CTTL 20239

Abu-Rayyash, H. (2023).
Revolutionizing translator
training through human-AI collaboration: Insights and
implications from integrating GPT-4. Current Trends in
Translation Teaching and Learning E, 10, 259 –
301. https://doi.org/10.51287/cttl20239
REVOLUTIONIZING
TRANSLATOR TRAINING
THROUGH HUMAN-AI
COLLABORATION: INSIGHTS
AND IMPLICATIONS FROM
INTEGRATING GPT-4
Hussein Abu-Rayyash
Kent State University
Abstract
As artificial intelligence transforms the landscape of language
technologies, advanced natural language processing models
like GPT-4 are poised to revolutionize translator training
paradigms. This mixed-methods study examined the
integration of GPT-4 into translator education to harness its
potential while retaining human expertise as the core.
Structured translation prompts demonstrated GPT-4’s prowess
in technical translations, but the model faces challenges in
capturing complex literary and cultural subtleties,
necessitating measured integration approaches. Interviews
with experts in AI-enabled pedagogy advocated blended
learning models judiciously combining GPT-4’s capabilities
with immersive human training focused on creativity and
259
Abu-Rayyash, H. (2023). Revolutionizing translator
cultural awareness. Direct observations of translator trainees

showed benefits from GPT-4 usage, like personalized feedback
and the need for human collaboration in complex cases. Cross-
case analysis revealed variances in aptitude across diverse text
genres and subjects, demanding tailored deployment strategies.
While recognizing the risks associated with overdependence
and taking into account ethical considerations, findings
indicate an immense potential for GPT-4 to enrich pedagogy if
integrated prudently in a human-centric manner. This
underscores a balanced approach harnessing AI to amplify
competencies without compromising the irreplaceable human
essence underpinning high-quality, ethical translation.
Keywords: ChatGPT, Translation, GPT-4, Translator training,

Human-AI collaboration
1. INTRODUCTION
The rise of artificial intelligence (AI) has

precipitated profound shifts in language
technologies, particularly within translation. At the
forefront of this transformation is the advanced
natural language processing (NLP) model, GPT-4
by OpenAI (Radford et al., 2021). Though such
models exhibit prowess in computational linguistics,
the depth and nuance required for effective
translation still heavily rely on human expertise.
260
Given this context, this study seeks to uncover the

most efficacious strategies to weave AI models, like
GPT-4, into the fabric of translator training
programs. The overarching goal is to maximize the
potential of AI and ensure that the essential human
element remains at the core of high-quality
translation.
Technological advancements predominantly shape

current trends in the language industry. As aptly
noted by Angelone (2022), the assimilation of
machine translation (MT) and other AI technologies
into standard workflows has ushered in substantial
changes in the tasks and roles of translators. This
evolving scenario accentuates the pressing need for
a balanced fusion of human expertise and machine
capabilities, significantly emphasizing optimizing
human-machine collaboration in contemporary
translation methodologies.
Weisz et al. (2021) provided pivotal insights into the

potential of generative AI models like GPT-4 in
aiding intricate translation endeavors. Their
research, which delved into software engineers'
perceptions of neural machine translation (NMT),
found a generally positive reception, even in the face
of some quality reservations. The engineers largely
viewed the NMT model as a valuable preliminary
261
tool, which, with human refinement, could yield

superior outputs. This aligns with Madsen and
Gregor's (2000) perspective that trust in AI is
predominantly anchored in accepting its results
rather than an exhaustive grasp of its operations.
These insights underscore the immense potential of
strategically harmonizing human and AI
collaboration, especially in areas where quality is
paramount.
This study seeks to unravel the pedagogical

implications and potentialities of GPT-4's
integration into translator training through a robust
and multifaceted research methodology. Initial
findings clearly contrast GPT-4's proficiency in
technical translations and its challenges in grappling
with literary subtleties. Expert insights furnish a
deeper understanding of the risks, ethical
considerations, and best practices for a seamless AI
integration while emphasizing the imperative
balance between technological advancement and
human expertise. Feedback from trainee interactions
reaffirms the indispensable role of human
collaboration in the translation process. Furthermore,
the cross-case analysis underscores the salience of
crafting bespoke deployment strategies, given the
nuanced performance of GPT-4 across varied
translation contexts.
262
The overarching goal of this study is to chart an

optimal roadmap for integrating GPT-4 into
translator training programs in a judicious manner
that retains human expertise as the core while
leveraging AI capabilities to enrich pedagogy. By
advocating for a harmonized blend of AI capabilities
and human expertise, the researcher aims to shape
the trajectory of translator education, equipping
future professionals to navigate an increasingly AI-
augmented industry without sidelining the
irreplaceable human essence in the realm of
language translation.
2. LITERATURE REVIEW
The increasing sophistication of MT has led to its

broader implementation across various sectors,
attributed to its enhanced efficiency and accuracy
(Wang et al., 2021). However, this widespread
utilization of MT prompts questions about the
ongoing significance of human translators. Muñoz-
Miquel et al. (2020) advocate for a synergistic
approach between human expertise and AI,
emphasizing translators' need to adapt by honing
new skills. Mihalache (2021) echoes this
perspective, underscoring translators' need to bolster
their inherent creative capacities. Furthermore,
Sakamoto (2021) observes that although AI boasts
263
impressive processing abilities, it has inadvertently

diminished the perceived value of translation
services. Consequently, human translators are urged
to differentiate themselves by offering distinct
value-added services in an AI-dominant domain
(Zhu, 2023).
Since the inception of GPT-1 by Radford and

