1. Introduction

Taboo expressions are among the most sensitive areas of language because they convey meanings beyond those associated with the words themselves. Taboo language includes terms or phrases that are perceived as offensive, rude, or socially inappropriate by members of a certain cultural group (Allan & Burridge, 2006; Jay, 1992). Cultural norms are not the same for all societies; thus, an expression that is taboo in one culture could be acceptable in another culture. This means that the translation of taboos goes beyond the simple task of finding equivalent terms.

In translating taboo expressions, the difficulty arises in the fact that the taboo expressions serve different functions in communication. Apart from conveying an expression of annoyance, a taboo expression could also be used to express anger, intimacy, humor, social identity, power struggle, or even opposition to prevailing ideologies (Jay, 2009). Taboo expressions in literature are normally used by writers with the intention of constructing their characters, bringing out social injustices, contesting against prevailing ideologies, or simply projecting their stylistic voice. The translator, thus, has to deal with the challenge of balancing between semantic accuracy and cultural appropriateness of the translated expression.

Previous studies have categorized the elements of taboo speech into various types such as profanity, obscenity, vulgarity, epithets, insults and slurs, scatology, and cursing (Jay, 1992). These categories do not denote separate groups of vocabulary but rather different levels of cultural and emotional sensitivity which call for diverse translation strategies. Therefore, the translation of taboo speech is not an easy task that only requires linguistic skills, but intercultural sensitivity as well (Allan & Burridge, 2006).

Recent developments in AI technology have revolutionized translation by way of developing advanced language models like ChatGPT and DeepSeek. Both have displayed excellent results in translation tasks for various languages. However, literary translation poses one of the most difficult challenges to AI because literary content often relies heavily on implicit cultural connotations and emotions. This problem becomes even more pressing in the case of taboo expressions, which involve making ideological choices about how language is used within specific cultural contexts. While many scholars claim that recent developments in generative AI show promising results in ordinary translation tasks, there is still insufficient research on its translation of literary content, especially taboo expressions in Arabic literature.

Additionally, previous work has been mainly concerned with measuring the quality of translation or comparison between AI and human translators in terms of their performance in translating technical and non-literary texts. Relatively little research has been conducted regarding taboo translation in literature, and even less has used a systematic approach for assessing the quality of translation to compare human translators with AI. It is especially important to take into account the need for context-based decisions in literary translation, which brings up a question about how well the current generation of AI is capable of maintaining the role of taboos in language.

In order to fill the gap mentioned above, the current research seeks to compare the English translation of Nawal El Saadawi’s Woman at Point Zero by Sherif Hetata (1983) with the translations produced by the AI translators ChatGPT and DeepSeek. A total of 50 taboo terms have been collected manually from the Arabic novel for analysis. To assess the quality of the translations, the MQM criteria will be employed, which include four main dimensions – namely, accuracy, fluency, terminology, and style. Moreover, some chosen examples will be analyzed in depth to see how translation strategies used by the human translator and the AI models impact the transmission of cultural meaning and stylistic elements.

This paper adds to the emerging body of research on literary translation through AI in several key ways. It is one of the first studies that compares human and generative AI translations of taboo terms in Arabic based on the MQM system. The paper addresses the issue of the current big language models’ handling of the conflict between adaptation and faithfulness in literary translation. Finally, it offers valuable information on both the potentials and constraints of AI-assisted literary translation.

Therefore, the present study attempts to answer the following research questions:

1. What is the nature of the differences between the human translator’s translation, ChatGPT’s translation, and DeepSeek’s translation of the taboo terms used in Woman at Point Zero in terms of accuracy, fluency, terminology, and style using the MQM approach?

2. What are the translation techniques adopted by the human translator, ChatGPT, and DeepSeek for translating taboo terms from Arabic into English and what are their implications on meaning transfer?

2. Literature Review

2.1 Theoretical Background

Taboo expressions are one of the most culturally delicate aspects of translation since they acquire their meaning not only based on their vocabulary but also on social conventions, religious beliefs, gender roles, ideologies, and contexts. Taboo expressions are words or phrases that are considered offensive, rude, or socially inappropriate in a particular society (Allan & Burridge, 2006; Jay, 2009). Yet, taboo language does not have an intrinsically offensive nature. Its offensiveness is dependent on who says it, whom it is directed at, the context in which it is used, and its pragmatic role. In the field of literature, taboo expressions may have several purposes; they may be used to describe a character, express rage, unveil violence, criticize social hypocrisy, or voice marginalized voices.

