A user study of neural interactive translation prediction

Machine translation (MT) on its own is generally not good enough to produce high-quality translations, so it is common to have humans intervening in the translation process to improve MT output. A typical intervention is post-editing (PE), where a human translator corrects errors in the MT output. Another is interactive translation prediction (ITP), which involves an MT system presenting a translator with translation suggestions they can accept or reject, actions the MT system then uses to present them with new, corrected suggestions. Both Macklovitch (2006) and Koehn (2009) found ITP to be an efficient alternative to unassisted translation in terms of processing time. So far, phrase-based statistical ITP has not yet proven to be faster than PE (Koehn 2009; Sanchis-Trilles et al. 2014; Underwood et al. 2014; Green et al. 2014; Alves et al. 2016; Alabau et al. 2016). In this paper we present the results of an empirical study on translation productivity in ITP with an underlying neural MT system (NITP). Our results show that over half of the professional translators in our study translated faster with NITP compared to PE, and most preferred it over PE. We also examine differences between PE and ITP in other translation productivity indicators and translators’ reactions to the technology.


Introduction
Interactive translation prediction (ITP) serves to allow translators to work with the output of machine translation (MT) systems by using it like an "auto-complete" feature.Rather than starting with a complete (but likely erroneous) translation which they then must post-edit (PE), a translator using ITP guides the translation process.They can accept a suggestion with a single keystroke, or reject it by typing an alternate translation.When a suggestion is rejected, the MT system recomputes its predictions from the given prefix and presents its new suggestions to the translator.
As described in Wuebker et al. (2016) and Knowles and Koehn (2016), this is done in neural interactive translation prediction (NITP) by feeding the translator's token(s) into the neural machine translation (NMT) model as conditioning context (rather than feeding in the rejected system predictions), then producing the rest of the translation token by token.Using reference text to simulate translators, both papers show that NITP outperforms ITP systems that are based on phrase-based statistical MT even when the underlying MT systems are of similar quality.
In this work, we investigate the use of NITP through a user study with professional English-Spanish translators.We integrate an NITP system into a web-based translation workbench (Fig. 1) and conduct a user study with eight professional translators.We find that most translators in our study prefer NITP to PE, and most would be willing to use it in their work.Over half of the translators translated faster with NITP than PE, but we do not find a significant difference between translation speed with NITP and PE overall.We provide some analysis of translator reactions to the tool, including a discussion of the potential relationship between translator experience with PE and their reactions to ITP assistance.

Related work
Our work focuses on neural interactive translation prediction.However, the earliest body of work on ITP, including the TransType and TransType2 projects (Langlais et al. 2000;Foster et al. 2002;Bender et al. 2005;Barrachina et al. 2009), predates the current wave of neural approaches to MT.Following those projects, approaches using static search graphs were proposed to allow for ITP from phrase-based statistical MT.In the search graph approach, the system seeks to find a match for the prefix (the partial translator input) in the search graph, backing off to edit-distance techniques when exact matches are not found (Koehn 2009;Koehn et al. 2014).An alternative approach for statistical phrase-based MT is to use constrained decoding (Wuebker et al. 2016).The proposed approach for NITP (described in more detail in the following sections) was introduced in Wuebker et al. (2016) and Knowles and Koehn (2016), who found in simulations that NITP outperformed phrase-based statistical ITP approaches, in terms of their accuracy in predicting the next word after a translator-generated prefix.Some computer-aided translation workbenches, including the research environment Fig. 1 Interactive translation prediction in casmacat: The system suggests continuing the translation with algoritmo puede escribir, which the user can accept by pressing the tab key casmacat1 (see Fig. 1) and the commercial tool Lilt2 contain implementations of ITP.
The effect of ITP on translation productivity has been assessed through simulations and empirical studies, with much of the focus placed on processing time and technical effort (relative to unassisted translation or PE).Macklovitch (2006) found that the TransType2 ITP research system increased translators' productivity in terms of processing time (relative to unassisted translation) by about 15-20%, and produced texts of comparable quality, while Barrachina et al. (2009), in a simulated setting, showed that ITP had the potential to reduce typing effort between about 55% and 80%.Koehn (2009) found that, relative to unassisted translation, both ITP and PE produced better quality, faster translations, but that ITP did not yet yield time gains to the level of PE.Similar findings of translators being slower overall in ITP than in PE are reported in Underwood et al. (2014), Green et al. (2014), Sanchis-Trilles et al. (2014), Alabau et al. (2016) and Alves et al. (2016).The number of participants in these research studies ranged from 5 to 32; language pairs investigated were English to Spanish, Portuguese, and German, and French to English.Except for Green et al. (2014), all studies were conducted on casmacat.Findings on the keystroke activity involved in ITP, however, are somewhat contradictory, with Sanchis-Trilles et al. (2014) finding it to be lower in ITP, and Alves et al. (2016) obtaining the opposite result.This is likely due to differences in how the interactive functionality was implemented, with the former producing a translation of the entire sentence-instead of word-by-word suggestions that need to be confirmed-every time a keystroke was made.Findings on cognitive effort are also mixed, with Alves et al. (2016) reporting that ITP involved more gaze fixation counts than PE, but that their total duration was lower, and Underwood et al. ( 2014) reporting that gaze duration across conditions was similar, though more gaze attention was placed on the target text than on the source text in the ITP condition.As for final translation quality, Alves et al. (2016) and Underwood et al. (2014) found ITP and PE to result in comparable quality measured in terms of edit distance, while Green et al. (2014) found that translations done using ITP yielded slightly higher BLEU scores than those done in PE.Finally, in terms of translator satisfaction, and unlike Macklovitch (2006) or Underwood et al. (2014), Koehn (2009) found that, overall, translators preferred ITP over PE.

