matches.
http://producthelp.sdl.com/sdl%20trados%20studio/client_en/Edit_View/TMs/EVWorki
ngwithTMsAbout_Translation_Memory_Matches.htm (consulted 11.04.2014).
 Tatsumi, Midori (2010). Post-Editing Machine Translated Text in A Commercial
Setting: Observation and Statistical Analysis. PhD thesis. Dublin City University.
 Teixeira, Carlos (2011). “Knowledge of Provenance and its Effects on Translation
Performance in an Integrated TM/MT Environment.” Bernadette Sharp, Michael Zock,
Michael Carl and Arnt Lykke Jakobsen (eds) (2011). Proceedings of the 8th
International NLPCS Workshop: Human-Machine Interaction in Translation.
Copenhagen: Samfundslitteratur, 119–130.

129
Biographies
Kristine Bundgaard is a PhD Fellow at the Department of Business
Communication, School of Business and Social Sciences, Aarhus
University, Denmark. Her research examines translator-computer
interaction in MT-assisted TM translation in a workplace context. She has
a background as a state-authorised translator between Danish and
German and teaches courses in translation and text production.
E-mail: kribu@bcom.au.dk
Tina Paulsen Christensen, PhD, is an Associate Professor at the
Department of Business Communication, School of Business and Social
Sciences, Aarhus University, Denmark, where she teaches translation
tools, professional translation and interpreting. With a background as
state-authorised translator between Danish and German, her research
interests cover computer-assisted translation, translation process
research, legal translation and court interpreting.
E-mail: tpc@bcom.au.dk

130
Anne Schjoldager, PhD, is an Associate Professor at the Department of
Business Communication, School of Business and Social Sciences, Aarhus
University, Denmark, where she teaches translation, translation theory
and English-language text production. With a background as state-
authorised translator between Danish and English, her research interests
are professional translation processes, strategies and tools.
E-mail: asc@bcom.au.dk

Translator-computer interaction in action — an observational process study of computer-aided translation

More Related Content

What's hot (7)

Similar to Translator-computer interaction in action — an observational process study of computer-aided translation (20)

More from antonellarose (9)

Recently uploaded (20)

Translator-computer interaction in action — an observational process study of computer-aided translation