Relating Production Units and Alignment Units
in Translation Activity Data
Michael Carl and Arnt Lykke Jakobsen
Dept. of International Languages Studies & Computational Linguistics
Copenhagen Business School, 2000-Frederiksberg, Denmark
Abstract. The definition and characterisation of Translation Units (TUs) in hu-
man translation is controversial and has been described in many different ways.
This paper looks at TUs from a translation process perspective: we investigate
the sequences of keystrokes which have been typed during translation production
and re-define TUs in terms of text production units (PUs). We correlate those
units with translation equivalences in the translation product, so-called alignment
units (AUs) and compare the translation performance of student and professional
translators on a small translation task of 160 words from English into Danish. In
contrast to what has frequently been assumed, our data reveals that TUs are rather
coarse, as compared to the notion of ‘translation atoms’, comprising several AUs,
and they are particularly coarse for professional translators.
1 Introduction
There is a large body of literature on segmentation in translation which can be separated
into two fundamentally different kinds: research into human translation processes seeks
to find basic segments of activities in the translation process, whereas others think of
the segments more statically as properties observable in the translation product i.e.
correspondences in pairs of texts as a result of a translation process. Accordingly, there
is a confusion in the usage of the term “translation unit” (TU), which sometimes refers
to the former and sometimes to the latter type of unit. However, it is by no means
clear that there is isomorphism between the units that a translator has in mind during
translation and the correspondencies which can be made out later in the final translation
product.
According to [1] translation units are lexicological units: they are signs, each with
de Saussures two components, the signifiant and the signifi
´
e. Such a unit is “the small-
est segment of the utterance where the cohesion of signs is such that they cannot be
translated separately” [1, p:16]. We will refer to this kind of unit, which can be detected
as translation equivalences in the final translation product, as Alignment Units (AUs).
In line with [2, p:14] we adopt the more dynamic view, seeing a TU as “the section
of text which the translator focusses on at any one time”. Similarly, for [3, p:254] the
TU is the “translators focus of attention at a given time in the translation process”. This
definition implies that TUs cannot be directly observed in the user activity data (UAD),
such as in the translator’s keystrokes or gaze movements, but relate to activations in
Carl M. and Lykke Jakobsen A.
Relating Production Units and Alignment Units in Translation Activity Data.
DOI: 10.5220/0003024400370046
In Proceedings of the 7th International Workshop on Natural Language Processing and Cognitive Science (ICEIS 2010), page
ISBN: 978-989-8425-13-3
Copyright
c
2010 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
the translator’s mind. However, we assume that there is no appreciable lag between a
translator’s focus of attention and what he is typing.
1
It therefore becomes possible to investigate units of translation production and as-
sume that they are related to TUs. We define a production unit (PU) as a sequence of:
1. successive keystrokes in time that are not interrupted by a pause longer that a given
PU segmentation threshold. Only deletion and insertion keystrokes are considered.
Navigation activities, using the mouse or combinations of keystrokes are ignored.
2
2. successive keystrokes in text that produce a coherent piece of text. Text producing
and deleting activities are part of the same PU only if they are in close proximity to
each other.
The boundary of a PU is thus defined to lie between two successive keystrokes that are
separated by more than a certain delay in time, or if the second keystroke contributes
to production or revision of a different piece of text [5]. Thus, deletions and corrections
are possible in one PU if they are within the vicinity of the current cursor position.
Note that not all conceivable PUs are equally likely to reflect a unit of the trans-
lator’s focus of attention. As is the case for AUs, TUs must also comprise a coherent
set of signs that “cannot be translated separately”, if, as [6] suggests, skilled translators
proceed “little at a time” combining source text (ST) and target text (TT) segments, so
that “each segment forms a fragment of bi-text in their minds”.
The paper investigates various PU segmentation thresholds to maximise the like-
lihood of the PU being part of a TU. We expect PUs to satisfy two properties: 1.
PUs should represent complete meaning entities rather than arbitrary sequences of
keystrokes. 2. PUs should align with AU boundaries, rather than crossing their lines.
