The term artificial intelligence (AI) was coined in 1956 by John McCarthy, who used the term artificial intelligence for the first time ([19]).

The study [of artificial intelligence] is to proceed on the basis of the conjecture that every aspect of learning or any other feature of intelligence can, in principle, be so precisely described that a machine can be made to simulate it. An attempt will be made to find how to make machines use language, form abstractions and concepts, solve kinds of problems now reserved for humans, and improve themselves.

The application of AI in higher education can be traced back to the 1960s when early computer-assisted instruction systems were developed and employed in universities ([5]). In recent years, AI’s potential in education has been increasingly recognized, and it has been adopted in various educational practices, such as in education ([3]), higher education ([4]), online education ([16]), etc.

The research concerning AI-based tools for teaching and learning in higher education has seen sustained exploration over the years ([4]; [11]; [29]).

The release of the ChatGPT LLM by OpenAI in 2022 particularly extended interest in AI to a broad civic audience and particularly to higher education institutions. [10] find that these comprehensive language models can serve as a supplement rather than a replacement for classroom instruction.

As a phenomenally popular AI, DeepSeek has rapidly captured widespread attention through its unique technological innovations and promotion strategies, becoming a cornerstone of digital infrastructure for enterprises. People are increasingly finding themselves enveloped by its influence: their educational tools, daily necessities, professional dependencies, and social interactions all intersect with DeepSeek ([14]).

[2] approach educational AI tools from three different perspectives: a) learner-facing, b) teacher-facing, and c) system-facing AI in education (AIEd). Among these, teacher-facing systems are used to support the teacher and reduce his/her workload by automating tasks of administration, assessment, feedback, etc. Of the 138 research articles [4] examined, only 17% of them focused on instructors, which is far from sufficient.

Assessment and evaluation were the most common uses of AIEd in higher education. [13] reported that ChatGPT can generate acceptable multiple-choice items based on the given reading materials. [15] used natural language processing to create a system that automatically created highly realistic short-answer questions. [20] used OpenAI GPT to generate reading comprehension items, and the generated items produced a similar level of difficulty and yielded strong discrimination power. [11] reviewed publications published between 2017 and July 2023 and highlighted several research gaps, including the need for more empirical studies to assess the effectiveness and impact of GenAI tools. The studies above demonstrate that AI technologies can be employed to generate various test items, but few studies have investigated AIG for translation tasks.

2.3.1. Lower Technical Barriers for Course-Based AIG

AIG was originally developed for large-scale assessments, and the creation and manipulating of item models and/or templates are needed, which is beyond the technical capabilities of ordinary teachers, especially language teachers who mostly major in liberal arts. The introduction of GenAI models in AIG can help teachers to overcome these technical barriers.

2.3.2. Course-Based Teachers Are Qualified Reviewers of Test Items

The generated test items by AIG need reviews by subject matter experts (SMEs), and the designers for course-based examinations are always the teachers of the courses, so they are qualified SMEs for the review of the generated test items of their courses.

2.3.3. Reducing Teachers’ Burden of Test Item Writing

The syllabus, knowledge, and skills of a course are relatively stable, and the course teachers have to prepare test items semester after semester. By introducing GenAI models, course teachers can generate many test items by providing proper prompts, which reduces their burden of repetitive tasks.

RQ1: Whether the English-Chinese and Chinese-English sentence translation items generated by GenAI models reflect the differences in language, culture, and thinking patterns between English and Chinese, and how about their content validity?

RQ2: Whether the English-Chinese and Chinese-English text translation passages generated by GenAI models reflect the differences in language, culture, and thinking patterns between English and Chinese, and what about their content validity and readability?

Listening, speaking, reading, writing, and translating/interpreting are basic skills of English as a foreign language (EFL) learners, among which translation is of vital significance for information input and output. Sentence- and text-translation (English-Chinese; Chinese-English) are the main test types. Thus, we put prompts into the seven GenAI models to check whether they can finish the specific generative tasks and how well their performances are.

Task 1: Generating English-Chinese Sentence Translation Items

Prompt 1: Based on a 10,000-word vocabulary base, generate 10 English-Chinese sentence translation items, paying attention to the differences in language, culture, and thinking patterns between English and Chinese. Reference translation and an explanation are required.

Task 2: Generating Chinese-English Sentence Translation Items

Prompt 2: It is the same as Prompt 1 except for the translation direction of Chinese-English.

Task 3: Generating English-Chinese Text Translation Passages

Prompt 3: Based on a 10,000-word vocabulary base, generate two English-Chinese text translation passages. The length is about 150 words, paying attention to the differences in language, culture, and thinking patterns between English and Chinese. Reference translation and an explanation are required.

