ArticlePDF Available

The effect of translation memory databases on productivity

Authors:

Abstract and Figures

Although research suggests the use of a TM (translation memory) can lead to an increase of 10% to 70%, any actual productivity increase must depends on the TM content. If the target renditions included in the TM database exhibit more free characteristics, this may adversely affect the translator’s productivity. This paper examines how productivity is affected by different kinds of TM databases. A pilot experiment was undertaken to investigate the impact of two different versions of a TM database – free vs. literal TMs. All participants translated the same source text but used different TMs. The results show that in the higher fuzzy-match categories, translators using the less literal TM did not gain as much speed as was the case when using a more literal TM.
Content may be subject to copyright.
From Translation Research Projects 3, ed. Anthony Pym, Tarragona: Intercultural Studies Group, 2011. pp. 63-73.
http://isg.urv.es/publicity/isg/publications/trp_3_2011/index.htm
The effect of translation memory databases
on productivity
MASARU YAMADA
Rikkyo University, Japan
Although research suggests the use of a TM (translation memory) can lead to an increase of
10% to 70%, any actual productivity increase must depends on the TM content. If the target
renditions included in the TM database exhibit more free characteristics, this may adversely
affect the translator’s productivity. This paper examines how productivity is affected by
different kinds of TM databases. A pilot experiment was undertaken to investigate the impact of
two different versions of a TM database free vs. literal TMs. All participants translated the
same source text but used different TMs. The results show that in the higher fuzzy-match
categories, translators using the less literal TM did not gain as much speed as was the case
when using a more literal TM.
Key words: translation memory, productivity, localization.
Introduction
The role of the technical translator has changed as a direct result of translation memory (TM)
technology. Translators are no longer focused on translating texts from scratch, but on recycling
previously translated texts: the essence of the technology is “text re-use” (García 2009). As TM
databases have changed into network-based systems, the translated texts are no longer locally
managed by translators but rather centralized by translation bureaus or Language Service
Providers (LSPs). In addition, due to the use of sections within localization projects,
independent players such as quality-assurance checkers and client reviewers make extensive
changes to the translated texts. By the time the texts are finalized, translators have lost control
over their own translations.
Under these circumstances, translators using a TM provided by the LSP must deal with the
imposed segments they have not generated themselves. This means more time checking and
editing, thus adversely affecting productivity. Although previous studies have shown that the
use of TM can lead to a 10 to 70% increase (Bowker 2005, Dragsted 2004, O’Brien 1998,
Somers 2003), the actual productivity gain must depend on the TM content.
Target renditions exhibit more free characteristics when a product is adapted, customized,
or highly localized, possibly due to reviewers’ extensive modifications in the course of the
localization process. If these texts are put into the TM database, it may have an impact on the
translator’s productivity. This present study will therefore examine how productivity is affected
when different renditions are put into the TM database.
Relevant literature on TM productivity
The desire to increase productivity is one of the main reasons for using a TM, and this aspect
has been investigated in empirical studies. According to O’Brien’s experiment (1998: 119),
anything from 10% to nearly 70% can be leveraged from the TM. Somers (2003: 42) states that
“while on occasion a TM product might result in a 60% productivity increase […], 30% may be
a more reasonable average expectation.” Dragsted (2004: 210) has indicated that the average
increase was 16 % for students and 2% for professionals. Bowker’s pilot study (2005: 17)
64
shows that translators without using a TM could not finish a 387-word translation within the 40-
minitue time frame, while participants using a TM completed the task.
Whatever the exact increase might be, it has been established that the use of TM increases
productivity. However, none of these studies has taken different types of TM content into
consideration. In Bowker’s experiment (2005), two different versions of the TM (original TM
vs. error-included TM) were prepared, but her main objective was to compare the quality of
products. No difference in productivity was recorded. I therefore decided to undertake a pilot
study with the aim of investigating the impact of different TM databases on translation
productivity.
Pilot study
The pilot study was carried in March 2009 using eight student translators as participants. They
came from the Translation and Interpretation program of the Monterey Institute of International
Studies in the United States.
