Integrating Artificial Intelligence-Assisted Translation Support into English Courses: Effects on Translation Accuracy, Perceived Stress, and Anxiety

Summary

This study examines how structured artificial intelligence-assisted translation support can be incorporated into English courses to improve translation accuracy while reducing perceived stress and anxiety among higher education students.

Abstract

English translation competence is an important academic skill in higher education, yet many students experience substantial linguistic difficulty, cognitive burden, and psychological pressure during translation tasks. This mixed-methods study examined whether structured artificial intelligence-assisted translation support integrated into English courses was associated with changes in translation accuracy, perceived stress, and anxiety among university students. A total of 525 undergraduate and postgraduate students participated in an eight-week instructional intervention. Baseline and post-intervention assessments included a translation accuracy test, the Perceived Stress Scale, and the Generalized Anxiety Disorder-7 scale. Quantitative analysis was conducted using paired comparisons, correlation analysis, regression analysis, and group-based comparisons, and was complemented by qualitative interviews to explore students’ learning experiences. After the intervention, participants showed improved translation accuracy and lower levels of perceived stress and anxiety. Higher frequency of tool use was associated with better translation performance, although the study design does not support strong causal inference. Qualitative findings further indicated that structured tool-assisted practice helped reduce cognitive overload, supported revision processes, and improved engagement with translation tasks. These findings suggest that integrating artificial intelligence-assisted translation support into English courses may offer a practical instructional approach for strengthening translation performance while supporting student well-being. The study contributes empirical evidence to current discussions of technology-enhanced translation pedagogy and provides an adaptable teaching framework for higher education contexts.

Introduction

English translation competence has become increasingly important in higher education as universities seek to prepare students for global academic exchange, international professional communication, and cross-cultural knowledge production¹. Yet translation is not a purely linguistic exercise. It also requires contextual judgment, cultural mediation, and sustained cognitive control, making it a demanding form of academic work for many learners². In classroom settings, these demands often generate considerable pressure, especially for students in disciplines such as linguistics, literature, and international studies, where translation tasks are closely tied to performance evaluation and academic identity³. Conventional translation instruction, although pedagogically valuable, can therefore become time-intensive, cognitively burdensome, and psychologically taxing when students are expected to manage complex source texts with limited immediate support⁴.

Recent developments in artificial intelligence-assisted language technologies have created new possibilities for addressing these instructional challenges⁵. In translation learning, such tools can provide rapid lexical, syntactic, and semantic support, thereby helping students reduce uncertainty during drafting and revision⁶. Some studies in English as a foreign language contexts have suggested that technology-assisted learning may improve language performance, strengthen learner confidence, and support emotional regulation during demanding tasks^7,8. At the same time, existing literature has also raised concerns about over-reliance on automated output, reduced critical engagement, and the continuing need for teacher guidance when these tools are introduced into formal learning environments⁹. As a result, the educational value of artificial intelligence-assisted translation support cannot be understood solely in terms of output efficiency. It must also be examined in relation to how such support is structured, how students interact with it, and how it shapes both performance and psychological experience during learning.

Despite growing interest in artificial intelligence-assisted translation, an important gap remains in the current literature. Much of the existing research has focused either on translation quality or on general attitudes toward educational technology, while relatively few studies have examined academic and psychological outcomes together within a clearly structured classroom intervention¹⁰. In particular, limited evidence is available on whether guided use of artificial intelligence-assisted translation support in university English courses is associated not only with improved translation accuracy, but also with reduced perceived stress and anxiety across a sustained instructional period. This gap is important because the pedagogical value of such tools depends not only on whether students produce better translations, but also on whether the learning process becomes more manageable, confident, and educationally sustainable.

Accordingly, the present study investigated the integration of structured artificial intelligence-assisted translation support into university-level English courses over an eight-week period. The study had two objectives: first, to examine whether students’ translation accuracy, perceived stress, and anxiety changed after the intervention; and second, to explore whether patterns of tool use were associated with translation performance and student experience. We further expected that structured implementation, rather than unguided tool exposure, would be associated with improved translation outcomes and more favorable psychological responses. The novelty of the study lies in its combined evaluation of cognitive and psychological dimensions within a single instructional design, using both quantitative assessment and qualitative inquiry to examine how artificial intelligence-assisted translation supports functions in real classroom practice.