We are looking for a PU segmentation threshold that maximises these properties.
In section 2 we describe the experimental setup for data acquisition, the hard and
software used to collect the UAD as well as details about the translation task. Section 3
looks into details and analyses a data segment in depth. Section 4 applies the analysis
technique to a set of 24 translations and provides a large number of PU correlations.
2 Data Acquisition
We base our research on a translation experiment in which 12 professional and 12 stu-
dent translators produced translations using the version of the Translog [7] software
3
developed as part of the EU Eye-to-IT project
4
. Translog presents the ST in the upper
part of the monitor, and the TT is typed in a window in the lower part of the monitor.
1
Note that this is an adapted version of the “eye-mind assumption” [4], which hypothesises that
“there is no appreciable lag between what is being xated and what is being processed” [4,
p:331].
2
Some people read electronic texts by moving the cursor on the fixated words. If navigation
keystrokes were to count in PUs, we would obtain long PUs consisting only of navigation
activity.
3
The keylog portion of the software can be obtained from www.translog.dk
4
cogs.nbu.bg/eye-to-it/
38
When the start button is pressed, the ST is displayed and eye movement and keystroke
data begins to be registered. The task of the translator is to type the translation in the
lower window. After having completed the translation, the subject presses a stop button,
and the translation, together with the translation process data, are stored in a log file.
This so-called User Activity Data (UAD) is then transformed into a relational data
structure which allows us to map eye movements and keyboard actions onto ST and
TT positions and vice versa: [8,9] describe how keyboard actions are related to the TT
words to which they contribute and how TT words are mapped on ST words, so that
for (almost) each keystroke, we can determine to which ST translation it contributes. In
order to do so automatically, the relational data structure requires, besides the prepara-
tion of the keyboard and eye movement data, also the alignment information between
the ST and the TT.
For the data set of the 24 translations we semi-automatically aligned the ST and TT
and converted the alignment and UAD into the required structure.
Killer nurse
| {z }
1
receives
| {z }
2
four
| {z }
3
life sentences
| {z }
4
z }| {
Mordersygeplejer
z }| {
modtager
z}|{
fire
z }| {
domme p˚a livstid
Fig.1. A segment of the product data shows four AUs aligning 6 ST words and 6 TT words; AUs
were manually aligned.
3 Analysis of Production Units
In this section we look more closely at a time segment of 20 seconds in which the
translation shown in figure 1 was typed. We develop and discuss properties of the PUs
and show how these properties generalise to the entire translation session. In section 4
we apply these criteria to all 24 translations.
Details of the keystroke segmentations are reproduced in tables 1 and 2. During
these 20 seconds, 56 keystrokes were produced. Each different PU segmentation thresh-
old groups the data differently with different properties of the produced PUs.
The 400ms segmentation in table 1 groups the 56 keystrokes into 10 PUs. Due to the
pauses of 894ms and 488ms after “y” and “j” respectively, the word “Mordersygeple-
jerske” is segmented into 3 PUs. Note that the suffix “ske” is later deleted and thus does
not appear in the final translation in figure 1. On the other hand, the two AUs “mod-
tager” and “fire” are written smoothly so that these two AUs are grouped into one PU.
The “start” column in table 1 gives the timestamp when the PU was started to be typed.
The “dur” column shows the time needed to complete the PU, and “pause” shows the
interval of time following the PU until the next keystroke was processed. Note that the
duration is 0 if the PU contains only one keystroke. The AU” column shows which
AUs are generated by each PU. Thus, the first three PUs all contribute to the translation
of the first AU
1
while the fourth PU contains AU
2
and AU
3
.
A number of corrections occur in the following segments: first the letter “s” is writ-
ten and then deleted in the next segment. The deletion is indicated in brackets “(s)”.
39
Table 1. Properties of PUs as generated with 400ms segmentation.