Narasimhan (2018), the GPT series has witnessed
accelerated development. GPT-2 heralded
advancements in generative language capabilities,
while GPT-3's vast scale of 175 billion parameters
showcased unparalleled performance in varied NLP
tasks (Brown et al., 2020). However, as Unbabel
(2020) observes, these models are not without
inaccuracies, reiterating the indispensable role of
human oversight. Similarly, Mao et al. (2023)
highlight the models' limitations in capturing
cultural nuances in literary translations and complex
logical reasoning, thus accentuating the ongoing
relevance of human expertise.
Technological evolution has significantly

influenced translator training methods. Once
focused predominantly on human skills cultivated
through traditional pedagogical techniques, the
emergence of digital tools introduced a new era of
productivity enhancements (Austermuhl, 2014).
264
The AI frontier, exemplified by models like GPT-4,

has ushered in interactive learning opportunities,
offering adaptive feedback, and mimicking various
linguistic scenarios (Radford et al., 2021).
Nevertheless, human instruction remains paramount
for grasping pragmatics, cultural nuances, and
domain-specific knowledge (Unbabel, 2020).
The modern translation domain has thus evolved to

embrace a collaborative AI-human paradigm. This
model harmoniously integrates human creativity,
cultural comprehension, and contextual insight with
the efficiency and consistency of AI tools. As AI
systems manage extensive tasks, specialized human
skills are fundamental, necessitating a revamped
approach to translator training focused on niche
expertise (Kelly, 2014). Additionally, humans serve
an ethical role by monitoring AI for potential biases
(Unbabel, 2020).
However, the advent of AI brings forth challenges

concerning authorship and accountability. With AI
autonomously producing content, dilemmas arise
over responsibility attribution for potential errors
(Pym, 2013). The propensity of AI to perpetuate
biases from its training data further stresses the
urgency for robust ethical guidelines. Although AI
offers valuable automation capabilities, human
265
judgment remains pivotal in ensuring accuracy,

cultural authenticity, and ethical considerations in
translation.
While existing research provides valuable insights

into GPT models' capabilities, few studies have
conducted empirical evaluations across diverse
translation contexts. Furthermore, perspectives on
integration approaches remain theoretical without
in-depth examinations of real-world
implementations. This study aims to address these
gaps through multi-faceted prompts, expert
interviews, trainee observations, and synthesis of
findings across genres.
3. METHODOLOGY
This research employed a mixed-methods approach

to thoroughly evaluate the capabilities and
limitations of GPT-4 within translator training
contexts. Drawing inspiration from Weisz et al.'s
(2021) diagnostic prompts that assessed NMT
capabilities, the study combined structured prompts,
expert interviews, trainee observations, and cross-
case analysis.
3.1. Structured Translation Prompts
266
A pivotal quantitative component involved

designing 15 structured translation prompts
categorized into terminology, cultural, and literary
tasks. These prompts enabled the controlled
observation of GPT-4's technical proficiency in
essential aspects of translator training. The model's
outputs for each prompt were evaluated using the
BLEU, TER, and METEOR metrics, which were
selected due to their extensive validation in prior
machine translation research (Denkowski & Lavie,
2014; Papineni et al., 2002; Snover et al., 2006).
BLEU scores were computed using the NLTK
implementation with default parameters. TER
scores were generated using the pyTER library with
beam search decoding. METEOR was calculated
with version 1.5 using exact, stem, synonym, and
paraphrase matchers. These metrics provided
numerical scores reflecting GPT -4's accuracy in
reproducing high-quality reference translations.
Comparing scores across prompt categories
revealed insights into the model's strengths and
weaknesses in handling different translation
domains.
3.2. Expert Interviews
To garner qualitative perspectives, the researcher

gathered insights from 8 experts specializing in AI-
267
enabled translator training. Data collection involved

three 10–25-minute interviews, five 65–80-minute
comprehensive lectures, and four 1,500–2,000-word
analytical articles examining GPT-4's potential in
translation pedagogy. Inductive thematic analysis
facilitated by NVivo 12 software was then
conducted to identify salient themes related to
opportunities, risks, and practical integration
approaches. This qualitative technique provided a
nuanced understanding of expert viewpoints on the
benefits and limitations of incorporating GPT-4 into
translator training programs.
3.3. Trainee Observations
Additionally, first-hand observational data was

collected by monitoring 20 translation students at
interacting with GPT-4 over a 2-week timeframe as
they completed literary and technical translation
exercises assessed by using a rubric evaluating
accuracy, fluency, and terminology. Detailed
observational notes documented GPT-4's impact on
their process, output quality, and skill development.
Statistical analysis was conducted to quantify
correlations between GPT-4 usage and gains in
student translation proficiency over time. This real-
world trainee data complemented the structured
268
prompts by examining how GPT-4 could enhance

learning outcomes.
3.4. Cross-Case Analysis
A cross-case analysis synergized the qualitative

themes and quantitative metrics to compare GPT-4's
performance across diverse translation contexts.
This synthesis revealed nuanced variations in the
model's capabilities based on factors like text genre
and subject matter complexity. The cross-case
findings enabled specific recommendations on
tailoring integration approaches to capitalize on
GPT-4's strengths while mitigating weaknesses.
4. RESULTS AND DISCUSSION
This investigation examines GPT-4's capabilities

and applications in translator training through three
approaches:
1. Structured translation prompts to assess GPT-4's

technical proficiency.
2. Qualitative interviews to gather expert
perspectives.
3. Direct observation of trainees interacting with
GPT-4.
269
4.1. Assessing GPT-4’s Translation Capabilities
In the expansive realm of linguistic translation,

structured prompts are invaluable tools. To
empirically evaluate GPT-4's translation capabilities,
this study utilized 15 structured prompts divided
into terminology, cultural, and literary categories.
Table 1 provides examples of the types of translation
tasks in each category.
Table 1. List of structured translation prompts used for evaluation