Protocol

This study was conducted within the School of Humanities and Arts, Civil Aviation Flight University of China, using de-identified educational data collected in the course setting. No directly identifiable personal information was included in the analytic dataset. All participants were informed of the study purpose and procedures, and written informed consent was obtained prior to participation. Participation was voluntary, and non-participation or withdrawal from the research component did not affect course attendance, grades, or access to regular instruction. Under the applicable research practice framework of Civil Aviation Flight University of China, studies based on de-identified minimal-risk educational data may not require formal ethical review. All de-identified materials were stored in password-protected files accessible only to the research team. The ethics and participant flow are shown in Figure 1.

1. Participant recruitment and baseline profiling

A total of 525 students were recruited from university English-related or translation-related courses. Undergraduate and postgraduate students from disciplines such as linguistics, literature, and international studies were included. All volunteers were screened before enrollment. Only students who were currently taking an English or translation course and had at least intermediate English proficiency based on institutional placement records, prior coursework, or equivalent academic evidence were included. Students with substantial professional translation experience were excluded. Age, gender, academic level, field of study, prior experience with AI-assisted translation support, and baseline frequency of tool use were recorded. Participant characteristics were presented in Table 1.

2. Instructional design and intervention workflow

An eight-week classroom intervention was implemented in which conventional translation instruction remained the primary teaching mode and AI-assisted translation support was used as a structured supplementary scaffold¹¹. In Week 0, the baseline assessment was completed. From Week 1 to Week 8, one translation task was assigned each week, and all participants were required to follow the same instructional sequence. At the end of Week 8, the assessment battery used at baseline was repeated. After the quantitative phase, semi-structured interviews were conducted with a purposively selected subgroup. The overall workflow is shown in Figure 2.

3. Baseline assessment

All baseline measures were administered in a supervised classroom setting under identical instructions and timing conditions. A translation task based on a source text of about 300 words was used. A passage containing both literal and culturally nuanced segments was selected so that students were required to demonstrate lexical accuracy, syntactic control, semantic fidelity, and contextual understanding¹². The use of external tools was prohibited during baseline testing. All completed scripts were anonymized before scoring. Each script was scored independently by two trained raters using a rubric covering lexical accuracy, grammatical and syntactic appropriateness, semantic completeness, and cultural contextualization. If the score discrepancy exceeded the predefined tolerance range, discussion was required until a consensus score was reached.

After the translation task, the 10-item Perceived Stress Scale was administered and scored according to the standard procedure, with higher scores indicating greater perceived stress¹³. The 7-item Generalized Anxiety Disorder scale was then administered and scored using the published method, with higher scores indicating greater anxiety severity¹⁴. Finally, a short translation-confidence measure was administered using a five-point Likert format. Representative items, score ranges, and variable definitions were provided in Table 2.

4. Standardized weekly integration of AI-assisted translation support

During each intervention week, one translation task was assigned that was matched as closely as possible in genre and difficulty to the other weekly tasks. Students were required to complete the task in the same order each week. First, the source text was read independently, and difficult lexical items, syntactic structures, or culturally loaded expressions were marked. Second, an initial draft or a brief note describing anticipated translation problems was prepared. Third, consultation of approved AI-assisted translation systems was allowed only during drafting and revision. Direct submission of unedited machine-generated output was not allowed.

Students were required to evaluate all retained or rejected machine suggestions using the same four criteria each week: lexical fit, syntactic naturalness, discourse coherence, and cultural appropriateness. Submission of the marked source text, the machine-assisted draft, and the final revised translation was required for every weekly task. After submission, a structured peer discussion was organized using the same prompts each week so that students explained where automated output was helpful, misleading, overly literal, or contextually inappropriate. This design followed current evidence showing that machine translation is most educationally useful when embedded in guided comparison and revision rather than used as a replacement for human judgment¹⁵.

5. Monitoring and categorizing tool use

AI-assisted tool use was tracked throughout the eight-week intervention. After each weekly assignment, participants were required to complete a brief online usage log reporting whether automated support was used, what function it served, how frequently it was consulted, and at what stage it entered the translation process. Standardized response categories were provided in Week 1 so that participants used the same reporting criteria throughout the study. At the end of Week 8, the weekly records were aggregated, and overall usage frequency was classified as daily, weekly, or occasional use. The operational definitions of these categories are shown in Figure 3.

