Synthesis and discussion

In what follows, the results of the analyses that address RQ1 and RQ1a are synthesized and discussed. To answer RQ1, based on a process analysis, all matches in the FAQ text and the Newsletter were categorized as belonging to the accept, reject or revise category. First, quantitative results for this categorization were presented. The analysis showed that, in the FAQ text, CM matches (100%) and 100% matches (94%) were mostly accepted which was also to be expected as the source segments are identical to the source segments stored in the TM. 95-99% matches were mostly revised (50%) or accepted (43%). Contrary to expectation, 85-94% matches were mostly rejected (55%). This, however, appeared to be

caused by three specific source text segments in the FAQ text which could be transferred directly to the target text without changes, and this appeared to make all translators use the Copy Source to Target function to reject these matches. 75-84% matches were mostly revised (90%). The same was the case for 70-74% matches (71%), although the remaining 29% of these matches were rejected. 79% of MT matches were revised, 12% were rejected and 9% were accepted. The fact that 9% of the provided MT matches were accepted is interesting from a productivity perspective, since these were instances where the MT engine produced suggestions that were acceptable to the translators without changes. In terms of the TM/MT threshold, the analysis of the FAQ text indicated that it might be preferable to set the threshold higher than 70%, since the translators deemed 29% of 70-74% matches not worth revising compared to 12% of MT matches. However, the 29% include one 70% match which was rejected by 6 of the translators. In any case, this needs to be related to results on, for instance, editing speed and thus needs further exploration. The Newsletter contained no CM matches, but the translators mostly accepted the 100% matches (86%). 95-99%, 85-94%, 75-84%, 70-74% and MT matches were all mostly revised (100%, 79%, 94%, 95% and 93%, respectively). Very few matches were rejected in the Newsletter, and 4% of MT matches were accepted without changes. Regarding the TM/MT threshold, there were only small differences between the translators’ choices in MT matches and low fuzzy matches. Thus, on the basis of this analysis, this might indicate that these matches are of a comparable quality.

However, as mentioned above, this needs to be seen in the light of results for editing speed, for example. If we compare the results of the categorization of matches in the two texts, we see that 100% matches are mostly accepted in both texts and that TM fuzzy matches (with one exception) and MT matches are mostly revised in both texts. However, whereas all TM fuzzy match types and MT matches were sometimes rejected in the FAQ text, only 75-84%

and MT matches were rejected in the Newsletter and only on a few occasions.

RQ1a was addressed through an analysis aimed at identifying characteristics of the translators’ interaction with the MT-assisted TM tool in the accept, reject and revise categories. Concerning the accept category, analysis of the FAQ text showed that the translators were typically able to accept matches without carrying out any observable research to verify the provided matches. The same was the case in the Newsletter. The accepted fuzzy TM matches and MT matches had different characteristics. For instance, some were acceptable condensed translations of the source-text segments, some contained what the translators appear to have deemed insignificant differences compared to the source-text segments, some were acceptable direct transfers of source-text segments and some were accepted although they did not appear to be entirely acceptable translations.

With respect to the reject category, it was interesting to note that in both the FAQ text and the Newsletter, no matches were rejected by use of “pure” deletion, i.e. where the

translators delete the match letter by letter or use the mouse or the keyboard to select the match and delete it all at once. This goes against what is typically assumed in the CAT literature, namely that when translators reject matches, they translate the source segment from scratch (cf. e.g. Bowker & Fisher 2010, p.61; Kenny 2011, p.467; Garcia 2015, p.81). The present study shows that this is not the case: in the FAQ text, the translators mostly rejected matches using the Copy Source to Target function, i.e. by replacing the match with the

source segment and then using this as the point of departure for their translation. In one case, a translator rejected a TM match by replacing it with an MT match. In the Newsletter, two matches were rejected because the Copy Source to Target function was used, and two because the TM matches were replaced by MT matches. It was also interesting to note that, in the FAQ text, all of the rejected matches had one of the following characteristics: 1) the entire source segment could be transferred directly to the target text segment and used as the translation in an unchanged form, 2) the source segment contained one or more red words, which should remain untranslated in the target text, or 3) the source segment contained a tag indicating the presence of a visual element. In the Newsletter, the source segment for the match, which two translators rejected by use of the Copy Source to Target function, also contained elements that could be transferred directly to the translation. Thus, the study suggests that certain elements such as source-text items which can be directly transferred to the translation, formatting and tags trigger the rejection of matches using the Copy Source to Target function. With respect to MT matches, this finding is also interesting in the context of Moorkens and O’Brien’s (2013) study which found that 81% of the

participating translators would like to have a shortcut for “one-click rejection of MT suggestion”. Using the Copy Source to Target function is, in effect, a one-click rejection of the MT match, but translators might also appreciate the possibility of quickly deleting the match and then translating from scratch.