The reason for using student translators rather than professional translators was mainly
convenience. As a visiting scholar at the Monterey Institute during the 2008 academic year, I
had access to a group of 8 students who volunteered to join the pilot experiment. The students’
ages ranged from the early 20s to the early 30s. All students were at Masters level: two subjects
were second-year students and six were first-year students. These students had diverse
backgrounds: some came directly after finishing their undergraduate degrees, others had a few
years of work experience, and one student had professional translating experience. Their
language background also varied: half of the students were Japanese native speakers while the
rest were English native speakers. Despite this diversity, all the students were highly proficient
in both English and Japanese, and we assume their translation skills to be at “near professional”
level. This assumption is not without precedent: when Tirkkonen-Condit (1991) compares the
translation behavior of professional and non-professional translators, the second-year students
represent the “professional translators”, and Bowker (2005) uses Masters students for her TM
error propagation analysis.
Our subjects’ TM skill level was, however, not at a proficient level. They were only novice
or moderate users of TM. Some of them had completed a course on Translation Memories at the
institute, others had not. To make sure that they were comfortable with the tool, I provided a
training session and exercise lessons prior to the experiment. At the end of the trainings, I did
not find major technical difficulties, nor did I see any when observing the actual screen
recordings of their behaviors. Nevertheless, as Ribas (2007) points out, the translators’ relative
computer literacy may affect their translation performance in regard to the quality. This factor
may thus be seen as a limitation of this experiment.
For the experiment, we prepared two different types of TM database for the same source
text. All participants translated the same source text, but they used different TM databases. The
first type was free-translation content (hereafter referred to as TM-F), which was based on
authentic material used in an actual localization project. The pre-translation entailed a number
of additions and deletions. The other type was more of literal-translation content (TM-L), for
which I made modifications on the basis of the TM-F database. The TM-L content was not
necessarily a literal translation: it was at the level of the current translation norm in the
localization industry. Some examples of the differences between TM-F and TM-L are shown in
Table 1.
In TM-F, for instance, the source word application in the first example was rendered as
program in the target Japanese text. In the second example, the source word current was not
translated. In the third example, the source phrase basing on the features and tasks of your
computer was eliminated in the target text. Although some may claim that these features may be
close to mistranslations in terms of formal correspondence, the TM-F content was, after all,
authentic and was accepted in the market.
It is important to note the “match rate” of the TM. In principle, the TM functions as a
database that stores previously translated content as paired source and target segments, and
65
retrieves the translation segment for “recycling”. The similarity level is indicated by the match
rate, based on the syntactical structure of the source text. For instance, if a new sentence is said
to be an “80 percent match” (fuzzy match) of an existing sentence, this represents the high
resemblance and only a few corrections are required by the translator in the target text. If the
new sentence is a “100 percent match” (exact match), this means that there is a high probability
of no change at all in the target text. The source text used in our experiment was identical for
both groups; therefore, the match rate for each sentence (or segment) was the same. However,
because different target renditions were prepared for each type of the TM database, I expected
different editing efforts to be required by translators.
Table 1: Sample sentences from TM-F and TM-L, with back-translation into English
ST TM-F (Free) TM-L (Literal)
Application
Configuration Wizard プログラム設定ウィザード
Program Configuration Wizard’
アプリケーション設定ウィザード
Application Configuration Wizard’
Obtaining information
on the current status of
components and tasks
and statistics on them
コンポーネントとタスクのステー
タスと統計情報の取得
‘The status of component and task, and
statistics information ACC obtaining. ‘
コンポーネントとタスクの現在のステ
ータスおよび統計の情報の取得
‘Information on component’s and task’s
current status, and statistics –ACC
obtaining.’
Its task is to help you
configure the initial
settings of the
application basing on
the features and tasks
of your computer.
