الفهرس 
IntroductionOnly 14 pages are availabe for public view
 |  | from | 186 |  |  |
| | | from | 186 | | |
Abstract
The increasing volume of translation work has led to the integration of machine translation, specifically neural machine translation (NMT), into the translation industry. Consequently, a common practice has emerged wherein machine-generated translations are utilized as a preliminary version, which is then manually refined by human translators. This practice is known as machine translation post-editing. However, it is imperative to ascertain whether the effort invested in post-editing is worthwhile and whether it poses greater demands than traditional translation methods. Previous studies have examined the effort required for both traditional translation and post-editing; however, their focus is primarily on statistical machine translation and the resulting findings are not consistent. Additionally, these studies have failed to address the cognitive effort specifically associated with translating Arabic texts. Understanding the cognitive effort involved in post-editing Arabic neural machine-translated texts and bridging this gap is crucial for minimizing mental fatigue and enhancing productivity among translators. Thus, this study aims to bridge this gap by investigating the cognitive effort expended in both NMT post-editing and traditional translation within the English-Arabic language pair, including the relationship between cognitive effort and quality and productivity in both methods. This investigation employs a combination of two approaches for cognitive effort: the process and the product, allowing for a comprehensive analysis of the cognitive effort involved.
The study aims at addressing the following key questions:
Which is more cognitively demanding, traditional translation or post-editing of Neural machine translation?
How does traditional translation compare to post-editing in terms of productivity using the word-per-minute metric as a measure?
How does traditional translation compare to post-editing in terms of quality using DQF-MQM error typology as a measure?
What is the level of cognitive effort required to review post-edited and traditionally translated Arabic texts using Temnikova’s (2010) error typology as a measure?
What is the correlation between cognitive effort in process and product on the one hand, and the correlation between cognitive effort and quality and productivity on the other hand?
The study poses the following hypotheses to the research questions:
Cognitive effort in post-editing is expected to be less than cognitive effort in Traditional translation.
The cognitive effort exerted during the review of post-edited texts is expected to be higher compared to reviewing the traditionally translated texts.
The productivity of post-editing is expected to be higher than that of traditional translation.
The quality of post-editing is expected to be lower than that of traditional translation.
There is a correlation between cognitive efforts in the translation process and the resulting translation product and a correlation between cognitive effort and quality and productivity.
To address the research question and validate the hypotheses, the study utilizes two documentary texts extracted from a UNDP report, each consisting of approximately 500 words (see Appendix B). The neural machine translation engine provided by Google is chosen as the machine translation tool. The study involves a group of 20 professional translators who were tasked with translating one of the two texts and post-editing the other. The translations and post-edits are conducted on the Matecat platform.
To investigate the first research question, cognitive effort is examined during the translation and post-editing processes using a metric called pause-to-word ratio (PWR). PWR is calculated by dividing the number of post-editing or translation pauses by the number of words post-edited or translated. The pause data was collected using the Inputlog software. The analysis of recorded pauses during the process reveals that participants exhibit a higher number of pauses and longer pause durations during translation compared to post-editing. The Pause-to-Word Ratio (PWR) further supports these findings, indicating higher cognitive effort in translation compared to post-editing.
Furthermore, to explore the cognitive effort through the product, word translation entropy (HTra) is employed. HTra refers to the level of uncertainty or variability in the available translations for a particular word. Higher entropy values indicate a greater number of translation options, signifying a higher cognitive effort exerted to produce the target text. Conversely, lower entropy values suggest a limited number of consistent translations, indicating lower cognitive effort. The analysis of word translation entropy in the final translated products has shed light on the cognitive effort involved. The results reveal that traditional translation often demands more cognitive effort due to the presence of multiple translation choices, whereas post-editing involves fewer options and, consequently, lower cognitive effort.
Concerning the second question, the study investigates the productivity of participants in two different translation methods. A questionnaire and a practical experiment using Inputlog are employed to gather data. The questionnaire responses indicate that participants perceive a significant increase in productivity during post-editing tasks compared to translating from scratch. The practical experiment validates these findings, showing that the average time spent on translation tasks is significantly higher than on post-editing tasks. The productivity rate for translation is 5.5 words per minute (WPM) (330 words per hour), while for post-editing, it is 11.5 WPM (690 words per hour). These results confirm that post-editing yields higher productivity rates compared to traditional translation.
As for the third question, the quality assessment of translation and post-editing tasks is carried out using the TAUS DQF-MQM error typology. The findings indicate that translation has a slightly higher number of errors compared to post-editing, but the difference is not statistically significant. When examining specific error types, accuracy errors are more common in translation, while post-editing shows slightly higher occurrences of style and fluency errors, often stemming from the machine translation system and not adequately addressed by participants. The results suggest that post-editing can help translators produce high-quality translations that are comparable to those created from scratch. However, traditional translation performs better in terms of style and fluency, indicating that post-editing may be more suitable for texts that require a more literal style, such as technical documents.
To address the fourth research question, the study employed Temnikova’s (2010) cognitive model to analyze the effort required for reviewing translated and post-edited texts based on the type and number of corrections. It gives a cognitive effort ranking for each correction. The findings of the study demonstrate that the cognitive effort ranking involved in correcting errors is higher in translated texts compared to post-edited texts, which suggests that post-edited texts require less cognitive effort to be reviewed than translated texts.
Regarding the fifth research question, the study identified correlations between cognitive efforts in both the translation process and the resulting translation product. Additionally, a correlation is found between productivity and cognitive effort. Specifically, a strong negative correlation is observed, indicating that higher levels of cognitive effort are associated with decreased productivity. This suggests that when participants engage in post-editing, they benefit from the machine translation output, resulting in fewer modifications (indicating low technical effort), increased efficiency (improved productivity), and reduced cognitive effort.
Lastly, the study provides recommendations for future research on cognitive effort, highlighting the importance of investigating various factors that can impact cognitive effort in both translation and post-editing. Examining additional variables such as text complexity, translator expertise, and language pair-specific challenges could provide valuable insights into the specific conditions that influence cognitive effort in translation and post-editing. These areas for future research will ultimately contribute to advancing the field and promoting a more efficient and effective translation process.