Concerning the revise category, quantitative results on the distribution of the revised matches into match-internal and match-external revision were first presented for both texts.

In the FAQ text, for all translators combined, match-internal revision was more frequent than external revision in 100% and 95-99% matches. For 85-94% matches, match-internal and match-external revision were almost equally frequent, and for 75-84% matches internal revision was most frequent. For both 70-74% and MT matches, match-external revision was most frequent. Thus, in terms of the FAQ text, the study suggests that translators primarily rely on their own judgement when editing TM matches with match values from 75% and up, whereas in 70-74% matches and MT matches they need other support than the proposed matches to produce an acceptable translation of the match in question. For 70-74% matches, match-external support was sought in 93% of cases, whereas it was 62% for MT matches. This is interesting as it supports the point mentioned above, namely that it might be advisable to set the TM/MT threshold higher than 70%. However, this would need to be explored further due to the low amount of data in the 70-74% match type and should be combined with results for other factors such as editing speed. In the Newsletter, in all TM match types, internal revision was more frequent than match-external revision, whereas match-match-external revision was more frequent in MT matches. Thus, in terms of the Newsletter, the study suggests that translators rely on their own judgement in TM matches and that in MT matches, they more frequently make use of resources or functionalities external to the match.

Next, the analysis explored the subcategories match-internal and match-external revision in order to contribute to identifying characteristics of the translators’ interaction with the MT-assisted TM tool. In terms of match-internal revision, focus was on the instances where the translators were provided with suggestions from the AutoSuggest function. The analysis

showed that in both the FAQ text and the Newsletter, the suggestions were seldom employed by the translators. This is interesting from a TCI perspective since it leads to speculations as to whether AutoSuggest suggestions are overall more of a hindrance to successful and frictionless interaction with the tool than they are a help. Ehrensberger-Dow and Heeb (2016) also found that the translator participating in their study tended to ignore the AutoSuggest suggestions, also in cases where she ended up typing the same word as the suggestion. The finding seems to be in line with the study by O’Brien et al., who found that when shown the AutoSuggest feature, translators reported that they “thought the feature could be useful, but also expressed the opinion that they would like to have the option of turning it off” (2010, p.3). This finding thus calls into question expectations that auto-completion functions such as AutoSuggest are “probably the most productive […] way of subsegment matching” (Reinke 2013, p.33). However, possible explanations for the relative lack of usefulness of the AutoSuggest function might also be that the sources from which the function retrieves the suggestions were not sufficiently large to predict successfully what the translators were writing, that translators prefer typing their translations themselves, and/or that they even prefer to ignore the suggestions.

In terms of match-external revision, the analysis showed that the translators used a number of different external actions. In both texts, the translators used the actions concordance search, Copy/Cut à Copy Source to Target à Insert, termbase search, Google search, search in online dictionary, search in local dictionary and pasting element copied from another segment. In the FAQ text, in addition to the actions already mentioned, the translators used the external actions Copy Source to Target, reference text and pasting element copied from the source segment. In the Newsletter, in addition to concordance search, Copy/Cut à Copy Source to Target à Insert, termbase search, Google search, search in online dictionary, search in local dictionary and pasting element copied from another segment, the translators used the external action Web page.

Examining all seven translators’ use of external actions revealed that, in the FAQ text, in 100% matches, the actions reference text and pasting element copied from another segment were most frequent. In 95-99% matches, the action Google search was most frequently used. In 85-94% matches, the external action concordance search was most frequently used, whereas the Copy/Cut à Copy Source to Target à Insert action was most frequently used in 75-84% matches. In 70-74% matches, these two actions were used in an equal number of cases. In MT matches, concordance search was the external action most frequently used. In the Newsletter, only one external action was used in 100% matches namely search in a local dictionary. No external actions were used in 95-99% matches. In the remaining match types (85-94%, 75-84%, 70-74% and MT matches), the action concordance search was most frequent. In both texts, a wider range of external actions was used in MT matches than in the other match types. In the following, the use of selected external actions is commented upon.