設定ウィザードはコンピュータへ
適切な保護設定を行うお手伝いをし
ます。
‘Configuration Wizard helps to
configure the protection settings properly
to your computer‘
アプリケーション設定ウィザードは、
コンピュータの特徴やタスク基づき、ア
プリケーションの初期設定を行うお手伝
いをします。
‘Application Wizard helps to configure the
initial settings of the application, basing on
the features and tasks of your computer
Hypothesis
Because the use of a TM forces the translator’s cognitive segmentation into smaller linguistic
chunks (cf. Dragsted 2004), the translator, if using TM-L, should be able to easily identify one-
to-one correspondences between the source and pre-translated texts. I therefore hypothesized
that TM-L would correlate with faster translation speeds than TM-F.
General experiment design
Translators were requested to translate a text of about 500 words from English into Japanese
using the TM. The text was from an anti-virus software manual, a topic normally encountered in
the localization industry. The translators were put into two sub-groups: TM-L and TM-F. All of
them were asked to translate the same source text, using either the TM-L or the TM-F database.
They were not notified of which TM database they would be using.
The experimental set-up was designed to reflect the translators’ natural work environment.
No time restriction was given for the task. The subjects were allowed to use their own
computers and were permitted access to their usual reference materials, including the Internet,
in addition to the TM provided.
All of the subjects’ operations on their PC screens were recorded using BB Flashback. It
recorded searches of electronic resources, cursor movements, clicks, and keystrokes as well as
the translations. The recorded data were analyzed to trace the history of each translator’s
activities. BB Flashback was installed on each subject’s computer and worked in the
background so that it did not affect the subject’s natural work environment.
The TM tool used for this experiment was SDL Trados 2007, the most common tool of this
kind in Japan and the market leader in the world localization industry. Nearly 80% of translation
service providers in Japan that use some kind of translation memories adopt SDL Trados (Japan
66
Translation Federation, 2008). According to Lagoudaki (2006: 21), the TM most used
worldwide is also SDL Trados.
Because this experiment was a pilot study, the sample size was small. A total of 8 students
was obviously not a high number, especially to assess the statistical significance of quantitative
analysis.
Results
The results of the pilot study are shown in Table 2. Contrary to my prediction, the overall
difference in speed between TM-F and TM-L was not highly significant. The average
production time shows TM-F 1:04.22 vs. TM-L 1:05.44, meaning that the overall production
speed with TM-F was actually marginally faster.
This may partly be attributed to two factors. The first factor was that subject A1 in TM-F
group recorded an exceptionally high productivity gain so that this translator’s speed
contributed greatly to the overall average speed in the TM-F group. The WPM (words per
minute) of A1 was 11.29, compared with a median of 7.48. The second cause was subject D1 in
TM-L group, who was the slowest translator of all, achieving only 6.88 WPM. The data without
these two translators would make the result TM-F 7.21 WPM and TM-L 8.13 WPM, which is in
line with our expectations. Nevertheless, such intentional manipulation cannot be an option here
unless there is a legitimate reason to do so.
Table 2: Translation productivity: TM-F vs. TM-L
Total Time
Mean Time
Mean WPM
TM-F
A1
0:45:10
1:04:22
8.23
A2
1:12:42
B1
1:11:46
B2
1:07:48
TM-L
C1
1:00:06
1:05:44
7.82
D1
1:14:05
D2
1:08:31
D3
1:00:15
Production speed for100% matches
Although the overall average data did not indicate any clear advantage of TM-L over TM-F, we
found some differences in translation speed by subdividing detailed data into match-rate
categories. Figure 1 illustrates the WPM for different match rates and a comparison between the
two types of TM databases.
67
Figure 1: Speed comparison: 100%, fuzzy, and no match (WPM)
Comparing the speed at the 100% category, we see that TM-F was still faster than TM-L.
Again, this was not in line with my prediction. However, detailed observation of the screen-
recording data shows that this difference was mostly due to the translators’ technical skill and
how they handled the 100% matches (EM=exact matches). Some translators were familiar with
short-cut key commands to semi-automatically skip the EM segments. Short-cut features should
reduce or eliminate any time spent on the EM segments, and translators who took advantage of
these functions normally paid little attention to these segments. If they were more cautious and
not in the habit of using the short-cut key features, they took some time to check EM segments.