The translation of taboos in literary texts is especially difficult since they tend to have symbolic meaning as well. It is possible to use literal translation, euphemisms, omission, substitution, or even culturalization when translating such words, but all of these ways influence the way in which the target audience will perceive the text (Zagood et al., 2022). While literal translation may keep the shock effect from the source text, it will be unnatural and too shocking for the target audience. Euphemisms may make the translation less shocking, but they also make the emotion weaker, and the ideology of the text is lost. Moreover, omission may cause the loss of essential meaning related to characterization, realistic depiction, and social criticism ( Lestari, 2023).

The problem is even more acute in the case of feminist literary translations. Feminist translation theory understands translation as an ideological practice rather than a simple linguistic transfer (Simon, 1996; Castro and Ergun, 2017). In a feminist text, the use of taboo terms could serve as a means to unveil gender oppression, exploitation of the body, violence, and control by patriarchy. The novel by Nawal El-Saadawi, Woman at Point Zero, can be used as an illustration in this regard since the body, sexuality, violence, and women’s resistance become key elements in a patriarchal society of Egypt. Taboo terms are integral parts of the feminist discourse in the novel, and their softening or neutralization will result in the loss of its radicalism.

Pragmatic equivalence is the key to this debate. Baker (2018) suggests that when translating from one language into another, a translator has to take into consideration the role of context, implicature, and communicative purpose. Likewise, House (2015) believes that translation quality has to be judged based on how well the translated text accomplishes the same communicative task as the source text. As far as taboo words are concerned, pragmatic equivalence implies not just conveying the lexical meaning of words but also the desired effect that might be shock, humiliation, anger, resistance, or any other emotion.

Recently, machine translation and artificial intelligence developments have added yet another aspect to this debate. Machine learning models like ChatGPT and DeepSeek are capable of delivering fluent and consistent translations. However, literary translation is quite a complicated task since it relies on implied meanings, styles, cultural memories, and ideological positioning. Scientific evidence recently gathered concerning machine literary translation demonstrates that post-editing of creative translations must balance efficiency and style/creativity preservation. For instance, a recent 2024 paper written by Macken explores the potential of ChatGPT in post-editing literary machine translations, while Castaldo et al. (2025) analyze LLM-generated literary translations along with their post-editing by professional translators.

2.2 Empirical Studies

From recent research on artificial intelligence-assisted translations, it is evident that large language models have demonstrated high levels of effectiveness in general translation, although their efficiency depends on the text types, language pairs, domains, and evaluation measures used. According to Zhao et al. (2023), ChatGPT was assessed using human assessment of its translation performance in various domains and was found to perform effectively in general translation, while it lacked domain-specific terminology and style in some cases. This applies to the current study due to the context-dependence of taboo terms in literary language.

Similarly, Hendy et al. (2023) analyzed the use of GPT in machine translation and found that the models can compete with neural machine translation. Nevertheless, there are studies comparing human translation, neural machine translation, and ChatGPT translation results, which show that while being good in general translations, LLMs do not have to be good at translating voice, ideology, or culturally meaningful nuances. This is important because the book under consideration implies the translation of taboo expressions used in the context of feminist theory, social criticism, and psychological trauma.

Research in machine translation of literature has paid attention to the limitations of machine translations. Guerberof-Arenas and Toral (2022) concluded that machine-translated literary texts require post-editing that might limit the creativity of the translators and affect the effectiveness of their translations negatively. The following works have kept addressing the role of AI for literary translators instead of taking their jobs. Macken (2022) has discussed literary translation as a process consisting of machine translation, post-editing, and revision.

Recent research has been centered around LLMs and literary post-editing. Macken (2024) analyzed the possibility of automatic literary post-editing by ChatGPT and concluded that LLMs could enhance the qualities of machine translations, although professionals would be required to ensure literary translation. Additionally, Castaldo et al. (2025) revealed that the use of LLMs for creating literary translations could save the time spent on editing while preserving a certain degree of creativity; however, the research is another proof that the evaluation and intervention of humans are required when it comes to creative translation.