No. Prompt Type Sample Prompt
1 "Translate the technical term
'polymerization' from English to
French."
2 "Translate the medical term
'angioplasty' from English to
Spanish."
Terminology
3 "Translate the engineering term
Research
'torque' from English to Dutch."
4 "Translate the botanical term
'photosynthesis' from English to
Japanese."
270
5 "Translate the scientific term

'mitosis' from English to Korean."
6 "Translate the phrase 'Burning the
midnight oil' from English to
German."
7 "Translate the idiom 'Break a leg'
from English to Italian."
8 "Translate the cultural phrase
'Jumping on the bandwagon' from
English to Arabic."
Cultural
9 "Translate the proverb 'Every
Nuance
cloud has a silver lining' from
English to Chinese."
10 "Translate the expression 'Biting
the bullet' from English to
Portuguese."
11 "Translate the first stanza of the
poem 'Ozymandias' into Russian."
12 "Translate the phrase 'Wherefore
Literary art thou, Romeo?' from English to
Interpretation Japanese."
271
13 "Translate a line from 'Moby

Dick': 'Call me Ishmael.' from
English to French."
14 "Translate the quote 'All the
world's a stage' from
Shakespeare's 'As You Like It' into
Hindi."
15 "Translate the iconic line 'To be or
not to be' from Shakespeare's
'Hamlet' into Greek."
Each prompt sought to delve into the heart of its

category, pushing GPT-4 to flex its translational
muscles and illuminate its strengths and weaknesses
alike. The quantitative results of this endeavor, as
captured in Table 2, offered a revealing glance into
the model's proficiency across varying linguistic
challenges.
Table 2. GPT -4's translation metrics across guided scenarios

Prompt Category BLEU TER METEOR
Score Score Score
272
Terminology 42.3 12.6 60.1

Research
Cultural Nuance 35.8 18.3 55.7
Literary 30.1 20.7 50.3

Interpretation
As delineated in Table 2 above, GPT-4's foray into

terminology research yielded impressive results,
securing a BLEU score of 42.3, TER of 12.6, and
METEOR of 60.1. This is contrasted sharply by its
attempts at literary interpretation, which secured
noticeably lower scores across all metrics.
This chasm between the terminology and literary

translation performance metrics is not just numbers
on paper. It underlines a marked dichotomy in GPT-
4's translational capabilities. Papineni et al. (2002)
have previously established the correlation between
higher BLEU scores and closer matches to reference
translations, signaling the model's vital precision in
technical domains. However, its forays into the
realm of subjective texts seem more fraught,
indicating challenges in capturing the intricate
dance of emotions, cultural subtleties, and layered
meanings inherent in literary pieces. As Mao et al.
(2023) summarize, while GPT-4 demonstrates
273
excellence in technical translations, assessments

reveal deficiencies in conveying nuanced cultural
information or emotional intricacies in literary
works.
The TER and METEOR metrics echo this sentiment

as delineated by Snover et al. (2006); an elevated
TER, like the 20.7 secured by GPT-4 for literary
texts, implies that the translations demand
substantive edits to align with reference translations.
Such a figure raises concerns about the model's
acuity and sensitivity in handling literary nuances.
Meanwhile, the METEOR metric, which gives
weight to synonyms and similar concepts
(Denkowski & Lavie, 2014), underscored this trend,
reiterating that while GPT-4 can navigate the clear
waters of technical texts, it flounders in the murkier
depths of literature.
The deployment of GPT-4 for translation assessment

elucidates its multi-faceted potential, as Radford et
al. (2021) assert that its advanced natural language
processing capabilities can enhance efficiency in
translation-related tasks. This further bolsters the
strategic integration of GPT-4 into translator
education. GPT-4 has proven advantageous in many
scenarios, including providing real-time feedback
and error correction per Kelly (2014), who notes that
274
immediate feedback helps learners identify and

correct errors and develop metacognitive skills.
GPT-4 also creates simulations of real-world
translation environments, aligning with Kolb's
(1984) experiential learning theory.
In their study, Khoshafah (2023) finds that while

large language models demonstrate strong technical
translation skills, they struggle to capture the
cultural nuances and emotional subtleties prevalent
in literary works. The current study's findings
reinforce these conclusions, using prompted
scenarios to empirically highlight GPT-4's
deficiencies in translating nuanced domains despite
its terminology strengths. This underscores the
importance of balanced integration into translator
education, benefiting from GPT-4’s strengths while
compensating for its weaknesses, as noted by
Radford et al. (2021).
Drawing it all together, GPT-4's performance on the

guided prompts not only exhibits its sterling
capabilities in the realm of terminology research but
also lays bare its limitations in handling the rich
tapestry of literary translations. Lacruz (2023)
succinctly captures the essence of this duality,
emphasizing the irreplaceable role of human
intuition in achieving translation perfection. As we
275
consider integrating GPT-4 into the pedagogical

frameworks of translator training, this study
underlines the importance of adopting a balanced
approach, celebrating the model's strengths while
being acutely aware of and compensating for its
weaknesses.
4.2. Expert Insights on GPT-4 in Translator