Subject A1 in the TM-F group, the fastest translator of all, made the most use of this feature.
That is probably why the overall processing speed with TM-F in the EM category was higher,
and it had nothing to do with the influence of the content included in the TM database.
We have not yet seen any significant difference between the databases in other matched
categories, other than the fact that translation speed was higher for the EM segments.
Fuzzy-match speed in detail
In order to analyze more closely the effect of two different types of TM, I measured each
individual translator’s speed for every 10% of the fuzzy-match ranges. Table 3 gives the mean
speed of individul translator sorted by the match rate, and Figures 2 and 3 show their behavioral
patterns.
Table 3: Individual translator’s WPM at detail match rate
TM-F TM-L
A1 A2 B1 B2 C1 D1 D2 D3
99-90 23.49 15.01 5.79 9.54 18.21 8.68 30.00 23.29
89-80 34.41 11.22 11.33 11.97 15.78 7.85 12.77 19.15
79-70 19.66 15.31 11.12 16.51 18.07 7.79 15.26 25.22
below 70 12.49 8.05 11.65 11.93 11.75 8.89 7.24 13.52
0.00 4.45 7.25 8.40 5.52 12.34 13.33 5.72 7.78
0,0
20,0
40,0
60,0
80,0
100,0
120,0
100 Fuzzy No Match
TM-F
TM-L
68
Figure 2: WPM change at detailed match rate for TM-F
Figure 3: WPM change at detailed match rate for TM-L
In the graph for TM-F (Figure 2) we see that the production speed with subject A1 was
much higher than the rest of the translators. A1’s speed began at 23.49 WPM in the 99-90%
match range, increased to 34.41 WPM at 89-80% category, then constantly dropped almost
proportionally to the decreasing match rate, until it reached no-match with 4.45 WPM. A1’s
dynamic range (the difference between the highest and the lowest WPM) was approximately 30
WPM.
Conversely, subjects A2, B1, and B2 exhibited much narrower ranges of speed change
through all fuzzy matches. Their overall trend appeared “flatter” than A1’s. For instance, A2’s
fastest peak came around 79-70% with 15.31 WPM, but the slowest speed at no-match showed
7.25 WPM. A2’s range difference was only about 8 WPM; the trend curve did not show as
much dynamic movement as A1’s. A similar tendency was also observed with B1 and B2.
This result suggests that the subjects in the TM-F group did not gain the same benefits from
each fuzzy match segment proposed by the TM. It is normally expected that translation speed at
a high match rate is faster than with lower matches, but this prediction was not applicable to the
case of TM-F. The only exception was noticed with translator A1, whose speed curve changed
almost proportionally to the match rate. However, even in the case of A1, the processing speed
at the 99-90% category was significantly slow. This implies that free TM content may have
reduced the translator’s segmentation recognition speed in higher match categories.
In contrast, the overall trend with TM-L (Figure 3) showed more consistency and a wider
range of speed leverage. The production speed increased almost in step with the increasing
match rate. An exception was found with subject D1, whose curve contradictedly went up as the
match rate decreased. D1’s result was something that should not be observed in professional use
of a TM. Perhaps D1 did not follow any proposed translations presented by the TM. This was
69
also evident from this translator’s post-experiment comments, which I requested participants to
submit after the experiment. Subject D1 stated “TM was of no use for me”.
Other than this exceptional case, however, translators C1, D2, and D3 indicated very
similar characteristics: the speed at 99-90% match was the highest or near highest, and then
dcreased toward no match, almost in proportion to the fuzzy-match rate.
The dynamic range in the case of TM-L was also wider. Translator D2 gave 30.00 WPM at
99-90% match category and 5.72 WPM at no match. The difference was over 24 WPM. D3’s
range was also over 17 WPM, although C1’s trend fell within the range of a little over 6 WPM.
Nevertheless, C1 still recorded a higher translation speed than the average for TM-F.
