Translation Revision and Post-editing

Industry Practices and Cognitive Processes

Maarit Koponen, Brian Mossop, Isabelle S. Robert, Giovanna Scocchera, Maarit Koponen, Brian Mossop, Isabelle S. Robert, Giovanna Scocchera

• Publisher: Routledge
• Year: 2020
• eBook ISBN: 9781000201574
• Topic: Languages & Linguistics
• Subtopic: Linguistics
• Index: Languages & Linguistics

Translation Revision and Post-editing looks at the apparently dissolving boundary between correcting translations generated by human brains and those generated by machines. It presents new research on post-editing and revision in government and corporate translation departments, translation agencies, the literary publishing sector and the volunteer sector, as well as on training in both types of translation checking work.

This collection includes empirical studies based on surveys, interviews and keystroke logging, as well as more theoretical contributions questioning such traditional distinctions as translating versus editing. The chapters discuss revision and post-editing involving eight languages: Afrikaans, Catalan, Dutch, English, Finnish, French, German and Spanish. Among the topics covered are translator/reviser relations and revising/post-editing by non-professionals.

The book is key reading for researchers, instructors and advanced students in Translation Studies as well as for professional translators with a special interest in checking translations.

PART I
Post-editing versus revision

1
PREFERENTIAL CHANGES IN REVISION AND POST-EDITING

Jean Nitzke and Anne-Kathrin Gros

In this chapter, we will investigate and discuss the phenomenon of preferential changes in revision and post-editing, which we call over-editing. Translators over-edit when they revise/post-edit more than is necessary, given certain guidelines. In both tasks, some translators feel the urge to improve all linguistic aspects because they want to achieve perfect quality, even though the guidelines state otherwise. It is very important that revisers and post-editors adhere to guidelines in order to make the process efficient and worthwhile, especially in regard to payment.

Over-editing is a phenomenon that is previously known from research on CAT tools, in particular from work with translation memory (TM) systems. Translators revise the suggested TM match too extensively, when the quality of the match would have been good enough with no edits (for exact matches) or just a few edits (for fuzzy matches), and this is costly in terms of time and consequently money. Mellinger and Shreve (2016) compare the behaviour of nine participants in a study who were asked to each translate seven segments with exact matches, seven segments with fuzzy matches, and seven segments with no matches. The match type was randomly chosen for each segment, and the participants did not know what kind of match they would be confronted with. However, they were instructed not to change a segment if they felt that it did not need to be changed. They were also allowed to delete the entire match if they felt that it would take too much time to edit it. The participants changed 60% of those exact matches which did not require any changes. On the other hand, they changed only 74% of the fuzzy matches, although all of them needed changes. Many of the unnecessary changes were preferential changes on a syntactic or lexical level. The explanation for this phenomenon offered by Mellinger and Shreve is that because the participants do not have to create their own translations when there are fuzzy and exact matches but instead have to compare source and target text, they actively look for mistakes. Further, the participants have their own translations in mind, which results in two competing versions of the target text segment. The participants tend to change the TM match into their own translation.

The same phenomenon can also be observed in PE and revision. Translators might create their own mental concept of the source unit with their own translation ideas and are then confronted with the machine translation or a translation created by another person (see Oster 2017 for information on priming and monitoring). These latter translations are not necessarily defective, but they do not always correspond to the translator’s representation. Over-editing was previously examined for PE in De Almeida (2013). In her study, she analysed the edits of 20 participants post-editing IT texts (ten translated into French and ten into Brazilian Portuguese). Her main categories were essential changes, preferential changes, unimplemented essential changes and introduction of errors. ‘Preferential changes’ are synonymous with what we refer to as over-editing. The participants made 45.16 preferential changes on average in a text with 1008 words; in other words, an unnecessary change was made every 22.32 words.

Specific instructions to avoid over-editing are frequently included in the PE brief. For example, one of the instructions in the study by Aikawa et al. (2012: 7) very explicitly states: “Avoid over-editing: don’t try to over-edit if the existing translation(s) … are grammatical and readable.” As over-editing is a rather abstract concept, the usefulness of such an instruction might be questionable. Participants are reminded that they should keep the editing process to a minimum, but the extent to which the translators adhere to the instructions remains unknown. Further, it might be difficult for translators who are not trained in PE to lower their quality standards and correct only what they are asked to correct. Similarly, it might be difficult to revise only what is necessary and to disregard personal style and habits.

This chapter will focus on the over-editing behaviour of translators in PE and revision scenarios. To that end, datasets from three studies will be analysed, in which the participants were required to perform one of the two tasks. In the following sections, we will first describe the three datasets and present the instructions which were given to the participants. Then we will outline how we identified over-editing instances and report the results for the individual tasks with their similarities and differences, which will be discussed in the last section.