6. Post-intervention assessment

At the end of Week 8, the same assessment battery used at baseline was repeated. A post-intervention translation text was used that was matched as closely as possible to the baseline text in length, genre, and difficulty. Administration conditions, timing, scoring rubric, and rater workflow were kept consistent with baseline procedures. The same Perceived Stress Scale, the same Generalized Anxiety Disorder scale, and the same translation-confidence measure were administered in the same order used at baseline.

7. Qualitative interview procedures and thematic analysis

After post-intervention assessment, 40 students were purposively selected for semi-structured interviews, ensuring variation in gender, academic level, and prior experience with AI-assisted translation support. Interviews were conducted face-to-face or through an approved online meeting format using the same interview guide for all participants. Questions were asked about tool-use strategies, perceived cognitive burden, confidence changes, evaluation of machine output, and concerns about over-reliance. All interviews were audio-recorded with participant consent and transcribed verbatim.

The transcripts were analyzed using thematic analysis. Two researchers were asked to read all transcripts independently, generate initial codes, compare coding outputs, merge overlapping codes, and refine candidate themes through repeated review of the transcripts. After the initial coding stage, inter-rater agreement was calculated, and Cohen’s kappa was reported. The analytic logic followed established procedures for thematic analysis in qualitative research¹⁶.

8. Quantitative and mixed-methods analysis

Descriptive statistics were used to summarize participant characteristics and baseline scores. Paired-sample comparisons were used to test pre-post differences in translation accuracy, perceived stress, anxiety, and confidence. Multiple regression models were used to identify predictors of post-intervention translation accuracy, with baseline performance, frequency of AI-assisted tool use, and psychological variables entered as independent variables. Chi-square tests were used to examine associations between usage categories and categorical outcomes, and Pearson correlation coefficients were used to assess the direction and strength of associations among translation accuracy, usage frequency, perceived stress, anxiety, and confidence. Statistical significance was set at alpha = 0.05 using two-tailed testing. The quantitative and qualitative findings were integrated at the interpretation stage according to standard mixed-methods procedures¹⁷.

Results

Primary outcomes

The primary objective of this study was to determine whether translation accuracy, perceived stress, anxiety, and translation confidence changed over the eight-week intervention period. Analyses based on 525 participants showed significant pre-post differences across all four primary outcome variables. As shown in Table 3, translation accuracy increased from 68.45 (SD = 8.12) at baseline to 82.76 (SD = 6.95) after the intervention. Perceived s...

Discussion

The present study examined whether structured use of translation support tools in university English courses was associated with changes in translation performance and psychological outcomes over an eight-week period. The results showed significant pre-post differences in all four primary outcomes. Translation accuracy increased after the intervention, while perceived stress and anxiety decreased, and translation confidence increased. The secondary analyses further showed that post-intervention translation accuracy was a...

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
AI-assisted translation and language support platform	OpenAI (web)	Web service; accessed on study schedule	Used during weekly drafting and revision tasks to generate translation suggestions and language feedback
Computer workstations / laptops	Institutionally available equipment	Equivalent acceptable	Used for translation tasks, data entry, coding, and statistical analysis
Digital audio recorder / recording device	Institutionally available equipment	Equivalent acceptable	Used to record interviews when required
Generalized Anxiety Disorder scale (7-item)	Original developers: Spitzer et al.	Standard instrument	Used at baseline and post-intervention to assess anxiety
Headphones for transcription and checking	Institutionally available equipment	Equivalent acceptable	Used during transcription and coder checking
Online translation platform	Google LLC (web)	Web service; accessed on study schedule	Used as an additional translation support source during the intervention
Perceived Stress Scale (10-item)	Original developers: Cohen et al.	Standard instrument	Used at baseline and post-intervention to assess perceived stress
Qualitative coding software	QSR International	Version 14	Used for thematic organization and coding support during interview analysis
Statistical analysis software	IBM Corp.	Version 31	Used for descriptive statistics, paired-sample tests, regression, chi-square tests, and Pearson correlations
Translation accuracy source texts	Study-developed	Version 1.0	Approximately 300-word passages matched in length and difficulty
Translation confidence measure	Study-developed	Version 1.0	Three-item self-report measure using a 5-point Likert format
Translation scoring rubric	Study-developed	Version 1.0	Four analytic domains: lexical accuracy, grammatical/syntactic appropriateness, semantic fidelity, cultural contextualization
Weekly usage log	Study-developed	Template v1.0	Used to record task stage, frequency, and function of support-tool use