From these observations, we can conclude that the different types of the TM database
seemed to have been affecting a productivity increase in fuzzy matches. Especially in the higher
fuzzy-match categories, translators using TM-F did not gain as much productivity leverage as
they did in the TM-L group. Hence, the overall dynamic range in TM-F was narrower than that
in TM-F.
The overall differences between TM-F and TM-F are shown in Figure 4, where subject A1
has been excluded from caclucation. As mentioned above, subject A1’s processing speeds for
fuzzy-match categories were much higher than the remainder of the participants in the same
group. Further investigation of this translator’s processing is needed. Given this, however,
Figure 4 still provides us with an overview of the productivity difference between TM-F and
TM-L.
Figure 4: Average speed for fuzzy/no-match categories
Figure 3 shows that the production speed with TM-L is equal to or higher than with TM-F in all
categories. Especially in the 99-90% match category, the speed for TM-F was significantly
lower, at approximately half that of TM-L.
Concluding remarks
In sum, the TM-L production speed for fuzzy match segments exhibited faster WPM than did
work with TM-F. That is, if a TM content is highly customized or localized as in TM-F, it may
reduce productivity.
The reason for the reduced speed has not been analyzed in this paper. It may be related to
the translator’s focus range or translation unit. Under the TM-F condition, where the target
renditions contained many deletions and additions, translators require more effort to recognize
one-to-one correspondence between the source and the target text. Because the use of TM
restricts the translator’s segmentation range to a sentence or smaller unit, chunk-level
recognition would be more difficult when using the free translation content (TM-F). This in turn
may imply that translators using a TM are actually working on a sub-segment unit rather than an
entire sentence or the discourse level.
70
In terms of efficiency in the actual practice of localization, if free segments are put into the
TM database, there is a chance that this may adversely affect the translator’s performance. The
freer the renditions in the TM, the less effective the localizability may be. In order to improve
the efficiency, it is necessary to review both the project workflow and the TM database, because
the TM databases, like translators themselves, are no longer isolated from the project: they are
part of the localization team.
References
Bowker, Lynne. 2005. “Productivity vs Quality: A pilot study on the impact of translation
memory systems”. Localisation Focus 4(1): 13-20.
Dragsted, Barbara. 2004. Segmentation in Translation and Translation Memory Systems: An
empirical investigation of cognitive segmentation and effects of integrating a TM system
into the translation process. Doctoral dissertation, Copenhagen Business School:
Samfundslitteratur.
García, Ignacio. 2009. “Beyond Translation Memory: Computers and the professional
translator”. The Journal of Specialised Translation 12: 199-214.
Guerberof, Ana. 2009. “Productivity and quality in the post-editing of outputs from translation
memories and machine translation”. Localisation Focus 7(1): 11-21.
Japan Translation Federation. 2009. Honyaku Hakusho 2009 [Translation White Paper 2009].
Tokyo: Japan Translation Federation.
Lagoudaki, Elina. 2006. Translation Memories Survey 2006: Users’ perceptions around TM
use. http://www.atril.com/docs/tmsurvey.pdf. Visited May 2010.
O’Brien, Sharon. 1998. “Practical experience of computer-aided translation tools in the
localization industry”. Lynne Bowker, Michael Cronin, Dorothy Kenny and J. Pearson
(eds) Unity in Diversity?: Current Trends in Translation Studies. Manchester: St. Jerome.
115-22.
Ribas, Carlota. 2007. Translation Memories as vehicles for error propagation. A pilot study.
Minor Dissertation. Tarragona: Intercultural Studies Group, Universitat Rovira i Virgili.
Somers, Harold. 2003. “Translation memory systems”. Harold Somers (ed.) Computers and
Translation: A Translator’s Guide. Amsterdam and Philadelphia: Benjamins. 31-47.
Tirkkonen-Condit, Sonja. 1990. “Professional vs. Non-Professional Translation: A Think-Aloud
Protocol Study”. M.A.K. Halliday, J. Gibbons, and H. Nicholas (eds) Learning, Keeping
and Using Language. Amsterdam: John Benjamins. 381-394.