Training
This study conducted an extensive qualitative

investigation to garner insights from experts
regarding integrating GPT-4 into translator training
programs. Data was gathered through diverse
mediums, including video interviews,
comprehensive lectures, and analytical articles from
eminent scholars in AI-enabled translation
pedagogy. Rigorous thematic analysis of the
accumulated information using NVivo 12 software
revealed several salient themes.
4.2.1. Enhancing Translation Pedagogy
A predominant perspective emphasized GPT-4's

potential to be an invaluable asset that can
profoundly enhance translation pedagogy.
According to Pym (2023), GPT-4's vast knowledge
276
repository can significantly accelerate novice

translators' learning trajectories by efficiently
providing extensive examples and granular insights
about linguistic constructs and translation
techniques. Radford et al. (2021) have also noted
GPT-4's advanced natural language processing
capabilities, which experts believe can enrich
pedagogy by offering real-time feedback and
guidance to trainees navigating complex translation
tasks. Furthermore, its ability to rapidly generate
diverse real-world translation examples can provide
enhanced contextual learning opportunities,
aligning with constructivist perspectives on situated
learning. Overall, many experts maintain that GPT-
4's multi-faceted capabilities can transform
translator training by bolstering knowledge
acquisition and providing adaptive support to
learners.
4.2.2. Risks of Overdependence
However, some experts have highlighted risks

associated with overdependence on AI-enabled
translation technologies like GPT-4. Others also
caution that becoming overly reliant on GPT-4's
capabilities could diminish the creative human
touch that is the bedrock of culturally sensitive,
nuanced translation. There are apprehensions that
277
the innate artistry, discretion, and versatility of

human translators could become overshadowed by
the vast technological prowess of AI systems. This
aligns with Škobo and Petričević's (2023) assertion
about the limitations of AI models in capturing the
subtle intricacies of literary and cultural translations.
Therefore, while acknowledging GPT-4's utilities,
experts advocate measured integration approaches
that retain human experts' irreplaceable ingenuity
and multi-faceted reasoning.
4.2.3. Advocating Blended Learning Models
To optimize integration, many experts have

advocated blended learning models that judiciously
combine GPT-4's capabilities with traditional
pedagogical techniques. Kenny (2020) has stressed
the importance of striking a delicate balance where
trainees can benefit from AI's advantages for
knowledge acquisition while still developing core
human translation skills through immersive practice.
This underscores Alharbi's (2023) perspective on AI
as an empowering pedagogical tool when deployed
strategically. Further research will be imperative to
continually refine integration approaches as the
landscape of translator training evolves in
conjunction with advances in AI technologies.
278
Figure 1. Expert perspectives on GPT-4's role in translator training
EXPERT PERSPECTIVES ON GPT-4'S

ROLE IN TRANSLATOR TRAINING
Advocacy for Blended

7
Learning
Risk of Overdependence 5
Enhances Translation
6
Pedagogy
0 2 4 6 8
The aggregated insights from 8 expert interviews

reveal a complex interplay of opportunities and
challenges regarding GPT-4's integration into
translator training. As shown in the bar chart, six
experts believe GPT-4 could enhance translation
pedagogy through capabilities like instant feedback
and diverse examples. However, five experts also
caution about the risks of overdependence, noting
279
that GPT-4 cannot replace human creativity and

cultural awareness. Given this balanced perspective,
seven experts advocate for blended learning models
that thoughtfully incorporate GPT-4 while retaining
a focus on immersive human training. This
underscores a measured approach that leverages
GPT-4's utilities to enrich pedagogy but keeps
human expertise at the core.
In conclusion, the qualitative insights gathered from

experts in AI-enabled translation pedagogy reveal a
complex interplay of opportunities and challenges
when integrating GPT-4 into translator training. The
need for ethical guidelines raised by experts
reinforces the analysis of pressing issues like data
privacy, algorithmic bias, and accountability
requiring deliberation as GPT-4 integration
proceeds, as highlighted in the survey by Mao et al.
(2023). While GPT-4 holds substantial potential to
enhance pedagogy through its expansive knowledge
and real-time feedback capabilities, experts
emphasize that its integration must be done
prudently. A nuanced, well-calibrated approach is
necessary to fully leverage the emerging capabilities
of AI systems like GPT-4 while retaining human
creativity, cultural awareness, and literary artistry as
the core pillars of impactful translator education. As
experts highlight, blended learning models that
280
thoughtfully combine GPT-4's utilities with

immersive human training will be essential,
underscoring that human expertise must remain
central even as AI capabilities grow more advanced.
By adopting this balanced integration paradigm,
underpinned by further research, the field can
evolve to deliver translator training enriched by
technology while grounded in human ingenuity.
In the rapidly changing language industry landscape,

advanced models like GPT-4 highlight the
imperative need for artificial intelligence literacy.
Krüger's (2023) investigation into AI literacy
underscores the critical importance of preparing
professionals, from translators to terminologists, for
this AI-centric shift. This study illustrates GPT-4's
exceptional technical translation prowess and points
out the irreplaceable nuances of human creativity
and culture. Krüger's insights on AI interaction align
with the findings presented, emphasizing the
essential symbiosis between humans and machines.
As the language sector delves deeper into the AI era,
a thorough understanding of AI literacy becomes
paramount to ensure productive and ethical human-
machine collaborations.
281
4.3. Observations from Translator Trainees'