Previous research in the area of taboo translation tends to concentrate on audiovisual translation, such as film subtitling and dubbing, as well as Netflix series and TV drama translation. It should be noted that such research is important and informative since it helps understand how tabooed language tends to get mitigated, censored, and adapted for target audiences. However, the results of such research cannot be applied directly to literary translation since, in contrast to audiovisual translation, it is not limited by time, space, and multimodality but rather depends on narrative voice and character description. That is why the current research is going to concentrate on Arabic feminist literary discourse.

Scholars who have studied the writings of Nawal El-Saadawi have largely centered their discussions around feminism, patriarchy, sexual violence, and the resistance of women. Analysis of Woman at Point Zero usually concentrates on the opposition of Firdaus to oppression from the institutions, sexually, and socially (Amireh, 2000; Balaa, 2018). Yet there has been insufficient research on how the taboo expressions in the book are translated into English, and how AI machines translate them. This is because the taboos in El-Saadawi’s writings are very related to the concepts of voice, body, oppression, and resistance.

Al-Yasin and Rabab’ah (2019) investigate how taboo language is rendered in American Hip Hop films that are subtitled into Arabic, with special interest in fan-produced subtitles. The study reviews 512 occurrences of English swearwords from three profanity-laden films, where official English subtitles are compared with those of Arabic fan-made ones. Based on Ljung’s (2011) swearing typology, the researchers classify the English swearwords according to function and intensity, and examine the way in which these expressions are translated into Arabic. Four major strategies used in fan subtitles are determined by the study: Literal translation of taboo expressions, cultural substitution (i.e., taboo replaced by a culturally acceptable equivalent), omission or deletion of taboo terms, and, lastly, euphemistic rendering (i.e., expressions minimizing the shock in the target language). The findings indicate that rather than using additions, the Arabic fan subtitlers preferred either substitution or omission as they were sensitive to the norms of the target audience. The study also highlights the emergence of tension between the need to retain the expressive strength of the source text and the need to be culturally conformist in the Arab context (Al-Yasin and Rabab'ah, 2019).

Abedaal and Al Sarhani (2021) investigate the translation of English swear words and taboo expressions into the Arabic subtitles. Using a mixed-method approach, they used a qualitative approach to identify translational strategies and quantitative methods to assess subtitle quality. The study comprises 40 cases of swear words and taboo expressions in the 2001 American crime drama “Training Day”. The analysis exploited Pedersen’s (2005, 2011) paradigms of a subtitling strategy type and a quality assessment model with the tweaks made by Abdelaal (2019). It was established that the most recurrent way used was the usage of omission and euphemistic translation, which was caused by cultural sensitivities as well as the conservatism of the Arab audiences. Other issues were: holding on to language, specifying, generalizing, and formalizing the informal language. Results of the quality assessment, which was designed taking into account the FAR model (Functional equivalence, Acceptability, Readability), showed that the majority of subtitles were at a good level of quality with minor incongruities that did not distort the impression about the viewer’s comprehension. The research finds that although omission and euphemizing are effective devices and are culturally constrained, they are still necessary to strike a balance of sensitive subtitling and retention of the expressive content of the source material. The authors suggest reworking the existing models of the quality assessments to incorporate the quantitative and qualitative criteria for a more integral analysis (Abdelaal and Sarhani, 2021).

Haider et al. (2023) examine how the text editors translate taboo expressions from the Jordanian Arabic Series Jinn to English. This series was selected because of its importance as Netflix’s first Arabic-Jordanian production of explicit content and culturally sensitive language. The cultural taboo expressions were translated into English using four main strategies: Transforming a source culture taboo into a target culture taboo of the same or greater or smaller strength, omitting the source taboo expression, replacing it with a neutral or less offensive one, and using euphemism or circumlocution. The results demonstrate that the majority of taboo expressions were borrowed using substitution and euphemism to fit the target culture norms. However, in many instances, the original level of offensiveness was achieved or increased. Moreover, the study indicates that the most commonly spoken implied functions of taboo language were ‘unfriendly suggestions’ and ‘noun supports’. It recommends future research on cross-cultural taboo-language transfer within subtitled and dubbed media (Haider et al., 2023).