... The use of TM is very necessary for computeraided translation (Yamada, 2011) and computational approaches for machine translation (Koehn and Senellart, 2010). Similar sentence pairs retrieved from a TM are also utilized as a type of knowledge to enhance the translation (Liu et al., 2019a;He et al., 2021;Khandelwal et al., 2021). ...
... Translators wishing to translate a sentence can benefit from fuzzy matching techniques to retrieve similar segments from the TM. These segments can then be revised, thereby improving productivity and consistency of the translation process (Koehn and Senellart, 2010;Yamada, 2011). The retrieval of similar examples from a TM has also proved useful in conventional (AR) neural MT systems; they can be injected into the encoder (Bulte and Tezcan, 2019; or as priming signals in the decoder (Pham et al., 2020) to influence the translation process. ...
... Translation memory (TM) is basically a database of segmented and paired source and target texts that translators can access in order to re-use previous translations while translating new texts (Christensen and Schjoldager, 2010). For human translators, such similar translation pieces can lead to higher productivity and consistency (Yamada, 2011). For machine translation, early works mainly contributes to employ TM for statistical machine translation (SMT) systems (Simard and Isabelle, 2009;Utiyama et al., 2011;Liu et al., 2012. ...
... Motivated by the use of Translation Memory in Computer-Aided Translation (Yamada, 2011) and its usage in computational approaches to Machine Translation (Somers, 1999;Koehn and Senellart, 2010;Khandelwal et al., 2020, inter alia), we retrieve similar examples to the test source from a datastore that includes pairs of the source text and their corresponding translations via BM25, an unsupervised efficient retriever to provide additional context to the model. We propose a novel incontext example selection and re-ranking strategy to maximize the coverage of the source n-grams in the retrieved examples. ...
... Since the world is not a snapshot once the training corpus is collected, we can never expect an everlarge model to capture everything in its parameters, even for LLMs like GPT-3 (Brown et al., 2020), and it is important to endow the model with access to an external memory bank to solve different NLP tasks (Lewis et al., 2020c). For translation task, long before machine translation, the localization industry has been proposing retrieval techniques to help human translators to get higher productivity and consistency (Yamada, 2011). Early works on machine translation mainly contribute to employing memory for statistical machine translation (SMT) systems (Simard and Isabelle, 2009;Liu et al., 2012). ...
Preprint
Full-text available
With direct access to human-written reference as memory, retrieval-augmented generation has achieved much progress in a wide range of text generation tasks. Since better memory would typically prompt better generation~(we define this as primal problem), previous works mainly focus on how to retrieve better memory. However, one fundamental limitation exists for current literature: the memory is retrieved from a fixed corpus and is bounded by the quality of the corpus. Due to the finite retrieval space, bounded memory would greatly limit the potential of the memory-augmented generation model. In this paper, by exploring the duality of the primal problem: better generation also prompts better memory, we propose a framework called Selfmem, which iteratively adopts a retrieval-augmented generator itself to generate an unbounded memory pool and uses a memory selector to pick one generated memory for the next generation round. By combining the primal and dual problem, a retrieval-augmented generation model could lift itself up with its own output in the infinite generation space. To verify our framework, we conduct extensive experiments across various text generation scenarios including neural machine translation, abstractive summarization and dialogue generation over seven datasets and achieve state-of-the-art results in JRC-Acquis(four directions), XSum(50.3 ROUGE-1) and BigPatent(62.9 ROUGE-1).
... For decades, the localization industry has proposed TM technologies in CAT tools to allow human translators to visualize one or several similar or usable translations extracted from the TM when translating a sentence, leading to higher productivity and consistency (Yamada, 2011). Hence, even though the retrieval methods of TM differ among CAT tools (Bloodgood and Strauss, 2014), human translators generally accept discounted translation rates for sentences with high fuzzy matches. ...