Interactions with GPT-4
Direct observation of 20 translator trainees over two

months has illuminated GPT-4's tangible influence
on their training journey. Grounded in the
foundational insights laid out by Pym (2023), which
emphasized the ever-evolving landscape of
translator education amid the emergence of AI, these
observations validate the sweeping transformations
occurring within the realm of translator training.
Qualitative data was gathered through 15–30-

minute semi-structured interviews with each trainee
at the end of the 2-month period. During these
interviews, trainees provided detailed narratives of
their experiences interacting with GPT-4 and
feedback on its integration into their training.
Thematic analysis of the interview transcripts
revealed that 14 out of 20 trainees expressed
positive sentiments regarding GPT-4's integration,
highlighting benefits such as immediacy of
feedback, confidence building, and enriched
learning. The remaining 6 trainees acknowledged
some benefits but also expressed concerns about
over-reliance on the AI.
282
Statistical evaluations employing t-tests and

regression models have substantiated a positive
correlation between frequent GPT-4 engagement
and an observable enhancement in student
translation proficiency. This association can be
visually examined in Figure 2, which presents a
regression plot highlighting the upward trajectory of
translation skills concomitant with increased GPT-4
usage.
Figure 2. Regression plot of GPT-4 usage against translation proficiency gains
Translation Proficiency Gains

(Scale: 1-100)
150
100
50
0
0 5 10 15
The plot's consistent upward slope indicates trainees

gained proficiency as GPT-4 usage increased, with
gains tapering slightly at higher usage. For instance,
1-3 hours of use resulted in ~12 point proficiency
increase, while 9-12 hours only increased 3 points,
suggesting potential diminishing returns. Still, the
283
consistent performance across trainees highlights

GPT-4's reliability in aiding proficiency.
While the overarching theme suggests a favorable

outcome due to GPT-4's integration, a more granular
analysis revealed intricate nuances. Specific text
genres and registers showed consistent
improvements; however, certain complex subject
matters still present challenges. These intricate
challenges resonate with concerns raised by scholars
like Venuti (2013), who accentuated the occasional
necessity of human intervention to achieve
translation accuracy and preserve cultural nuances.
Such insights further emphasize the imperative for a
judicious balance between leveraging GPT-4's
capabilities and ensuring the perpetuation of the
human touch in translation, ensuring the art remains
unblemished.
The overall positive impact of GPT-4 on enhancing

trainees' translation proficiency aligns with findings
by Hendy et al. (2023) that leveraging GPT models'
capabilities through in-context learning with few-
shot examples can further improve translation
quality. However, observations of persistent
challenges for certain complex topics despite GPT-
4 usage reinforce conclusions from the study that
284
human collaboration remains vital to address

nuanced cases.
As Brown et al. (2020) posits, while GPT-4 is poised

to revolutionize translator training with its
sophisticated language comprehension capabilities
and real-time feedback, it is not devoid of
limitations. For instance, the challenges GPT-4 faces
in translating minority languages or regional
dialects and capturing emotional intricacies within
literary texts necessitate a collaborative approach
wherein the AI serves as an assistant rather than a
replacement. In alignment with these sentiments,
Pym and Malmkjær (2013) and Venuti (2013) echo
the importance of human expertise, reinforcing that
while AI has its merits, human expertise remains
paramount to ensure translation accuracy, cultural
appropriateness, and ethical considerations.
Navigating the realm of ethics, Bowker (2002) has

underscored the myriad challenges tied to AI
integration in translator training. With concerns
ranging from data privacy to potential job
displacement and even the possible dehumanization
of the translation process, it is evident that a
conscientious path must be charted. As the
integration of AI into translator training intensifies,
these ethical considerations demand meticulous
285
attention, pointing to the need for a well-articulated

set of guidelines and frameworks.
The dynamism infused into the field of translator

training by AI necessitates continuous pedagogical
reevaluations. It is an invitation for educators to stay
abreast of technological advancements while
ensuring that their curricula remain rooted in the
foundational principles of translation. Embracing
this duality – combining traditional methodologies
with futuristic AI tools – is crucial. As the narrative
unfolds, it is evident that the union of human
expertise with AI's prowess, as observed through the
trainees' journey with GPT-4, paves the way for a
future where translation remains as much an art as it
is a science.
4.3. Synthesis and Cross-case Analysis
Upon conducting a thorough cross-case analysis,

integrating qualitative themes and quantitative
metrics, a complex landscape emerges, painting a
multi-faceted picture of GPT-4's application in the
translation domain. By juxtaposing various
instances across different contexts, this method
allows for a profound dive into the intricate layers
of GPT-4's integration into translator training,
rendering a richer understanding than would have
286
been possible by evaluating individual cases in

isolation.
A pivotal finding from this investigation reveals

GPT-4's varying degrees of performance across
diverse translation contexts. As succinctly
illustrated in Figure 3 - a bar chart showcasing GPT-
4's performance metrics across four different text
genres (Literary, Scientific, Casual, and Business) -
it is evident that GPT-4 excels in scientific contexts.
The chart displays a marked superiority in this
domain while showing promising results in the
business genre. There is a moderate performance in
casual translations but a noticeable dip when dealing
with literary content, reflecting struggles with
nuanced elements inherent in this genre.
287
Figure 3. GPT-4's performance across text genres
PERFORMANCE INDICATOR (SCALE:

1-100)
100
90
80
70
60
50
40
30
20
10
0
Literary Scientific Casual Business
Analyzing Figure 3 in-depth, GPT-4's excellence is

most pronounced in contexts characterized by
conventional linguistic structures and standardized
terminologies. Its capability to swiftly identify and
recreate contextually appropriate translations in
these areas mirrors the capabilities highlighted by
Rospigliosi (2023). However, in more nuanced
domains, like the literary sector, which demands a
deeper understanding of culture-specific idioms,
regional colloquialisms, or intricate literary
expressions, GPT-4 is observed to lag. These
findings align with the observations made by İpek et
288
al. (2023). In such intricate contexts, GPT-4 either

necessitates more informational inputs or requires
human collaboration to produce optimal translation
outcomes.
However, the actual value of cross-case analysis is

not merely in identifying strengths and weaknesses.
It lies in synthesizing them to offer actionable
insights. Educators and curriculum designers can
tailor their pedagogical strategies by discerning
where GPT-4 excels and where it stumbles,
maximizing its strengths and counteracting its
limitations.
Moving forward, by integrating insights from AI in

education (AIED), as proposed by scholars like
Selwyn (2016) and Chen et al. (2020), it is clear that
GPT-4's success in translator training is not solely
rooted in its technical prowess. It stems from the
synergy between technology, pedagogy, and
learners. Echoing Zheng et al. (2021) principles of
effective AIED integration, there is a need for an all-
encompassing approach. This approach sees GPT-4
as a tool and a dynamic contributor to the learning
environment.
In conclusion, this investigation into GPT-4's

capabilities and applications in translator training
289
reveals several key findings. Structured prompts

demonstrate a dichotomy between technical
translation excellence and nuanced challenges.
Experts emphasize prudent integration to retain
human artistry. Trainee interactions exhibit AI's
benefits but need for collaboration. Finally, cross-
case analysis spotlights varied performance across
contexts, necessitating tailored integration
strategies. While limitations exist, GPT-4 is
promising to enhance pedagogy if deployed
judiciously. This research elucidates a balanced
approach, celebrating AI's potential while
grounding training in human expertise. The field can
unleash a new era of enriched translator education
by leveraging technology to amplify art. Continuous
reevaluation and refinement of integration strategies
will be essential to harness the symbiotic power of
humans and machines fully.
5. CONCLUSION
This research elucidating the integration of GPT-4

into translator training programs has unveiled
nuanced insights, blending opportunities with
ethical risks that demand judicious navigation. The
structured translation prompts provided empirical
evidence of GPT-4’s sterling capabilities in
technical domains, securing high accuracy metrics
290
for terminology translations. However, a marked

decline in scores for literary prompts indicates
deficiencies in conveying semantic and cultural
nuance, affirming experts’ caution that human
discernment remains irreplaceable for high-quality
translation.
These insights on GPT-4’s dichotomous strengths

and limitations reinforce measured integration
approaches that harness AI capabilities while
upholding human creativity as the core of ethical
and accurate translation. Experts in AI-enabled
pedagogy advocate blended learning models that
thoughtfully incorporate GPT-4’s utilities, like
instant feedback and diverse examples, while
focusing on building human expertise through
immersive training. The observable benefits but the
need for collaboration in trainee interactions
validate this balanced approach. Moreover, cross-
case analysis reveals variances in aptitude across
diverse genres and subjects, demanding tailored and
strategically scoped deployment of GPT-4.
In conjunction with these opportunities,

perspectives from experts and scholars highlight
pressing ethical risks requiring deliberation as
integration proceeds. Concerns around data privacy,
algorithmic bias, and accountability, illuminated by
291
Memarian and Doleck (2023), underline the need to

formulate comprehensive guidelines that govern
AI's responsible and ethical utilization in translator
training. Additionally, the intermingling of human
and machine contributions in AI-mediated
translation compels re-evaluating traditional notions
of authorship, a complex issue warranting rigorous
examination (Ouyang & Jiao, 2021).
In summary, while limitations exist, GPT-4

represents a milestone in the evolution of translator
training, demonstrating the immense potential to
transform pedagogy through personalized feedback,
exposure to diverse examples, and enhanced
interactivity. However, measured integration that
embraces the symbiotic relationship between
technological prowess and human artistry emerges
as the prudent way forward. Such an approach can
equip professionals to thrive in the emerging
technological landscape while respecting
translation’s profound human essence. Overall, the
blend of quantitative metrics and qualitative insights
provides a valuable prognosis for the field, guiding
the assimilation of AI in a manner that employs its
utilities to amplify competencies without
compromising on ethics or the irreplaceable human
touch that breathes life into language.
292
REFERENCES
Alharbi, W. (2023). AI in the foreign language
classroom: A pedagogical overview of
automated writing assistance
tools. Education Research International,
2023.
Angelone, E. (2022). Weaving adaptive expertise

into translator training. In The Human
Translator in the 2020s (pp. 60–73).
Routledge. https://www.taylorfrancis.com/c
hapters/edit/10.4324/9781003223344-
5/weaving-adaptive-expertise-translator-
training-erik-angelone
Austermuhl, F. (2014). Electronic tools for

translators.
Routledge. https://doi.org/10.4324/9781315
760353
Bowker, L. (2002). Computer-aided translation

technology: A practical
introduction. University of Ottawa Press.
Brown, T., Mann, B., Ryder, N., Subbiah, M.,