Task 4: Generating Chinese-English Text Translation Passages

Prompt 4: It is the same as Prompt 3 except for the translation direction of Chinese-English.

Three kinds of GenAI models are used in this research, including one localized model Deepseek-r1:1.5b, four online models, DeepSeek R1, DeepSeek V3.1, Ernie Bot and QwQ-Plus; and three advanced models, including Qwen-VL, ChatPDF, and QWQ-Plus.

Depending on the purpose of testing, tests can generally be categorized into aptitude test, achievement test, diagnostic test, proficiency test, and exit test. Among these, achievement test is always used to assess the students’ success in learning a foreign language, and it is usually directly related to a specific foreign language course. Therefore, it has been suggested that achievement test should be based on the specific course syllabus and teaching materials ([21]).

According to testing theory, a test has content validity if its content constitutes a representative sample of the language skills, structures, etc., with which it is meant to be concerned ([8]). To ensure content validity, the skills or constructs to be tested are typically outlined in detail for test developers’ reference ([21]).

Based on the course syllabus of translation and some authoritative textbooks in China ([30]; [6]; [17]), the English-Chinese translation strategies mainly include the selection, extension, and commendatory or derogatory meaning of words, conversion of parts of speech, amplification, repetition, omission, affirmation and negation; division and combination of sentences; and the Chinese-English translation strategies include equivalent translation, amplification, omission, combination translation, conversion of parts of speech, transformation of expression, commendatory or derogatory translation, translation of Chinese idioms, proverbs, and two-part allegorical sayings, subject prominence and topic prominence, passive and active voice, cohesion and coherence, etc. The operational definition of content validity for translation tasks refers to how well the task items cover all relevant parts of the construct of translation competence.

In order to review the content validity of the generated sentence translation items and text translation passages, two subject matter experts are invited. They hold Ph.D. degrees in linguistics and translation, respectively, with at least 20 years of teaching experience in universities related to translation and interpreting courses.

First, they are required to review the generated translation items and passages respectively according to the syllabus of English-Chinese and Chinese-English translation and interpreting courses. Second, they discuss and reach an agreement when the generated items are difficult to categorize. A checklist for RQ1 and RQ2 is as follows:

1) Does the test item reflect the difference between Chinese and English languages?

2) Does the test item reflect the difference between Chinese and English cultures?

3) Whether the test item reflects the differences in thinking-pattern features between Chinese and English?

4) What specific translation strategy is used in the English-Chinese sentence translation item?

5) What specific translation strategy is used in the Chinese-English sentence translation item?

6) What are the text type and topic of the Chinese-English translation passage?

7) What is the text type and topic of the English-Chinese translation passage?

4.3.1. Content Validity of English-Chinese Sentence Translation Items

The generated English-Chinese sentence translation items by the seven GenAI models are listed in Table 1. As for task 1, Deepseek-r1:1.5b failed in the translation direction; what it generated were not English-Chinese but Chinese-English sentence translation items. The other six GenAI models completed task 1 fairly well, including a variety of topics, such as cultural image, technological term, legislative text, euphemism, etc. The translation skills mentioned above are employed in different translation items. The first English-Chinese sentence translation items generated by the six GenAI models are chosen at random and listed in Table 1 below.

Table 1. Generated English-Chinese sentence translation items.

(1) ST: The professor’s lecture was such a white elephant that most students dozed off halfway through.

TT: 教授的讲座华而不实，大半学生听到一半就昏昏欲睡。(DeepSeek R1)

(2) ST: The arbitration award shall be final and binding on both parties.

TT: 仲裁裁决应是终局的，对双方均具有约束力。(DeepSeek V3)

(3) ST: The revolution in technology has led to a surge in remote work opportunities.

TT: 技术领域的革命已经导致远程工作机会激增。(ERNIE Bot)

(4) ST: Time is money.

TT: 一寸光阴一寸金。(Qwen)

(5) ST: He’s a real night owl — still working at 2 a.m.

TT: 他是个十足的夜猫子，凌晨两点还在工作(Qwen-VL)

(6) ST: The circular economy aims to close the loop between production and consumption by reusing resources indefinitely.

TT: 循环经济的目标是通过无限循环利用资源，实现生产与消费之间的闭环。((Qwen-Plus)

In example 1, “white elephant” means “a possession that is useless or troublesome, especially one that is expensive to maintain or difficult to dispose of” ([22]), and there is no such cultural image in Chinese, so it is translated into Chinese “华而不实 hua er bu shi” by liberal translation.

In example 2, the usage of the modal verb shall in legislative text is the key, which means “1. has a duty to; more broadly, is required to.” and “This is the mandatory sense that drafters typically intend and that courts typically uphold.” ([7]) The modal verb shall is translated into Chinese “应 ying”, showing its mandatory force.