Thesis
In an increasingly global world, more situations appear where people need to express themselves in a foreign language or multiple languages. However, for many people, writing in a foreign language is not an easy task. Machine translation tools can help generate texts in multiple languages. With the tangible progress in neural machine translation (NMT), translation technologies are delivering usable translations in a growing number of contexts. However, it is not yet realistic for NMT systems to produce error-free translations. Therefore, users with a good command of a given foreign language may find assistance from computer-aided translation technologies. In case of difficulties, users writing in a foreign language can access external resources such as dictionaries, terminologies, or bilingual concordancers. However, consulting these resources causes an interruption in the writing process and starts another cognitive activity. To make the process smoother, it is possible to extend writing assistant systems to support bilingual text composition. However, existing studies mainly focused on generating texts in a foreign language. We suggest that showing corresponding texts in the user's mother tongue can also help users to verify the composed texts with synchronized bitexts. In this thesis, we study techniques to build bilingual writing assistant systems that allow free composition in both languages and display synchronized monolingual texts in the two languages. We introduce two types of simulated interactive systems. The first solution allows users to compose mixed-language texts, which are then translated into their monolingual counterparts. We propose a dual decoder Transformer model comprising a shared encoder and two decoders to simultaneously produce texts in two languages. We also explore the dual decoder model for various other tasks, such as multi-target translation, bidirectional translation, generating translation variants, and multilingual subtitling. The second design aims to extend commercial online translation systems by letting users freely alternate between the two languages, changing the texting input box at their will. In this scenario, the technical challenge is to keep the two input texts synchronized while taking the users' inputs into account, again with the goal of authoring two equally good versions of the text. For this, we introduce a general bilingual synchronization task and implement and experiment with autoregressive and non-autoregressive synchronization systems. We also investigate bilingual synchronization models on specific downstream tasks, such as parallel corpus cleaning and NMT with translation memories, to study the generalization ability of the proposed models.
... Translation memory (TM) is basically a database of segmented and paired source and target texts that translators can access in order to re-use previous translations while translating new texts (Christensen and Schjoldager, 2010). For human translators, such similar translation pieces can lead to higher productivity and consistency (Yamada, 2011). For machine translation, early works mainly contributes to employ TM for statistical machine translation (SMT) systems (Simard and Isabelle, 2009;Utiyama et al., 2011;Liu et al., 2012. ...
Preprint
Retrieval-augmented Neural Machine Translation models have been successful in many translation scenarios. Different from previous works that make use of mutually similar but redundant translation memories~(TMs), we propose a new retrieval-augmented NMT to model contrastively retrieved translation memories that are holistically similar to the source sentence while individually contrastive to each other providing maximal information gains in three phases. First, in TM retrieval phase, we adopt a contrastive retrieval algorithm to avoid redundancy and uninformativeness of similar translation pieces. Second, in memory encoding stage, given a set of TMs we propose a novel Hierarchical Group Attention module to gather both local context of each TM and global context of the whole TM set. Finally, in training phase, a Multi-TM contrastive learning objective is introduced to learn salient feature of each TM with respect to target sentence. Experimental results show that our framework obtains improvements over strong baselines on the benchmark datasets.
... Furthermore, motivated by the use of Translation Memory in Computer-Aided Translation (Yamada, 2011) and its usage in computational approaches to Machine Translation (Somers, 1999;Koehn and Senellart, 2010;Khandelwal et al., 2020, arXiv:2212.02437v1 [cs.CL] 5 Dec 2022 inter alia), we retrieve similar examples to the test source from a datastore that includes pairs of source text and their corresponding translations via BM25, an unsupervised efficient retriever to provide additional context to the model. ...
Preprint
Full-text available
Large-scale generative models show an impressive ability to perform a wide range of Natural Language Processing (NLP) tasks using in-context learning, where a few examples are used to describe a task to the model. For Machine Translation (MT), these examples are typically randomly sampled from the development dataset with a similar distribution as the evaluation set. However, it is unclear how the choice of these in-context examples and their ordering impacts the output translation quality. In this work, we aim to understand the properties of good in-context examples for MT in both in-domain and out-of-domain settings. We show that the translation quality and the domain of the in-context examples matter and that 1-shot noisy unrelated example can have a catastrophic impact on output quality. While concatenating multiple random examples reduces the effect of noise, a single good prompt optimized to maximize translation quality on the development dataset can elicit learned information from the pre-trained language model. Adding similar examples based on an n-gram overlap with the test source significantly and consistently improves the translation quality of the outputs, outperforming a strong kNN-MT baseline in 2 out of 4 out-of-domain datasets.