1. The studies

The analysis in this chapter is based on the datasets of three different studies, all of which were conducted with English source texts and German as the target language. There were no strict time restrictions for any of the tasks. They all were conducted in Translog II (Carl 2012), a program used to record keystrokes, mouse activities and gaze data with the help of an eyetracker (either tobii TX300 or SMI mobile). The final product as well as the eyetracking and keylogging data were connected via the alignment tool YAWAT (Germann 2008). All the studies included one or more questionnaires to gather information on the participants and to check how satisfied the participants were with their own work. All sessions¹ of each participant were recorded on the same day, usually one after the other without a break.

Study 1

The first subset of data comes from an experiment in which the participants were asked to perform three different tasks: translate from scratch, full post-edit (FPE) and light post-edit (LPE) texts from either the technical or the medical fields (only the PE sessions will be analysed here). Table 1.1 presents the instructions for the tasks, which were adapted from the TAUS PE guidelines (Massardo et al. 2016). The three technical texts were excerpts from a dishwasher manual, while the three medical texts were taken from documentation included with a vaccine against measles, insulin for the treatment of diabetes patients and a medication for the treatment of cancer. All texts were approximately 150 words long. The MT output was created by Google Translate (at the time, it was still a statistical rather than a neural MT engine).

TABLE 1.1 PE instructions for Study 1 in line with TAUS PE guidelines

Instructions for LPE	Instructions for FPE
The output of ‘good enough’ post-editing (or ‘light’ post-editing) is comprehensible and accurate, but not stylistically compelling. The text may sound like it was generated by a computer, syntax might be somewhat unusual, grammar may not be perfect, but the message is accurate.	The level of quality is generally defined as being comprehensible, accurate, stylistically fine, though the style may not be as good as achieved by a nativespeaker human translator. Syntax is normal, grammar and punctuation are correct.
Use as much of the raw MT output as possible! Aim for semantically correct translations. Ensure that no information has been accidentally added or omitted. No need to use correct terminology, as long as it is clear what is meant. Edit any inappropriate or culturally unacceptable content. Apply rules regarding spelling. Don’t implement corrections that are of a stylistic nature only. Don’t restructure sentences solely to improve the natural flow of the text.	Aim for grammatically, syntactically and semantically correct translation. Ensure that no information has been accidentally added or omitted. Ensure that key terminology is correctly translated and used consistently. Edit any inappropriate or culturally unacceptable content. Use as much of the raw MT output as possible. Apply rules regarding spelling, punctuation and hyphenation. Ensure that formatting is correct.

The participants consisted of 12 advanced translation students for the technical texts and 9 for the medical texts, all at the University of Mainz, Faculty of Translation Studies, Linguistics and Cultural Studies, Germersheim. They had undergone at least two years of translation training and had passed at least one exam on translating in the relevant field. They were German native speakers who had studied English as a first or second foreign language. Some of them had some PE experience, but only a minority. Each participant had to translate one text from scratch, light post-edit another text and full post-edit a third text. The technical texts were therefore each translated four times, light post-edited four times and full post-edited four times, the medical texts three times (Table 1.2). For further information on the study, see Čulo and Nitzke (2016).

TABLE 1.2 Distribution of number of participants by task and text

	Translation from scratch	Light PE	Full PE
Technical text 1	4	4	4
Technical text 2	4	4	4
Technical text 3	4	4	4
Medical text 1	3	3	3
Medical text 2	3	3	3
Medical text 3	3	3	3

Study 2

This study was also conducted at the University of Mainz, Faculty of Translation Studies, Linguistics and Cultural Studies in Germersheim in 2012, on behalf of the Center for Research and Innovation in Translation and Translation Technology (CRITT), Copenhagen Business School, Denmark. The data² are included in the CRITT-TPR database (https://sites.google.com/site/centretranslationinnovation/tpr-db), which collects translation process data for different tasks and in different languages. The source texts consisted of four newspaper articles and two sociology texts with different levels of complexity. The length of the texts varied between 100 and 150 words. The MT output was again created by Google Translate.

In total, 24 participants took part in the study, 12 of them professional translators (with university degrees and some professional work experience) and 12 translation students (students of the university with only a little professional work expe...

Table of contents

• Cover
• Half Title
• Title Page
• Copyright Page
• Contents
• List of contributors
• Introduction
• Part I Post-editing versus revision
• Part II Non-professional revision and post-editing
• Part III Professional revision in various contexts
• Part IV Training
• Bibliography
• Index