Neural machine translation
Here we focus on an encoder-decoder with attention (one commonly used neural architecture), as described in Bahdanau et al. (2015) and implemented in the Nematus NMT tool (Sennrich et al. 2017).We use byte-pair encoding (BPE; Sennrich et al. 2016) to perform translation at the subword level.In the encoder, the preprocessed input sentence is passed through two recurrent neural networks (one left-to-right, one right-to-left), which are then concatenated together such that the hidden state (h t ) associated with each input token (x t ) contains information about that token and its full input sentence context.The decoder produces the output sentence one token at a time (left to right), conditioned on the previously produced tokens and an attention mechanism, which serves as a soft alignment to the representations of the input.
At step t, the conditional probability of generating output token y t given the full input sequence x and the previously output tokens ŷ1 , . . ., ŷt−1 is: where g is a non-linearity and c t and s t are the context vector and hidden state, respectively.The vector c t is a weighted average of all encoder hidden states h j , with weights generated by the attention mechanism.

Neural interactive translation prediction
In NITP, instead of conditioning the prediction of each token on the previous model predictions { ŷ1 , . . ., ŷt−1 } (as is done in standard NMT decoding), we condition on the true translator-generated prefix {y * 1 , . . ., y * t−1 }.This results in a new conditional probability equation: That is, the conditioning context is now the one produced by the human translator rather than the one produced by the MT system's predictions.In practice, we generate more than just the next predicted token to show to the translator, so it is better described as follows: given a translator prefix of length m, and some number n of next tokens which we wish to show to the translator, we have two equations.
In Eq. 3, we see that the word immediately following the user-generated prefix is conditioned on the user-generated prefix.In Eq. 4, we see that all subsequent words are conditioned on a user-generated prefix followed by predicted words (until such time as the translator accepts or rejects them).
If a translator rejects a suggestion and provides their own, there are two possible cases: either the translator has added a complete word to the translation, or they have added a partial word.In the case of a complete word, we follow Eqs. 3 and 4. That word becomes part of the prefix, and the generation of the subsequent tokens is conditioned on it.
If, however, the translator has only generated a partial word (which we will call a character prefix), this is slightly more complicated.We provide some additional technical detail here.We must first determine whether this character prefix is the prefix to any item in our (subword) vocabulary.If it is the prefix of at least one vocabulary item, we predict the completion to this word (or subword) by selecting the highest probability item in the vocabulary that starts with our character prefix (this can be described as a modification to the softmax and/or as a mask applied to the distribution prior to performing the softmax).Given the character prefix r * : where 1(y t ) = 1 if y t starts with the string r * 0 otherwise (6) We then continue predicting the remaining tokens in the standard fashion.
In the case that the character prefix is not the prefix to any item in our vocabulary, we must first apply BPE to it. 3Once BPE has been applied, we have the model consume (forced decode) all but the last subsegment.This last subsegment could be a complete vocabulary item on its own, or again a prefix to a vocabulary item.Thus we return to our approach of predicting the highest probability vocabulary item which has the last subsegment as a prefix, and then continue prediction.This approach, as described in the Letter Prediction Accuracy section of Knowles and Koehn (2016), eliminates the need to further modify our decoder, while maintaining the character-level interactions expected in ITP.Knowles and Koehn (2016) also propose speed-related improvements, which we discuss in Appendix B.