Analysis of the individual translators’ use of external actions showed that, in both the FAQ text and the Newsletter, concordance search was the only external action used by all

translators, and for most of the translators, it was the action most frequently used. Thus, the

study suggests that concordance search is the preferred external action for translators. This result ties in with Karamanis et al. (2010; 2011), for example, who found that the

concordance search is the primary resource used by translators when they encounter a translation problem. As noted above, frequent use of the concordance search might be motivated by the translators wanting to check the MT suggestions against the contents in the TM, i.e. against texts translated for the same client. Zapata (2016, p.142) notes a similar kind of behaviour in his study of translators’ use of a biconcordancer during post-editing, namely that translators often double-checked solutions proposed by an MT system.

The Copy/Cut à Copy Source to Target à Insert action was used by four of the translators in the FAQ text and by one translator in the Newsletter. The analysis showed that this action always occurred in segments where the source text contained either a word that could be transferred directly to the target text (formatted in red) or a tag indicating the presence of a visual element that had not been transferred to the match. Thus, it seems that the Copy/Cut à Copy Source to Target à Insert action is an example of what Risku refers to as iterative operation patterns, namely “observable, iterative, regular patterns that stand out as behavioural patterns in the way the translator works” (2014b, p.348), since this action was applied in similar situations and by the same translators, who to a varying extent seemed to apply this action as a cognitive routine and who each either used copy or cut. This point is further supported by the observation that, in the FAQ text, for the four translators in question, a number of matches either were rejected using the Copy Source to Target function, or the Copy/Cut à Copy Source to Target à Insert action was used during match-external revision. The Newsletter did not contain any words formatted in red, nor did it contain formatting or tags. With this in mind, it was interesting to note that the Copy/Cut à Copy Source to Target à Insert action was used only once in the Newsletter. This further supports the aforementioned suggestion that certain elements (words which can be transferred directly to the target segment, formatting and tags) trigger this action. Also, in the Newsletter, only two segments were rejected using the Copy Source to Target function, which points in the same direction.

The action pasting element copied from the source segment was only used in the FAQ text. It is noteworthy that in all cases, the “element” consisted of one or more words that should remain untranslated in the target text and/or of a tag indicating the presence of a visual element. Moreover, the two translators with most actions of this type (A and H) were two of the translators who did not use the Copy/Cut à Copy Source to Target à Insert action, indicating that the pasting element copied from the source segment action might be an iterative operation pattern for Translators A and H similar to how the Copy/Cut à Copy Source to Target à Insert action seemed to be an iterative operation pattern for Translators C, D, E and G. The pasting element copied from the source segment action was not used in the Newsletter.

In both texts, more than half of the translators carried out Google searches. In the FAQ text, none of these searches was followed by a translator visiting a Web page that came up as the result of these searches. This was, however, the case in a few instances in the Newsletter.

Thus, the study shows that in most cases, translators find it sufficient to use Google to check

possible Danish translations of English source-text items and check the frequency of certain Danish phrases and that only in a few cases do they go beyond Google and carry out further research. Thus, these searches were considered examples of what Jimenéz-Crespo refers to as using the “Web as Corpus” (2015, p.47). The actions termbase search, search in online dictionary and search in local dictionary were not frequent in any of the texts.

By juxtaposing the process analysis with the translators’ comments in the retrospective interviews, this enhanced the ability of the process examples given throughout the analysis to provide deeper insight into the translators’ interactions with the MT-assisted TM tool.

The examples made it possible to follow the unfolding of the translation, and the comments from the retrospective interviews on the translators’ thoughts about their processes and arguments for their solutions helped to clarify these processes. For example, some examples suggested that the translators tried to avoid unnecessary typing, while others illustrated how translators dealt with what they seemed to experience as translation problems by using different match-external resources to reach a terminological decision. Also, a number of examples showed that the arguments presented by translators for their translation choices related to the context in which the target text was intended to be used. The translators also pointed out some undesirable issues regarding MT, for example that words were left out of the MT matches, and the analysis also suggested that integration of the MT engine with the termbase caused problems in places. These and other issues are further addressed in Section 5.8 concerning RQ7. Finally, several of the examples indicated that the translation processes were not linear in the sense that the translators returned to a previous segment during the editing of an exemplified segment or returned to an exemplified segment later in the process. This is explored further in the analysis in Section 5.4.