NR start dur pause type AU PU
1 10485 1397 894 SAWU 1 Mordersy
2 12776 944 488 WUWU 1 geplej
3 14208 717 432 WUSA 1 erske
4 15357 2177 2719 SASA 2,3 modtager fire
5 20253 0 702 SAWA 4 s
6 20955 0 4141 WUWU 4 (s)
7 25096 118 886 WUWU 4 li
8 26100 192 669 WUWU 4 (li)
9 26961 1051 536 SASU 4 domme p˚a
10 28548 907 749 SUSA 4 livstid
Then “li” is produced and then deleted “(li)”. Finally the translation domme p˚a livstid”
is typed but segmented into 2 PU due to the delay of 536ms after “p˚a”.
The degree to which a PU coincides with a word in the TT translation or an AU
boundary in the ST is indicated by its type. A PU can start and/or end at a word bound-
ary. For instance “livstid” is a complete word and PU
10
starts and ends at a word sepa-
rator. “Mordersy” in contrast is the beginning of a word while “erske” is the ending of
that word. Accordingly, PU
1
starts at a word boundary and PU
3
ends with it.
In addition, a word (or segment) in the target language can start and/or end at an
AU boundary. For instance, “livstid” is the last part of a compound which is grouped
in AU
4
, and so it ends but does not start at the boundary of AU
4
. “domme p˚a” is
the beginning of that same compound and so PU
9
starts, but does not end, at the AU
4
boundary.
Thus the type of a PU consists of four positions (bits), each of which can take two
values: The first two positions indicate properties for the beginning of the PU, and the
last two positions indicate properties of its end. The values indicate whether or not the
PU aligns with word boundaries in the target language, and whether or not it aligns with
boundaries of the AUs of which it is a translation. These values are represented by the
letters [SWAU] which have the following meanings:
S first/last character of PU was a word separator (space, comma, semicolon, colon)
or immediately following a separator.
W first/last character of PU was not a word separator (and not immediately following
a separator)
A first/last character of PU was at an AU boundary.
U first/last character of PU was not at an AU boundary.
Thus, “SAWU”, as in the first line of table 1 indicates that the PU “Mordersy” starts
at the beginning of a word (S), and it aligns the beginning of an AU (A). The last two
letters indicate that this PU ends in the middle of a word (W) and in the middle of an
AU (U). Ideally, as discussed in the introduction, a PU should start and end with a word
separator and/or an AU boundary, such as “modtager fire” in line 4 of table 1. Segmen-
tations in the middle of a word would (perhaps) indicate that attention is focussed on
spelling or typing problems, rather than on translation. Thus, a PU of type “WUWU”
(e.g. line 2: “geplej”) indicates that the segment neither starts nor ends at a word or
40
an AU boundary. Such segments provide little insight into the cognitive processes of
translators, since they do not coincide with meaning units, as e.g. words and AUs do.
However, in the introduction we have argued that PUs should represent signs, with a
signifi
´
e, which is difficult to see in the case of “geplej” or “li”.
Four out of the 10 segments in table 1 are of this type, indicating that a 400ms
segmentation threshold does not correspond to the “cognitive” rhythm of segmentation
that we are looking for.
Table 2. Segmentation 800ms (above) and 1500ms (below) of the same 56 keystrokes from ta-
ble 1. The notation “s(s)” and “[s]” are equivalent, meaning typing and deletion of the bracketed
expression.
start dur pause type AU PU
10485 1397 894 SAWU 1 Mordersy
12776 4758 2719 WUSA 1,2,3 geplejerske modtager fire
20253 702 4141 SAWU 4 s(s)
25096 118 886 WUWU 4 li
26100 4104 3064 WUSA 4 (li)domme p˚a livstid
10485 7049 2719 SASA 1,2,3 Mordersygeplejerske modtager fire
20253 702 4141 SAWU 4 [s]
25096 5108 3064 WUSA 4 [li]domme p˚a livstid
The 800ms PU pattern in table 2 generates only half the number of PUs of those
for the 400ms segmentation. Only one (20%) of them is a “WUWU” segment while the
remaining four (80%) either start or end with a word separator. This pattern is even more
obvious in the 1500ms segmentation of table 2, where all segments show a linguistically
plausible beginning or end. On the other hand, segmentation with longer thresholds
includes more characters and subsumes more than one AU. Thus, the average lengths
of 400ms, 800ms and 1500ms segments of the first 56 keystrokes are 5.6, 11.2 and 18.7
keystrokes, respectively.