NITP system and study setup
We integrated an implementation of NITP based on Nematus into the open-source casmacat translation workbench (Alabau et al. 2014), which uses a similar layout and keyboard combinations to many commercial CAT tools (albeit without some common features like spell check, integrated dictionary or concordancer functionality) and which was also used in a number of the studies described in Sect. 2. We then conducted a longitudinal empirical study with a threefold purpose: (a) comparing translation productivity in ITP with that of PE; (b) investigating whether translation productivity in ITP improved as translators became familiar with the ITP technology, and (c) collecting translators' impressions of ITP.
We trained an English-Spanish NMT model using the attention-based encoderdecoder toolkit Nematus (Sennrich et al. 2017). 4We preprocessed the data using the standard preprocessing scripts: tokenization, truecasing, and BPE (Sennrich et al. 2016).The system is trained on Europarl v7 (Koehn 2005) and News Commentary v10 data, 5 which comprised the WMT13 training data for English-Spanish.This training set contains 3.95 million sentence pairs, over 102 million source tokens, and over 106 million target tokens.We used the WMT12 News Test data for validation.The system has a BLEU score of 29.79 (beam 12, less than 1 BLEU below the best score from WMT13) or 28.40 (beam 1) on the WMT13 test set and a reference-simulated word prediction accuracy of 59.1% (beam 1). 6 Eight Castilian Spanish professional translators (referred to as TrA through TrH) participated in the study.The original sample size was reduced by about 17% due to technical (server down) issues invalidating two translator-session combinations and to TrB producing unusable translation activity data by not adhering to instructions (we nevertheless report data on TrB's background and the feedback provided by him as they may help put said nonadherence in context-see Sect.5.3 for more details).The study consisted of eight sessions spanning four weeks; in the first session, translators engaged in PE (N=201 sentences); in the next seven sessions, they engaged in ITP (N=1349).From the first session we obtained a PE baseline against which we compared translation productivity in the ITP setting; potential learning effects derived from repeated ITP sessions were assessed by examining the indicators collected in the remaining seven sessions.
Eight news texts, controlled for length and syntactic complexity, 7 were selected for the user study.They dealt with a range of topics like politics, technology, business, and life and style.Texts had on average 29.13 sentences (SD = 1.24), 822.75 tokens (SD = 37.48), and a dependency length of 103 (SD = 2.99) and were assigned randomly to translators, while ensuring that each text was presented only once in each session and only once to each translator throughout the study.Translators were asked to produce publication quality translations with two specific guidelines: (1) use as much of the MT output as possible, as in Massardo et al. (2011), and (2) do not engage in preferential changes that do not improve the quality of the text.

Translator interactions
Translators were provided with detailed instructions about the study, including compensation, interaction modes, and translation quality expectations, via participant information sheets and a help page (see Appendix A).Translators conducted a warmup task consisting of PE and interactively translating five sentences prior to the main task, to make sure they familiarized themselves with the translation environment and with the interaction modes.
The casmacat system logs all keystrokes, mouse clicks, and movements between segments in the interface, along with timestamps.The system also logs requests to the translation server, source data, initial translation data, and final translation output produced by the translators.While the underlying translation system vocabulary consists of subword segments, user interactions are performed at the character level (by typing individual characters) and at the whole-word level (by hitting tab to accept a suggestion).All byte-pair operations are performed behind the scenes and are not shown to the user.
In the user interface (UI), shown in Fig. 1 in ITP mode, translators see a source sentence on the left and a space to enter their translation on the right.They translate the document sentence-by-sentence.During PE, the right side is initially populated with MT output, which the translator then edits, as in a standard word processor.During ITP, a floating box to the right and below the translator-produced prefix shows the next three suggested words.The translator can accept a word using the tab key, or type a new word.