Olimat et al. (2023) analyze the subtitling of sexually sensitive Arabic utterances for English-speaking viewers in two Netflix series: Jinn and Al-Rawabi School for Girls. The study examines 284 cases of sexually sensitive or culturally taboo language found in the original Arabic dialogue, in terms of how they are translated in the subtitles placed by Netflix for the English version. The researchers identify five major strategies utilized in translation process: Direct translation of the taboo lexeme from taboos to taboo lexeme direct standards, substitution of culture in this case, replacing the Arabic phrases that offend with the more contextually appropriate adventures of English, generalization that, a reduction in specifics and impact, omission which removes the sensitive term from the text and euphemistic rendering to palliate that disruptiveness of the original text.

The outcomes demonstrate that even though direct translation and cultural substitution were the most widely employed and effective strategies, the excessive use of generalization and omission often resulted in diluted meanings and a diminished effectiveness on the target audience. Furthermore, the study covers the other data set from the American Sitcom How I Met Your Mother, and here the authors describe the subtitling for the comedic and the sexually suggestive dialogue. It is observed that the timeliness of conveying explanatory subtitles during short micro dialogic sections is dramatically compressed in its comedic feel and the delivery of comedy. In conclusion, the results indicate that translating culturally sensitive content, such as sexual or religious themes, is a complex task. The translator must accurately convey the intended meaning of the original text while simultaneously showing sensitivity and respect toward the target culture. Striking this balance presents a significant challenge, especially when there are notable cultural differences between the source and target audiences (Olimat et al., 2023).

Badmus et al. (2024) investigate taboo expressions in the Idoma language pragmatically, and they focus on the courtesy and indirect nature of such terms. The research comprises 215 examples of taboo expressions distributed with reference to their social, sexual, and cultural relevance. The survey views those expressions as words that are regularly used by adults to teach their young children the virtues of moral behavior, but are avoided in a public setting, for social reasons. The authors use a typology of taboo language, distinguishing between those used in private and those in the public domain, and dealing with the indirect or euphemistic ways in which they are mellowed. The same strategies that are used by the speakers are also categorized according to the study: use of taboo terms directly, omission of taboo terms, circumlocutions or euphemisms, and culturally specific terms that describe community perception. The findings suggest that taboo language should remain for private conversation, and critical documentation of such expressions is very important for the preservation of sources of the Idoma language. The research also calls for greater analysis on the cultural and Ideological character of taboo language in the teaching of the youth on proper use and avoidance of such jargon (Badmus et al., 2024).

Shormani (2024) analyses the accuracy of translating Arabic oath phrases into English by ChatGPT and whether ChatGPT can capture swearing nuances or not. The research analyses 30 Arabic oaths drawn from literary sources, namely, it compares the transcripts of the AI and the human translations. The research states that there are really many weaknesses in how ChatGPT represents the subtlety of Arabic oaths, namely the inability to understand the religious figure specific to culture, the non-oath aspect of the Arabic language. Translation strategies assessed include direct translation of oaths, contextual alignment to fit in issues concerning the target language, removal of culture-specific elements or expressions, as well as literal translation of non-oaths’ aspects. The finding points to the fact that the translations of the Arabic oaths made using ChatGPT never guarantee the desired cultural as well as emotional impact of the Arabic oaths, which requires further improvement of the AI-based translation systems. The study concludes with proposals towards obtaining more data for subsequent training on the Arabic oath expressions and their cultural contextualization to build a better-performing system (Shormani, 2024).

Brahimaj (2024) investigates the sociolinguistic nature of taboo language in Kosovar society. The research involves the usage of a combination of quantitative and qualitative methodologies (mixed-methods approach), meaning that the given study permits having an overall idea about the usage and understanding of expressions that can be attributed to the taboo class. Data collection was made by a structured survey based on ten questions, in which patterns of usage and frequency of usage, social acceptability, and perceived offensiveness of taboo words were sought. The study was based on a sample consisting of 101 subjects of both sexes whose ages ranged from 17 to 60 years. Other than personal use of taboo language, the study acknowledged its relevance in context as well as the external influences that make an individual prone to uttering such language, like media exposure. Notwithstanding, no specific theoretical framework is referred to by name, yet the research is carried out in the context of the sociolinguistic approach to language use and identification of the way taboo language works on the informational level in social relations and how cultural and societal norms form a part of its use. It also tries its hand at language evolution in ways that implicitly concur with theories that deal with the dynamics and adaptiveness of language in society.