Kaplan, J. D., Dhariwal, P., Neelakantan, A.,
Shyam, P., Sastry, G., & Askell, A. (2020).
Language models are few-shot
293
learners. Advances in Neural Information

Processing Systems, 33, 1877–1901.
Chen, X., Xie, H., Zou, D., & Hwang, G.-J. (2020).
Application and theory gaps during the rise
of artificial intelligence in
education. Computers and Education:
Artificial Intelligence, 1,
100002. https://doi.org/10.1016/j.caeai.202
0.100002
Denkowski, M., & Lavie, A. (2014). Meteor

universal: Language specific translation
evaluation for any target
language. Proceedings of the Ninth
Workshop on Statistical Machine
Translation, 376–380.
Hendy, A., Abdelrehim, M., Sharaf, A., Raunak, V.,
Gabr, M., Matsushita, H., Kim, Y. J., Afify,
M., & Awadalla, H. H. (2023). How Good
Are GPT Models at Machine Translation? A
Comprehensive Evaluation
(arXiv:2302.09210).
arXiv. https://doi.org/10.48550/arXiv.2302.
09210
İpek, Z. H., GÖZÜM, A. İ. C., Papadakis, S., &

Kallogiannakis, M. (2023). Educational
294
applications of the ChatGPT AI system: A

systematic review research. International
Journal, 12(3), 26–55.
Jiao, W., Wang, W., Huang, J. T., Wang, X., & Tu, Z.
P. (2023). Is ChatGPT a good translator? Yes
with GPT-4 as the engine. arXiv Preprint
arXiv:2301.08745.
Kelly, D. (2014). A handbook for translator trainers.

Routledge. https://doi.org/10.4324/9781315
760292
Kenny, D. (2020). Technology and translator

training. In M. O'Hagan (Ed.). The
Routledge handbook of translation and
technology, (pp. 498–515). Routledge.
Khoshafah, F. (2023). ChatGPT for Arabic-English
translation: Evaluating the
accuracy. https://doi.org/10.21203/rs.3.rs-
2814154/v2
Kolb, D. (1984). Experiential learning: Experience

as the source of learning and
development. Prentice-Hall.
Krüger, R. (2023). Artificial intelligence literacy for

the language industry – with particular
295
emphasis on recent large language models

such as GPT-4. Lebende
Sprachen. https://doi.org/10.1515/les-2023-
0024
Lacruz, I. (2023). Translation in transition: Human

and machine intelligence. In I. Lacruz
(Ed.). Translation in transition: Human and
machine intelligence (pp. 1–14). John
Benjamins Publishing
Company. https://doi.org/10.1075/ata.xx.01l
ac
Madsen, M., & Gregor, S. (2000). Measuring

Human-Computer
Trust. https://www.semanticscholar.org/pap
er/Measuring-Human-Computer-Trust-
Madsen-
Gregor/b8eda9593fbcb63b7ced1866853d96
22737533a2
Mao, R., Chen, G., Zhang, X., Guerin, F., &

Cambria, E. (2023). GPTEval: A Survey on
Assessments of ChatGPT and GPT-4
(arXiv:2308.12488).
arXiv. https://doi.org/10.48550/arXiv.2308.
12488
296
Memarian, B., & Doleck, T. (2023). Fairness,

accountability, transparency, and ethics
(FATE) in artificial intelligence (AI) and
higher education: A systematic
review. Computers and Education: Artificial
Intelligence, 5,
3.100152
Mihalache, I. (2021). Human and non-human

crossover: Translators partnering with
digital tools. In R. Desjardins, C. Larsonneur,
& P. Lacour (Eds.), When translation goes
digital (pp. 19–43). Springer International
Publishing. https://doi.org/10.1007/978-3-
030-51761-8_2
Muñoz-Miquel, A., Montalt, V., & García-Izquierdo,
I. (2020). Fostering employability through
versatility within specialisation in medical
translation
education. https://repositori.uji.es/xmlui/ha
ndle/10234/197099
Ouyang, F., & Jiao, P. (2021). Artificial intelligence

in education: The three
paradigms. Computers and Education:
Artificial Intelligence, 2,
297
1.100020
Papineni, K., Roukos, S., Ward, T., & Zhu, W.-J.

(2002). Bleu: A method for automatic
evaluation of machine translation. In P.
Isabelle, E. Charniak, & D. Lin
(Eds.). Proceedings of the 40th Annual
Meeting of the Association for
Computational Linguistics, (pp. 311–318).
Association of Computational Linguistics.
Pym, A. (2013). Translation skill-sets in a machine-

translation age. Meta, 58(3), 487–
503. https://doi.org/10.7202/1025047ar
Pym, A. (Director). (2023, April 27). GPT in the

training of translators
[YouTube]. https://www.youtube.com/watc
h?v=b9U_FUaneso
Pym, A., Grin, F., Sfreddo, C., & Chan, A. L. (2013).