Table 3 provides those figures for the entire translation by P13 with various segmen-
tation thresholds for PU patterns. It shows the number of PU segments, their average
length in characters (#char), the percentage of linguistically plausible SASA and S.S.
segments and the percentage of the implausible WUWU segments. Under the 400ms,
800ms and 1500ms segmentation, translator P13 produced 98, 39 and 23 PUs. The
optimum PU segmentation threshold seems to be around 800ms to 1000ms, where a
maximum number of segments are linguistically plausible, and at the same time the
segments do not comprise too many AUs.
4 Segmentation in Writing
This section investigates various PU segmentation thresholds for all 24 translations.
As mentioned earlier in section 2, an English text of 160 words was translated by 12
professional and 12 student translators into Danish. As shown in table 4, there is a
large variance in both overall time needed for the translation and in the number of PUs
produced. Professionals took on average slightly more than 5 minutes to translate the
41
Table 3. Number and properties of PUs for translator P13, generated under various segmentation
thresholds.
thresh. #PU #char %WUWU %SASA %S.S.
200ms 355 2.6 27.04 33.80 37.75
400ms 98 9.4 16.33 37.76 44.90
800ms 39 23.5 2.56 43.59 64.10
1000ms 30 30.5 0.00 50.00 73.33
1500ms 23 39.8 0.00 47.83 69.57
2500ms 14 65.4 0.00 42.86 71.43
text (320 seconds), whereas students took more than 6 minutes (379 seconds). That
is, on average professional translators needed 84% of the time needed by students to
produce the translation. For both groups there was approximately a factor of 3 between
the fastest and slowest translator.
Table 4. Translation time (T-time) and number of PUs for different kinds of segmentation for the
12 professional and 12 student translators.
Professionals #PU
Tr. T-time 400ms 800ms 1500ms
P15 170636 70 18 3
P14 209142 105 38 13
P2 258782 96 38 15
P13 265762 98 39 23
P20 281148 162 50 22
P3 316730 138 62 32
P1 349750 107 33 18
P21 352497 154 69 33
P7 362404 141 78 50
P8 379272 114 58 33
P9 389931 145 71 47
P19 510366 177 84 46
av. 320535 126 53 28
Students #PU
Tr. T-time 400ms 800ms 1500ms
S6 228584 107 39 15
S18 260454 133 50 17
S16 285898 125 57 31
S24 322110 158 81 38
S11 350663 131 70 43
S4 353340 174 63 31
S12 374499 127 76 50
S10 377134 111 71 49
S17 411156 132 68 35
S23 425860 176 99 48
S22 507021 170 105 56
S5 654681 218 116 70
av. 379283 147 75 40
Similarly, on average students produced the translations with 30% more PUs than
professionals, and there was a factor of almost 4 between the smallest and largest num-
ber of PUs produced for both groups. These relations are also plotted in figure 2. The
black rectangular symbols in figure 2 indicate the relation between the translation time
and the number of segments produced with the 400ms PU segmentation threshold, the
triangular symbols those of the 1500ms PU threshold, and yellow squares and diamonds
represent student and professional translators, respectively. All translation thresholds in-
dicate a strong correlation between translation time and the number of segments. That
is, the more the translation is fragmented into a larger number of segments, the longer
is the translation time.
5
5
This correlation is not necessary: many fast-typed segments with short pauses in between them
could amount to the same overall translation time as few segments with long inter-segmental
pauses.
42
Fig.2. The graph shows a strong correlation between the translation time (horizontal) and the
number of PUs (vertical) for three types of segmentation.