... For this reason, translators need training on using TMs to gain an ideal benefit from the functions of them (p.13). Later, Yamada (2011) investigated the effect of TMs with different features (e.g. free and literal ones) on translator's productivity within the context of localization and concluded that localizability might be adversely affected if the TM contained "freer renditions" and this might in turn reduce the productivity (p. ...
Article
Full-text available
Machine-translated segments are increasingly included as fuzzy matches within the translation-memory systems in the localisation workflow. This study presents preliminary results on the correlation between these two types of segments in terms of productivity and final quality. In order to test these variables, we set up an experiment with a group of eight professional translators using an on-line post-editing tool and a statistical-based machine translation engine. The translators were asked to translate new, machine-translated and translation-memory segments from the 80-90 percent value range using a post-editing tool without actually knowing the origin of each segment, and to complete a questionnaire. The findings suggest that translators have higher productivity and quality when using machine-translated output than when processing fuzzy matches from translation memories. Furthermore, translators' technical experience seems to have an impact on productivity but not on quality.
Article
Full-text available
Translation has historically been performed by bilinguals equipped with specialised topic knowledge. In the mid 20 th century, textual theory and discourse analysis saw emphasis on a top-down, whole-text approach that paved the way for modern professional translators as linguistic transfer experts. This professionalisation was further driven by the digital revolution in the 90s which caused a huge increase in translation demand, and the creation of purpose-designed translation tools—principally translation memory (TM). However, the same technological processes that briefly empowered the professional translator also signalled a return to a bottom-up approach by concentrating on the segment. Twenty years on, translation tools and workflows continue to narrow this focus, even tending towards simple post-editing of machine translated output. As a result, topic-proficient bilinguals are again entering mainstream translation tasks via simplified translation management processes and crowdsourcing approaches. This article explores these recent trends and predicts that, over the next decade, professional translators will find it increasingly difficult to survive as linguistic transfer experts alone.
Segmentation in Translation and Translation Memory Systems: An empirical investigation of cognitive segmentation and effects of integrating a TM system into the translation process. Doctoral dissertation
  • Barbara Dragsted
Dragsted, Barbara. 2004. Segmentation in Translation and Translation Memory Systems: An empirical investigation of cognitive segmentation and effects of integrating a TM system into the translation process. Doctoral dissertation, Copenhagen Business School: Samfundslitteratur.
Translation Memories Survey
  • Elina Lagoudaki
Lagoudaki, Elina. 2006. Translation Memories Survey 2006: Users' perceptions around TM use. http://www.atril.com/docs/tmsurvey.pdf. Visited May 2010.
Practical experience of computer-aided translation tools in the localization industry
  • Sharon O'brien
O'Brien, Sharon. 1998. "Practical experience of computer-aided translation tools in the localization industry". Lynne Bowker, Michael Cronin, Dorothy Kenny and J. Pearson (eds) Unity in Diversity?: Current Trends in Translation Studies. Manchester: St. Jerome. 115-22.
Translation Memories as vehicles for error propagation. A pilot study. Minor Dissertation. Tarragona: Intercultural Studies Group
  • Carlota Ribas
Ribas, Carlota. 2007. Translation Memories as vehicles for error propagation. A pilot study. Minor Dissertation. Tarragona: Intercultural Studies Group, Universitat Rovira i Virgili.
Computers and Translation: A Translator's Guide. Amsterdam and Philadelphia: Benjamins
  • Harold Somers
Somers, Harold. 2003. "Translation memory systems". Harold Somers (ed.) Computers and Translation: A Translator's Guide. Amsterdam and Philadelphia: Benjamins. 31-47.