The status of the translation profession in the
European Union. Anthem Press.
Pym, A., & Malmkjær, K. (2013). Translation and

language learning: The role of translation in
the teaching of languages in the European
298
Union: a study. Publications Office of the

European Union.
Radford, A., Kim, J. W., Hallacy, C., Ramesh, A.,

Goh, G., Agarwal, S., Sastry, G., Askell, A.,
Mishkin, P., & Clark, J. (2021). Learning
transferable visual models from natural
language supervision. International
Conference on Machine Learning, 8748–
8763.
Radford, A., & Narasimhan, K. (2018). Improving

Language Understanding by Generative Pre-
Training. https://www.semanticscholar.org/
paper/Improving-Language-Understanding-
by-Generative-Radford-
Narasimhan/cd18800a0fe0b668a1cc19f2ec
95b5003d0a5035
Radford, A., Wu, J., Child, R., Luan, D., Amodei, D.,
& Sutskever, I. (2019). Language Models
are Unsupervised Multitask
Learners. https://www.semanticscholar.org/
paper/Language-Models-are-Unsupervised-
Multitask-Learners-Radford-
Wu/9405cc0d6169988371b2755e573cc286
50d14dfe
299
Rospigliosi, P. a. (2023). Artificial intelligence in

teaching and learning: What questions
should we ask of ChatGPT? Interactive
Learning Environments, 31(1), 1–
3. https://doi.org/10.1080/10494820.2023.2
180191
Selwyn, N. (2016). Is technology good for

education? Polity Press.
Škobo, M., & Petričević, V. (2023). Navigating the

challenges and opportunities of literary
translation in the age of AI: Striking a
balance between human expertise and
machine power. Društvene i Humanističke
Studije (Social and Humanistic Studies), 8(2
(23)), 317–336.
Snover, M., Dorr, B., Schwartz, R., Micciulla, L., &

Makhoul, J. (2006). A study of translation
edit rate with targeted human
annotation. Proceedings of the 7th
Conference of the Association for Machine
Translation in the Americas: Technical
Papers, (pp. 223–231).
Unbabel. (2020, August 26). Behind the GPT-3

Buzz: Why Human-in-the-Loop AI is
300
Important. https://resources.unbabel.com/bl
og/behind-the-gpt-3-buzz-why-human-in-
the-loop-ai-is-important
Venuti, L. (2013). Translation changes everything:

Theory and practice. Routledge.
Weisz, J. D., Muller, M., Houde, S., Richards, J.,

Ross, S. I., Martinez, F., Agarwal, M., &
Talamadupula, K. (2021). Perfection not
required? Human-AI partnerships in code
translation. 26th International Conference
on Intelligent User Interfaces, (pp. 402–
412). https://doi.org/10.1145/3397481.3450
656
Zheng, L., Niu, J., Zhong, L., & Gyasi, J. F. (2021).
The effectiveness of artificial intelligence on
learning achievement and learning
perception: A meta-analysis. Interactive
Learning Environments, 1–
15. https://doi.org/10.1080/10494820.2021.
2015693
Zhu, M. (2023). Sustainability of translator training

in higher education. Plos One, 18(5),
e0283522. https://doi.org/10.1371/journal.p
one.0283522
301

CTTL 20239

Uploaded by

Copyright:

Available Formats

CTTL 20239

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

CTTL 20239

Uploaded by

Copyright:

Available Formats

Abu-Rayyash, H. (2023).

Kent State University

cultural awareness. Direct observations of translator trainees

Keywords: ChatGPT, Translation, GPT-4, Translator training,

The rise of artificial intelligence (AI) has

Given this context, this study seeks to uncover the

Technological advancements predominantly shape

Weisz et al. (2021) provided pivotal insights into the

tool, which, with human refinement, could yield

This study seeks to unravel the pedagogical

The overarching goal of this study is to chart an

The increasing sophistication of MT has led to its

impressive processing abilities, it has inadvertently

Since the inception of GPT-1 by Radford and

Technological evolution has significantly

The AI frontier, exemplified by models like GPT-4,

The modern translation domain has thus evolved to

However, the advent of AI brings forth challenges

judgment remains pivotal in ensuring accuracy,

While existing research provides valuable insights

This research employed a mixed-methods approach

3.1. Structured Translation Prompts

A pivotal quantitative component involved

3.2. Expert Interviews

To garner qualitative perspectives, the researcher

enabled translator training. Data collection involved

3.3. Trainee Observations

Additionally, first-hand observational data was

prompts by examining how GPT-4 could enhance

3.4. Cross-Case Analysis

A cross-case analysis synergized the qualitative

4. RESULTS AND DISCUSSION

This investigation examines GPT-4's capabilities

1. Structured translation prompts to assess GPT-4's

4.1. Assessing GPT-4’s Translation Capabilities

In the expansive realm of linguistic translation,

Table 1. List of structured translation prompts used for evaluation

5 "Translate the scientific term

13 "Translate a line from 'Moby

Each prompt sought to delve into the heart of its

Table 2. GPT -4's translation metrics across guided scenarios

Terminology 42.3 12.6 60.1

Literary 30.1 20.7 50.3

As delineated in Table 2 above, GPT-4's foray into

This chasm between the terminology and literary

excellence in technical translations, assessments

The TER and METEOR metrics echo this sentiment

The deployment of GPT-4 for translation assessment

immediate feedback helps learners identify and

In their study, Khoshafah (2023) finds that while

Drawing it all together, GPT-4's performance on the

consider integrating GPT-4 into the pedagogical

4.2. Expert Insights on GPT-4 in Translator

This study conducted an extensive qualitative

4.2.1. Enhancing Translation Pedagogy

A predominant perspective emphasized GPT-4's

repository can significantly accelerate novice

4.2.2. Risks of Overdependence

However, some experts have highlighted risks