The PUs produced by professionals are, on average, longer and the time needed per
PU is (on average) higher for professionals than for those of the students.
Table 5. Average duration and length of PUs with different segmentation thresholds, and the
average typing speed (keystrokes per second) for professional and student translators. (Typing
averages are high because only within-PU keystroke intervals were counted and because PUs
with only one keystroke are recorded with no time duration).
Professionals Students
400ms 800ms 1500ms 400ms 800ms 1500ms
average PU duration in ms 1001 3113 6896 854 2216 5024
average PU length in chars 7.46 17.61 33.55 6.37 12.54 23.24
average (PU chars/PU dur.) 8.22 5.44 4.83 7.18 5.70 5.02
median (PU chars/PU dur.) 7.15 5.61 4.77 6.97 5.53 4.80
Average duration and length (in characters) for various PU segmentation thresholds
is given in table 5. Depending on the PU thresholds, students generate, on average,
15% to 30% shorter segments in length as well as in duration. They also produce more
segments than professional translators, as we have already shown above in table 4 and
in figure 2.
However, the typing speed in terms of characters per time within each PU does not
seem to vary much between professionals and students. The values in table 5 indicate
that average and median inner-segment typing speed between successive keystrokes de-
creases with a growing PU threshold. For instance, at 800ms segmentation threshold,
professionals produce 5.44 characters, while students produce 5.7 keystrokes per sec-
ond. The typing behaviour of students is more fragmented than that of professionals,
with more pauses longer than the segmentation threshold, but when the segments are
43
actually typed, the speed with which successive keystrokes are produced seems to be
identical for both groups.
6
With a segmentation threshold of 1500ms the longest PU was produced by translator
P15, with 183 characters. With a threshold of 800ms the longest PU was produced by
translator P14, comprising 18 AUs and the following 153 characters
7
:
fordi Norris ikke kunne lide at arbejde med [g]ældre mennesker. Alle hans ofte
var s[v][a]vagelige ældre kvin[ger ]der med hjerteproblemer. ALle ville
8
This TU starts with a subordinate clause, that is, the second half of a sentence, it
then contains a whole sentence and ends with the beginning of a third sentence at word
position 150.
Table 6. Number and properties of PUs for different segmentation thresholds: SASA: PUs start
and end with a word separator and an AU boundary, S.S.: PUs start and end with a word separator,
WUWU: PUs start and end in the middle of a word.
200ms 400ms 800ms 1000ms 1500ms 2500ms
#PU 8060 3293 1557 1245 842 524
%SASA 11.39 37.16 48.96 49.72 41.92 37.40
%S.S. 24.54 45.68 53.64 54.22 55.11 50.0
%WUWU 36.72 20.77 7.32 6.27 4.75 4.96
A more detailed listing of the types of PU is given in table 6. The table shows that the
type and distribution of PUs change under different PU thresholds. It gives an overview
of the number of PUs produced with 6 different thresholds and shows the percentage
of linguistically more and less meaningful segments, similar to table 3. The table does
not make a distinction betwee students and professional translators. With increasing
segmentation time, the number of generated segments decreases, and the percentage of
meaningful segments increases. A dramatic change of this effect can be observed up to
ca. 800ms: the meaningless “WUWU” segments fall below 10% and the linguistically
coherent ones grow beyond 50%. Beyond this margin, values change less quickly.
5 Conclusions
The paper investigates activity data of student and professional translators: An English
text of 160 words was translated by 12 professional and 12 student translators into Dan-
ish.
9
All keystrokes and eye movements were recorded using the Translog software.
6
Note that the typing speed was computed as the lapse of time between two (or more) successive
keystrokes, so that a PU consisting of a single keystroke would count as 0ms duration. This
explains the relatively fast typing speed.
7
deletions are in square brackets. The notation “[ger]” means that first “ger was produced
(including blank) and then deleted.
8
The source text of this passage is: “that Norris disliked working with old people. All of his
victims were old weak women with heart problems. All of them could”
9
The ST is reproduced in the Appendix.