Operationalization
Translation productivity was measured through eleven variables in three categories: temporal effort (Processing Time), technical effort (Manual Insertions, Manual Deletions, Navigation and Special Key Presses, Mouse Clicks, and Tokens of MT Origin) and final translation quality (MQM Score, 8 Accuracy Issues, Fluency Issues, Minor Issues, and Major Issues).More specifically, Processing Time was measured in seconds; Manual Insertions and Manual Deletions were the count of alphanumeric characters manually inserted or deleted, respectively, by the translator; Navigation and Special Key Presses were the count of navigation (up, down, left and right), control (ctrl, alt, shift) and tab key presses.Mouse usage was measured via the count of Mouse Clicks.Tokens of MT Origin measured the count of tokens in the final, target text that were accepted by the translators exactly and not changed after being accepted as suggested by the ITP system, or, in the case of PE, that were left unedited (i.e., not altered or moved around).Lastly, the MQM manual annotation framework allowed us to assign to each post-edited/interactively translated sentence a measure of translation quality defined via: (a) an MQM Score (0-100%) according to which a Pass (≥ 95%) or a Fail (< 95%) status was assigned and (b) the frequency of issues, classified according to their type (Accuracy and Fluency) and severity (Minor: issues that do not impede understanding of the text; Major: those that make the text difficult to understand and Critical: those that render the content unusable). 9We also separately examine (but do not model) word prediction accuracy for each translator.
We collected translators' impressions of NITP, through a questionnaire.They rated the following on a 5-level Likert scale: I prefer ITP to PE; ITP is less tiring than PE; As the study progressed, I took better advantage of the ITP suggestions; ITP helps me translate faster than PE; ITP helps me translate to better quality than PE and I would use ITP in real-life scenarios.They also answered open questions: Do you have any suggestions for improvement of any aspect of interactive translation's use? and Please provide any additional comments about your experience with interactive translation prediction.

Results and analysis
We examine our data in three different ways.We begin with a quantitative analysis of our overall sample results, both averaged and broken down by individual translator.We then build mixed-effects models and examine what they can tell us about our data.Finally, we take a look at translators' impressions and feedback about the tool.

Sample results
Table 1 shows summary statistics (mean and standard deviation) for translation productivity indicators, broken down by translation condition.As Table 1 indicates, sample results for eight out of the eleven variables are favorable to ITP.Note that no Critical issues were found in any translation condition, likely due to participants being professional translators.
Exploratory graphs did not show consistent trends over time in ITP in any of the measured variables except for Mouse Clicks, which showed a steady decrease, from the first ITP session (M = 0.34, SD = 0.40), gradually to the last ITP session (M = 0.28, SD = 0.46), possibly indicating that translators change how they interact with the computer in ITP over time.
As Table 2 shows, the effect of ITP on individual translators' productivity indicators varies.All translators made more Navigation and Special Key presses and fewer Manual Deletions in ITP, and all but two (TrC and TrD) made fewer Mouse Clicks in ITP.The increase in Special Key presses is directly attributable to the use of the tab key to accept translation suggestions in the ITP interface.Additionally, all but one translator (TrA) produced texts with more Fluency Issues in ITP, and all but one translator (TrC) produced texts with fewer Adequacy Issues in ITP.
We observe a wide range of word prediction accuracy scores (obtained by rerunning ITP as a simulation on the final translator output) for both ITP and PE, showing (as also shown in the tokens of MT origin) that the usefulness of the sugges-9 Quality annotation was conducted by the second author of this work.tions varies by translator.In all cases, the word prediction accuracy for a translator using ITP is higher than the reference-simulated overall word prediction accuracy (59.1%).While there is not a strict correlation between the positivity of translator reactions to ITP and word prediction accuracy or Tokens of MT Origin, the three translators with the highest word prediction accuracy do agree strongly or agree that they would use ITP in real-life scenarios, while the translator with the lowest word prediction accuracy strongly disagreed.The two translators with the most PE experience would use ITP and have high word prediction accuracy scores, which may suggest that they are adept at using machine translation output in their translations.Four translators were faster in ITP, the same number (though not the exact same set) that applied fewer Manual Insertions and made more use of MT in ITP, as measured by Tokens of MT Origin.10This is similar to earlier studies that have found notable between-translator variation.We discuss potential reasons for variation in Sect.5.3.