Al-Harahsheh and Shehab (2024) assess the ideological, cultural, and social implications related to the epithet parts of El-Saadawi's Woman at Point Zero's English cover pages. The study employed Pierce’s model of sign process. The study found that these epigraphic elements reflect El-Saadawi’s literary style in writing and the deep social, ideological, and cultural implications that she struggles to convey about the Egyptian society’s view toward women. In addition, it found that these epigraphs affect the readers’ perception and interpretation of the novel’s core message, theme, style, and its cultural context.

At broader conclusions, the study reports that the taboo language is common in the Kosovar society and is dynamically influenced by factors such as age in society, social environment, mood, media, among others. The use of taboo expressions is accompanied by anger or pain, and it depends on the context or the audience. The study is centered around the dynamism of language and the possibility of some change in the norms related to the changing opinions on the taboo expressions. According to the results, it is necessary to continue considering the connection between the utilization of taboo language and the general cultural and social transformations (Brahimaj, 2024).

Fitriansyah (2023) investigates the taboo in the Gran Torino Movie, which looks at the translation of such words in the film. The study assessed the subtitling of 270 taboo words that were subtitled into Arabic. Experts were brought together in FGDs to determine the quality of translations. The research divides taboo words into ten groups, and among them are sexual acts, profanity, insults about race, and scatological references. They served four main functions: They involve the option of connecting, disconnecting, or distinguishing words or speakers; in addition, they were called auxiliary, abusive, expletive, or humorous by some. Most taboo words went through changes in meaning and role when moved from one language to another. More often than not, the approach taken for translation involved reduction, with established equivalent, exploitation, discursive creation, and literal translation being the next most popular. Still, reduction was the primary method used, but this lowered the quality of translations. From the assessment of the authors, Nababan et al. (2012), the scores were 1.78 for accuracy, 2.12 for acceptability, and 1.92 for the total quality, suggesting a poor overall result. It summarizes that avoiding taboo vocabulary is okay, but using substitutions is more helpful for preserving the meaning and intensity (Fitriansyah, 2023).

The reviewed literature, therefore, reveals three major gaps. First, recent studies on taboo translation remain heavily oriented toward audiovisual translation, with limited attention to literary texts. Second, research on El-Saadawi has emphasized feminist themes but has rarely examined taboo expressions from a translation-quality perspective. Third, although recent studies on ChatGPT, LLMs, machine translation, and post-editing are rapidly increasing, few have examined how generative AI translates culturally sensitive taboo expressions in Arabic feminist literature. The present study addresses these gaps by comparing Sherif Hetata’s human translation of Woman at Point Zero with translations generated by ChatGPT and DeepSeek, using the MQM framework to evaluate accuracy, fluency, terminology, and style, while also considering pragmatic equivalence and feminist literary context.

3. Methods and Procedures

3.1. Data collection

The dataset for the current study was compiled out of 50 Arabic taboos manually collected from the original Arabic novel titled “Woman at Point Zero” by Nawal El Saadawi. These taboos were then paired with the official English translation of the book provided by Sherif Hetata (1983). The same taboos were translated via ChatGPT and DeepSeek with the same set of prompts used for all three translations in order to ensure consistent comparison. The 50 collected taboos belong to various categories, including sexual, gender, violent, body, religion-related, and other culturally sensitive taboos relevant to the feminist discourse of the novel.

All 20 expressions (See Appendix A) that were extracted for analysis were used for the quantitative study through the MQM approach. With this in-depth analysis, the researchers were able to analyze and compare the quality of the translation of the three translation results and find out the most prevalent patterns and errors in translation. The overall MQM scores from the whole corpus are shown in Figure 1.

For the qualitative analysis, eight examples were chosen through purposive sampling. This was done based on pre-defined criteria of methodology rather than convenience. First, the chosen examples are representative of the different categories of taboos, namely sexual, gendered, violent, and culturally-sensitive. Second, these examples show the various methods used by the human translator, ChatGPT, and DeepSeek to translate taboo expressions, which include, but not limited to, literal translation, euphemization, omission, cultural substitution, and semantic change. Third, the examples were selected as they have interesting linguistic, pragmatic, and cultural features that would allow discussing the issues in the translation of taboo expressions. Lastly, the chosen examples are those that had the highest frequency rates for the MQM error types found in the qualitative analysis of the full dataset. Therefore, the qualitative analysis aimed to exemplify and explain the tendencies revealed throughout the analysis of all 50 expressions rather than to substitute the results of the quantitative analysis. The use of only eight examples, however, allowed conducting a deep enough analysis without exceeding the scope of a journal article.