44
We investigate and correlate three types of unit: production units (PUs) of keystrokes
where no two keystrokes are separated by a pause longer than a given threshold, align-
ment units (AUs) which reflect translation equivalences in the translation product, and
translation units (TUs) which represent the translators’ focus of attention.
The goal was 1) to determine a threshold for the segmentation of PUs such that
they most likely conform to criteria of TUs, and 2) to investigate properties of PUs for
students and professional translators.
Various thresholds of maximal delay between successive keystrokes are explored to
group sequences of keystrokes into PUs. PUs are considered intelligible if their bound-
aries coincide maximally with linguistic (i.e. word) boundaries in the target language
and with the boundaries of translation atoms as defined by AUs. Our investigationshows
that pauses in writing activity of approximately 1000ms length produce segments of
maximal linguistic plausibility, and are, thus, indicative of cognitive processing units.
That is, a new TU is likely to have started if a pause of 1 second or more can be observed
with no keystroke.
Our findings can be summarised as follows:
The number of PUs correlates strongly with translation time: the longer the trans-
lation time, the more fragmented is the translation into segments (figure 2).
Professional translators produce translations more quickly than students (table 4).
Professional translators produce longer PUs than students in terms of time, as well
as in terms of the number of characters (table 5).
Professional and student translators type PUs at approximately the same speed (ta-
ble 5).
Longer PUs coincide better with word boundaries than shorter PUs (table 6)
Only a small percentage of PUs coincide with a single AU. Rather than producing a
minimal unit or a “translation atom”, we find that translators produce maximal seg-
ments, which seem to increase with the capacity and training of the translator. In a
further study, we intend to investigate properties of these segments in more depth so as
to construct an inventory of cognitive operations associated with the PUs.
References
1. Vinay, J.P., Darbelnet, J.: Stylistique compar´ee du franc¸ais et de l’anglais. Didier, Paris
(1958)
2. Bennett, P.: Translation Units in Human and Machine. Babel 40 (1994) 12–20
3. Alves, F., Vale, D.C.: Probing the Unit of Translationin time: aspects of the design and
development of a web application for storing, annotating and querying translation process
data. Across Language and Cultures 10 (2009) 251–273
4. Just, Carpenter, P.: A theory of reading: from eye fixations to comprehension. Psychological
Review 87 (1980) 329–354
5. Carl, M., Kay, M., Kristian T.H., J.: Long Distance Revisions in Drafting and Post-editing.
In: Cicling, Iasi, Romania (2010)
6. Harris, B.: Bi-text: A New Concept in Translation Theory. Language Monthly 54 (1988)
8–10
7. Jakobsen, A.L.: Logging target text production with Translog. In: [10]. (1999) 9–20
45
8. Carl, M.: Triangulating product and process data: quantifying alignment units with keystroke
data. Copenhagen Studies in Language 38 (2009) 225–247
9. Carl, M., Jakobsen, A.L.: Towards statistical modelling of translators activity data. In-
ternational Journal of Speech Technology 12 (2010) http://www.springerlink.com/content/
3745875x22883306/.
10. Hansen, G., ed.: Probing the process in translation: methods and results. Volume 24 of
Copenhagen Studies in Language. Copenhagen: Samfundslitteratur (1999)
Appendix: Source Test
Killer nurse receives four life sentences
Hospital Nurse Colin Norris was imprisoned for life today for
the killing of four of his patients. 32 year old Norris from
Glasgow killed the four women in 2002 by giving them large
amounts of sleeping medicine. Yesterday, he was found guilty
of four counts of murder following a long trial. He was given
four life sentences, one for each of the killings. He will have
to serve at least 30 years. Police officer Chris Gregg said
that Norris had been acting strangely around the hospital. Only
the awareness of other hospital staff put a stop to him and to
the killings. The police have learned that the motive for the
killings was that Norris disliked working with old people. All
of his victims were old weak women with heart problems. All of
them could be considered a burden to hospital staff.
46