Mixed-effect models of translator productivity
Data was analyzed with mixed-effect models, 11 a type of regression model useful for the analysis of grouped data, that describe the effects on a response variable of one or more explanatory variables by incorporating both fixed and random effects.Fixed effects measure population effects, while random effects control for variations in the measured variables across subjects (translators) and items (sentences).Given that our exploratory results showed, as is common in these kinds of studies, between-subject and between-sentence variations in all variables observed, mixedeffect models were deemed appropriate to analyze the data.Additionally, our study presented a number of missing observations (see Sect. 5.3 for details), making mixedeffects models a better choice over rm-ANOVA as the former are better equipped for handling unbalanced designs caused by missing data.Additionally, the inclusion of random effects often leads to more precise estimates of the fixed effects (Fahrmeir 2013).
Translation Condition (PE/ITP) and ITP Session (2 to 8) were treated as fixed effects to address the questions of translator productivity and change over time respectively, with translators and sentences modeled as crossed random effects.To minimize Type I errors, following Barr et al. (2013), the structure of the random effects was kept maximal (all possible random effects that the design justified, and that data allowed, were included, i.e. by-subject and by-item random intercepts and slopes).Where this structure proved too complex it was progressively simplified by removing the random effects with the lowest SD.Untransformed Processing Time was modeled with robust linear mixed-effect models fit with robustlmm (Koller 2016).All other response variables were modeled with generalized linear mixed-effects models, fit with lme4 (Bates et al. 2015).Confidence intervals and p values for the fixed effects were obtained with Wald tests.
Table 3 contains the coefficients estimated by the models addressing translator productivity.Note that ITP is the reference category against which PE is compared, and as such, the point estimates next to PE represent the difference between PE and ITP.All response variables were entered in the models on a by-sentence basis.They were not normalized by sentence length, similarly to Läubli et al. (2013) because the variation introduced by sentence length was already captured by the inclusion of by-sentence random effects.Accordingly, the coefficients, and their transformations if applicable, should be interpreted on a sentence-level.
Note that, while sample results for Processing Time narrowly favor ITP, as shown in Table 1, this result is heavily influenced by TrG logging the slowest Processing Time in PE, as shown in Table 2 in that same section.The robust model presented in Table 3 downweighted the Processing Time observations for TrG in the PE condition the most of all; removing TrG's data would make Processing Time favorable to PE in both sample results (M I T P = 4.34, SD I T P = 3.41; M P E = 4.08, SD P E = 5.28) and model results, with the average sentence in PE taking − 17.15 s to process (CI [− 26.19, −8.11]) than the average sentence in ITP.
As Table 3 shows, ITP significantly decreases Manual Deletions and Mouse Clicks, and significantly increases Navigation and Special Key Presses and Fluency Issues.When thinking of the nature of the ITP and PE tasks, findings relative to temporal and technical effort are intuitive.Translators may automatically insert full translations in ITP without manually deleting any text: they may only need to perform Manual Deletions if they want to change a previously accepted translation suggestion or their own typed translation.In terms of Navigation and Special Key Presses, to insert a one- word MT suggestion in ITP, the translator has to press tab (a Special Key), whereas in PE, the text is already on the target side box.As for Mouse Clicks, translators usually click on the places in the target text where they are going to apply corrections to the text, and, with no initial static text to correct, translators do not need to use the mouse as much as in PE.
Regarding final translation quality, our model indicates that Fluency Issues are more than twice as frequent in ITP as in PE.It should be noted that the implementation of casmacat used in this study did not have a working spell checker, something that very likely contributed to the presence of fluency issues in the final texts, done both in PE and especially in ITP. 12 Specifically, the biggest contributors to Fluency issues in ITP were style (35%) and spelling, i.e., awkward language (34%), followed by minor grammar issues (30%), with the remaining 1% being major grammar issues.In PE, the biggest contributor to Fluency Issues were style issues (54%), followed by minor grammar issues (23%), spelling (15%) and major grammar issues (8%).Note that all spelling and style issues were classified as being of minor severity.Finally, to keep this finding in perspective, it is important to bear in mind that the frequency of Fluency Issues was only one of the levels on which translation quality was measured.In fact, except for one translator-session combination (TrC in the 1st ITP session) the overall MQM score stays consistently above 95%, the minimum arbitrary quality threshold.
In terms of improvement over time, none of the models could determine whether productivity indicators improved over time in ITP, with only Mouse Clicks showing a downward, but not quite significant (p = .06),trend.A look at the standard errors and the width of the confidence intervals of these non-significant models (not included here for reasons of space) shows that, given the potential effect sizes, larger samples would be needed to clarify the nature of the relationships between variables.