The choice to use Nawal El Saadawi’s Woman at Point Zero is due to the presence of many taboo terms that are intricately related to problems of gender, sexuality, violence, religion, and oppression in society. These aspects of the book make it suitable for analyzing the ways through which human and modern AI translators translate culturally-sensitive texts in literature.

3.2. Data Analysis

This research utilized a mixed-method analysis approach, which is a combination of qualitative and quantitative research designs. During the initial phase of the analysis, the translations done by Sherif Hetata, ChatGPT, and DeepSeek were assessed quantitatively based on the MQM empirical metrics framework. The MQM was chosen since it offers an international model for evaluating translation quality. MQM assesses translation quality in relation to accuracy, fluency, terminology, and style; hence, it is important in evaluating translations of culturally sensitive literature.

The translation errors were assigned based on their severity levels in terms of MQM. Errors of minor severity (weight = 1) are those errors that have no influence or very little influence on the overall understanding or reading of the translated text. Errors of major severity (weight = 5) have a considerable effect on the usability or understandability of the translation, but the meaning is maintained nonetheless. The errors of critical severity (weight = 10) are those that change the meaning of the source text significantly, thus having the most significant impact on translation quality. All of the 50 items were assigned their score for MQM in a consistent manner.

After performing quantitative analysis, the study carried out a detailed qualitative analysis of the eight selected examples. The objective of this phase was to provide a rationale for the results of the quantitative analysis by examining how each translation dealt with cultural, pragmatic, and ideological aspects of the taboo expressions in question in terms of the literary context in which they occurred.

The identification and comparison of the translation strategies used by the human translator, ChatGPT, and DeepSeek in this research have been conducted using the classification developed by Khoshsaligheh and Ameri (2014). According to this classification, taboo expression translation is divided into four main strategies, namely: (1) taboo to taboo; (2) deletion; (3) taboo to non-taboo; and (4) euphemism. In order to conduct an in-depth comparison of translation quality and strategy usage, all eight selected examples have been analyzed on the basis of two classifications – the MQM one and the strategy taxonomy.

Findings of the study

The analysis focused on how the taboo translations were subtitled and translated in terms of fluency, accuracy, and the ability to deal with the social and cultural context of these expressions. The MQM framework was used to accurately identify the types of issues and categorize errors, to assess the quality of the translation. There was a clear contrast between AI translation and human translation, highlighting the strengths and limitations of each type of translation.

The above expression in Table 1 is considered a vulgarism since it describes female genital mutilation through the depiction of the injury to the body. In this case, the human translator succeeded in accurately translating the taboo and maintaining the feminist interpretation of El Saadawi’s message about the violation of women's rights. There are no MQM errors in this translation apart from the inaccurate translation of the word مطواة into “small knife” rather than into its actual English translation. However, this translation still conveys the full meaning and emotional intensity of the message without any MQM errors. On the contrary, ChatGPT offers a better translation of مطواة as “pocketknife.” Thus, ChatGPT manages to convey both semantic and pragmatic equivalence with no MQM errors. Nevertheless, DeepSeek keeps the taboo reference explicit; however, مطواة is translated as “dagger,” which constitutes a terminology error since the lexical item used does not correspond to the actual type of weapon used. As for the translation strategy, human translation and ChatGPT use taboo-for-taboo transfer; they use the domestication strategy and maintain the face-threatening acts of the expression in the target language. However, DeepSeek uses the combination of literal translation and lexical substitution.

The expression in Table 2 falls into the category of vulgarity because it denotes an injury of a physical nature and the feeling of pain and suffering, which underlines the protagonist's trauma and the feminist representation of violence done to women in the novel. Human translation manages to keep all the information provided in the source text, including the sequence of actions, imagery, and simile, without any changes, making it a fully coherent text with no MQM errors. The translator employs the domestication strategy to transfer the taboo meaning to the target readers. ChatGPT, despite being fluent, makes a significant mistake in terms of accuracy because it uses the pronoun "it" where "him" should be used in order to make a coherent passage. DeepSeek offers the most accurate rendering of the source text, making sure that referents are correct as well as imagery and metaphors are kept with the help of such lexemes as "waterskin," "stagnant water," and "grease." It also employs the domestication strategy. There are no MQM errors in the translation.