In example 3, the noun surge means “a sudden increase in amount or number”, and was translated into the Chinese verb “激增 ji zeng”, which employed the conversion of parts of speech.

In example 4, the proverb “Time is money” means “time is a valuable resource, therefore it’s better to do things as quickly as possible.” ([22]), while its Chinese translation is an equivalent Chinese proverb “一寸光阴一寸金 yi cun guang yin yi cun jin”, with a metaphor in it. What’s more, the adapted translation was employed in example (4).

In example 5, the phrase “night owl” means “a person who enjoys staying up late at night”, and its Chinese equivalent “夜猫子 ye mao zi” has the same meaning, which illustrates the cultural universals of animal metaphor in this respect.

In example 6, the logic in English is from aim to means, while the logic in Chinese is from aim to result via means. Therefore, the sentence order was adjusted in the English-Chinese translation process.

From the examples above, we can conclude that the content validity of the English-Chinese sentence translation items is justified.

4.3.2. Content Validity of Chinese-English Sentence Translation Items

As for task 2, Deepseek-r1:1.5b and ERNIE Bot failed. What the former generated were only some Chinese words or phrases as the source text items, with full Chinese translations. What the latter generated was “Please translate… (words or phrases) into English.” The other five models completed task 2 fairly well, including a variety of topics, such as cultural images, technological terms, legislative texts, euphemisms, etc., whose topics are listed in Table 2. The translation skills mentioned above are employed in different translation items. The first Chinese-English sentence translation items generated by the five GenAI models are chosen and listed as follows:

(7) ST: 塞翁失马，焉知非福。

TT: When the old man of Sai lost his horse, who could have known it was not a blessing in disguise? (DeepSeek R1)

In example 7, “塞翁失马 sai weng shi ma” is a Chinese allusion. The English translation retains the core image of “塞翁 sai weng (the old man on the frontier)” and “马 ma (his horse)”, and then a Western expression, “a blessing in disguise”, was employed to reveal its true meaning. The literal translation strategy was used by DeepSeek R1, but it mistranslated the word “塞 sai” as a surname.

(8) ST：这位作家是文坛常青树，笔耕不辍六十载。

TT: This writer is a literary evergreen who has kept writing diligently for sixty years. (DeepSeek V3.1)

In example 8, “常青树 chang qing shu” is a literary allusion, meaning that some writer is still popular even when he or she is fairly old. Chinese and English share this literary allusion, so the literal translation was employed by DeepSeek V3.1, which vividly maintains the plant metaphor.

(9) ST: 他虽然年纪大了，但精神矍铄，每天坚持晨跑五公里。

TT: Although he is advanced in age, he remains mentally and physically vigorous, jogging five kilometers every morning without fail. (Qwen)

In example 9, the Chinese words “年纪大了 nian ji da le” cannot be translated into “old”, which is possibly derogatory, so the expression “advanced in age” was employed; and there is no equivalent expression for the Chinese four-character idiom “精神矍铄 jing shen jue shuo”, so it was paraphrased as “mentally and physically vigorous”.

(10) ST：他这人外强中干，一遇压力就原形毕露。

TT: He’s all bark and no bite; under pressure, his true colors show. (Qwen-VL)

In example 10, the literal translation for the two Chinese four-character idioms “外强中干 wai qiang zhong gan” and “原形毕露 yuan xing bi lu” would be somewhat stiff, i.e., “be outwardly strong but inwardly weak” for the former ([25]) and “be revealed for what one is” for the latter ([25]), so liberal translation was employed by Qwen-VL.

(11) ST：中秋节吃月饼的习俗象征团圆和丰收。

TT: The Mid-Autumn Festival tradition of eating mooncakes symbolizes family reunion and a bountiful harvest. (QwQ-Plus)

In example 11, the traditional Chinese festival “中秋节zhong qiu jie” symbolizes “family reunion”, but not “bountiful harvest”. The QwQ-Plu made a mistake here, which is called the hallucination of GenAI tools ([26]).

Based on the analysis above, the content validity of the Chinese-English sentence translation items is also justified.

Table 2. Generated Chinese-English sentence translation items.

Undergraduate students majoring in translation and interpreting, after professional training, are capable of handling the translation of comprehensive texts in current affairs and news, as well as professional texts of general difficulty in fields such as politics, economy, society, and culture ([31]). Therefore, this section will check the content validity of the generated passages for English-Chinese and Chinese-English text translation, respectively, in terms of topics, discourse classification, and text readability. In order to assess text readability within a uniform framework, the source texts of task 3 and the target texts of task 4 are analyzed.

4.4.1. Topics, Discourse Classification, Statistics, Readability, and Content Validity of the Generated English-Chinese Translation Passages

After inputting prompt 3 into the GenAI models, the topics, discourse classification, passage statistics, and readability details of the generated passages are listed in Table 3.