Translators' impressions
Translators' impressions of ITP were overall very positive.Out of eight translators, five agreed (TrA, TrC, TrE, TrF) or strongly agreed (TrD) that they preferred translating with ITP assistance over PE.Five translators agreed (TrA, TrC, TrD) or strongly agreed (TrE, TrF) that they would use ITP in real-life translation scenarios.Six translators agreed (TrD, TrE, TrF, TrG,TrH) or strongly agreed (TrA) that they took better advantage of ITP suggestions as the study progressed.Three translators agreed (TrG,TrH) or strongly agreed (TrD) that ITP was less tiring than PE, with one strongly disagreeing (TrB) and the rest giving neutral answers.Translators' perceptions of their own individual translation speed under ITP relative to PE showed a high number (five) of neutral responses, highlighting perhaps the difficulty of making this kind of judgment.Their answers showed differences between translators' perceived and actual quality in both conditions, with only one of the five non-neutral responses matching the annotated translation quality level.
Translators' answers to the open questions reveal a number of valuable insights into various aspects of this study.Two translators (TrB, TrH) considered the speed with which translation suggestions appeared to be a hurdle when translating.While the vast majority of translation suggestions were passed to the interface in under 100 ms, these translators may have encountered a slower translation, experienced network lag, or encountered the end of a full suggestion (end-of-sentence token generated on the backend) without realizing this, and found themselves waiting.We provide additional notes on speed in Appendix B. Four translators (TrB, TrC, TrD,TrH) pointed out the orthographic, grammar, translation, style, and discourse-level issues of the MT suggestions.Three translators (TrA,TrD,TrE) identified desirable UI features such Footnote 12 continued by one translator.Spelling errors (including Spanish vs. Catalan spelling differences) were also introduced naturally by translators.It is possible that translators did not catch these errors before continuing to the next sentence, perhaps due to the lack of spell checker or if they were less thorough than they would typically be in checking their translations.
as keyboard shortcut customization and search and replace options.Three translators (TrC, TrE, TrF) indicated that the varying level of MT quality from sentence to sentence made some MT suggestions for some sentences confusing, which led TrF to opt in such cases for a PE-style solution (i.e., accepting all suggestions and then post-editing the complete sentence).Three translators mentioned how some time had to elapse before making the most out of ITP: "As the experience went on [ITP] helped me finish the tasks in a shorter time and with a higher level of confidence in the quality of my work."(TrA); "By the end of the study I found [ITP] to be a user friendly and straightforward tool" (TrF); "I had the distinct feeling that, on average, the suggestions were more and more spot on as I proceeded"13 (TrD).
Two translators (TrA, TrG) noted the cognitive and translation process differences between ITP and PE, such as ITP resulting in "less time researching terminology" (TrG) and it involving "a mental process different to PE, consisting of constantly comparing ITP's suggestions to the translator's own mental translations, a process that, while seemingly complex, nevertheless sped up translation times" (TrA).Two translators (TrB, TrF) mentioned how not being able to see, in principle, the whole machine-translated text in ITP slowed the overall translation workflow, because otherwise one could decide in one look whether or not the MT output was going to be helpful. 14Finally, two translators (TrA, TrG), expressed their worries about the translator's role and imprint in an MT-centered scenario: how in such scenarios, MT priming means "the voice of the translator is lost" (TrG), and how the user-friendliness and speed of the ITP system may generate overconfidence on the translator side and "lead to mistakes or wrong decisions if the required exigence and rigor levels are not there, on the user's side" (TrA).
Overall, translators' positive feedback towards ITP is consistent with Koehn (2009), and with Langlais et al. (2000).Only one translator (TrB) openly rejected ITP, as evidenced by him strongly disagreeing to all close-ended questions, and expressing negative views in the open questions.TrB chose, against task instructions, to ignore the ITP assistance altogether after just one ITP session, not accepting a single token the ITP system suggested afterwards, instead consistently typing his own translations, even when they matched.The translation activity data produced by TrB was deemed invalid and discarded, as any measures collected would not be representative of working in ITP, but rather of unassisted translation. 15TrB's negative perception may have been partly due to speed reasons, as reported in his feedback; nevertheless, it seems a fairly harsh judgment after having tried ITP only once.It may be that some translators are not willing to engage with PE or ITP, possibly because they already have a working routine they are comfortable with.In this sense, the views expressed by Vasconcellos and León (1985), O'Brien (2002), Rico Pérez and Torrejón (2012) and De Sutter (2011) that PE requires that the translator has a positive attitude towards MT resonate for ITP. 16n attempting to relate translators' feedback to their own background and quantitative results, our data can give us some useful insights into their potential relationship.While Moorkens and O'Brien (2015) reported that PE was perceived by some translators as a tedious activity, and that experienced translators were more likely to have negative views of PE than novice translators, in our study, we did not find any indication that experienced translators had negative views towards ITP, provided they also had PE experience.In fact, the more experienced translator (TrA), both in terms of length of experience (> 10 years) and translation volume in the previous 12 months (> 55k words)-who had between two and five years' PE experience-expressed, as detailed above, consistently positive views of ITP.In terms of translation productivity indicators, as shown in Table 2, TrA logged the fastest PE time and the second fastest ITP time of all translators.TrA also produced the highest quality texts in the PE condition and the highest quality texts of all translators in the ITP condition.
Regardless of their translation experience, professional translators with little or no PE experience though, may be more reluctant to engage in ITP.The two translators who expressed negative views of ITP-TrG to a minor degree, and, much more markedly, TrB-had, respectively, less than two years' PE experience or no PE experience.Finally, there is some indication that translators who have formal PE training or provide PE services frequently benefited the most from ITP.In fact, of the four translators who were faster in ITP than in PE, two have PE industry certifications (TrC [TAUS]; TrF [SDL]) and one (TrE) provides frequent PE services.