A variety of topics are generated, including political philosophy, AI ethics, urban loneliness, digitization of traditional crafts, etc., which cover the comprehensive texts and professional texts required, and the GenAI can generate texts on various topics.

Discourse classification mainly consists of argumentation, exposition, narration, and description ([9]). The generated passages are all argumentation except text 7, and most of the generated argumentative passages include elements of claim, evidence, and conclusion ([18]).

Example (12):

[1] The rapid development of generative AI has sparked intense ethical debates [2]. While these systems demonstrate remarkable creativity—producing original art or composing music—their “black box” nature raises concerns about accountability [3]. When an AI-generated image violates copyright laws, who bears responsibility [4]? The programmer, the user, or the algorithm itself [5]? This dilemma is compounded by cultural differences: Western frameworks emphasize individual liability, whereas Eastern philosophies often view responsibility as collective [6]. Furthermore, the anthropomorphic tendency to describe AI as “learning” or “thinking” obscures its mechanistic essence, potentially misleading the public [7]. Resolving these issues requires not only technical transparency but also cross-cultural dialogue to redefine ethical boundaries in the digital age. (DeepSeek R1)

In example 12, sentence 1 constitutes the claim, sentences 2 - 4, 5, and 6 are the evidence, and sentence 7 the conclusion.

Passage statistics cover length, number of sentences, words per sentence, and the percent of difficult words, respectively. Nine texts have more than 100 words; the other five texts have fewer than 100 words, especially texts 1 - 2 and texts 7 - 8, which have only about 30 or 50 words, respectively. Among them, nine texts have at least seven sentences, which are conducive to the complete structure of an argumentation, except for texts 1 - 2 and texts 7 - 8. The percent of difficult words ranges from 30% to 40%, except for texts 1 - 2 and texts 7 - 8.

As for text readability, Flesch Reading Ease is discussed ([27]). Flesch Reading Ease scores text readability on a 100-point scale, and there are seven levels of text difficulty: 0 - 29 (very difficult), 30 - 49 (difficult), 50 - 59 (fairly difficult), 60 - 69 (standard), 70 - 79 (fairly easy), 80 - 89 (easy), and 90 - 100 (very easy). According to the Flesch Reading Ease scores, texts 3 - 6 and texts 9 - 13 are very difficult, and text 14 is difficult.

Through the analysis of topics, discourse classification, statistics, and readability, we can conclude that the content validity of the generated English-Chinese translation passages is justified.

Table 3.Generated English-Chinese text translation passages.

4.4.2. Topics, Discourse Classification, Statistics, Readability, and Content Validity of the Generated Chinese-English Translation Passages

After inputting prompt 4 into the GenAI models, the topics, discourse classification, passage statistics, and readability details of the generated passages are listed in Table 4.

A variety of topics are generated, including sharing bikes, the Dragon Boat Festival, AI in medical care, urbanization, and traditional villages, etc., which cover the comprehensive and professional texts required, and the GenAI tools can generate texts on various topics.

As for the discourse classification, the generated passages are all argumentative.

As for the passage statistics, ten texts are more than 100 words, and the other two texts are less than 100 words, except texts 1 - 2. Among them, ten texts have at least 6 sentences, which helps to guarantee the complete structure of an argumentation, except texts 7 - 8. The percentage of difficult words ranges roughly from 20% to 35%.

As for text readability, Flesch Reading Ease is discussed ([27]). According to the Flesch Reading Ease scores in Table 4, texts 3 - 6 and texts 9 - 11 are very difficult, and texts 12 - 14 are difficult.

Through the analysis of topics, discourse classification, and readability, we can conclude that the content validity of the generated Chinese-English translation passages is justified.

Summarizing the generated results from Tables 1-4, we can see that DeepSeek V3, DeepSeek R1, Qwen, Qwen-VL, and QwQ-Plus completed the four generating tasks fairly well. More than ninety percent of the test items can be used directly, with only some minor corrections needed, but they are models for general purposes, so more models for vertical sectors, i.e., higher education, should be developed for higher quality generating results.

Table 4. Generated Chinese-English text translation passages.

Deepseek-r1:1.5b is a lightweight model for decoder-only, and it is good at generating texts in Chinese, so it can explain to some degree why it failed the four tasks of translation questions.

Li Yanhong, the CEO of Baidu, put it, “Baidu’s Ernie large model is a highly localized large language model in the Chinese market. This means that the Ernie that Baidu is currently developing will be more suitable for the Chinese language and the Chinese market than models developed abroad” ([28]). Therefore, it does not excel at text generation in English, and it failed tasks 2 - 4 of translation questions.