Conclusions
As is usual with research studies on translation processes, the empirical work discussed here presents several limitations which may have had an influence on the obtained results.Specifically, while casmacat is suitable for research on translation processes due to its extensive logging capabilities and it being open-source and web-based, it does not present features that are common in commercial CAT tools, such as multiple search and replace or spell checkers, to which professional translators are accustomed.This limited functionality may have contributed to slowing down translators and, most likely, to the presence of spelling issues.Additionally, cost and convenience motivated our sample size, quality assessment and language pair choices, therefore restricting the application of our findings.
With the above limitations in mind, the ITP study presented here and Daems and Macken (2019) in this special issue are, to the best of our knowledge, the first empirical studies investigating human translators' productivity in a NITP setting.Overall, our results point at ITP being a viable alternative to PE considering translators' feedback and temporal, technical and translation quality indicators.Translation requires,

Table 1
Summary statistics for translation productivity indicators in ITP and PE

Table 2
Translators' main translation productivity indicators and impressions

Table 1 ,
Processing Time is measured in seconds per source token; Manual Insertions, Manual Deletions, Navigation and Special Key Presses, and Mouse Clicks are measured as counts per source token.Tokens of MT origin as measured as counts per 100 source tokens; MQM Score as a percentage, and translation issues, as count per 1000 source tokens.Word Prediction Accuracy (WPA) is a percentage.

Table 3
Summary parameters, standard errors and significance of models (translator productivity) Poisson family and log link function was applied to Mouse Clicks, Fluency, and Adequacy Issues; binomial family and logit link to Pass Status (a dichotomization of the MQM score); and Poisson family and square root link to all remaining variables.Observation-level random effects were added to Poisson models to model overdispersion ***p < .001,*p < .05,all two-tailed