The above example is regarded as an instance of obscenity because it features a clear bodily reference with the notion of physical intimacy and the protagonist's repulsion towards the male character. Human translation successfully conveys the message of the original text with all imagery and communicative value being preserved, which leads to zero MQM mistakes. Clearly, the translator utilizes the domestication strategy. Meanwhile, ChatGPT makes many changes to the original phrase, including some events that do not appear in the source text, such as "I curled my lips to spit in his face" and "he caught my lips with his teeth". Such modifications lead to the distortion of events' sequence and constitute a significant accuracy problem according to the MQM guidelines, leading to the lack of semantic and pragmatic equivalence. Nevertheless, DeepSeek manages to reproduce both the bodily reference and the events' sequence via phrases like "viscous saliva" and "I thrust it back into his mouth" without creating any MQM errors. Concerning the translation strategy, taboo-to-taboo transfer was used by the human translator and DeepSeek, while interpretive mistranslation occurred in ChatGPT.

The above statement in example 4 is considered to be an example of offensiveness because it clearly refers to the woman's body. It further supports the feminist nature of this novel, where females appear to be vulnerable. The human translator has maintained the same message in the target language, but the phrase "من غير مشد" has been skipped, which makes the statement less offensive. Therefore, this is a minor omission error according to the MQM criteria. However, the translator transfers the implicature or the intended meaning accurately to the target readers. On the other hand, ChatGPT has kept the same source meaning intact through the translation "My breasts show through it without a bra". Likewise, DeepSeek has maintained the source meaning through the phrase "unsupported," which, although less obscene, semantically conveys the source message. Still, AI translation maintains the implicature or the intended meaning in the target text. None of the AI translations contains any MQM errors. Talking about translation strategy, both the AI translations have used the taboo-to-taboo strategy, while human translation is based on the partial omission strategy, thus making it less pragmatic.

This example is considered to be obscene language as the text includes a direct reference to the female body by means of the noun nipple. Thus, the given example reveals the problem of female objectification and sexual vulnerability described in the novel. As far as the human translation is concerned, it can be said that it perfectly captures the intended meaning of the source text and maintains its pragmatic effect since there is no information loss or distortion in the translation. ChatGPT translation is also faithful, and there are no MQM errors in it since the expression "My nipple, a prominent dark circle" demonstrates lexical precision and does not involve any modification of the original text. Both human translator and ChatGPT employ a domestication strategy. However, the translation produced by DeepSeek is flawed with one serious mistake, as the translator used the expression "In his dream: a dark, circular breast" in the place of "حلمة الثدي" thus adding some new information to the source text and distorting the meaning of the text. Such semantic distortion is caused by the fact that it changes both the propositional meaning and the pragmatic effect of the text, making it less faithful to the source text. Pragmatically speaking, it is clear that the translation fails to capture the communicative intention of the source text since "In his dream" is hallucinated, and the inaccurate "breast" instead of "nipple" distorts the picture. Concerning the translation strategy, it should be noted that in this case, the human translator and ChatGPT use taboo-to-taboo transfer while DeepSeek makes an interpretive mistranslation.

The phrase used here is an example of an epithet since it expresses disgust and animosity against a bunch of men through spitting on their pictures. The human translation captures the literal, intended, and derogatory meaning of the source text without any errors in terms of lexemes and grammar, just as ChatGPT does. Thus, there are no MQM errors in both translations. They both use a domestication strategy. However, DeepSeek creates the unneeded addition of "boasting" in its translation that is not present in the source text and distorts the context in the way that it attributes one of the features not mentioned in the source text. Such an error is considered to be a major error of accuracy since it alters the propositional meaning but keeps the message intact. As far as the translation strategy is concerned, human translation and ChatGPT use taboo-to-taboo transfer, keeping the insult as it is, while DeepSeek uses the combination of direct transfer and addition.

In this case, the term used in example 7 is classified as an obscenity since it involves explicit descriptions of physical contact. The translation into human language has captured the pragmatic meaning and style of the source text, especially the terms “old, gnarled hand,” which portray the image of يده المعروقة. On the other hand, ChatGPT and DeepSeek have maintained the explicitness in the translation, but each has made one major terminology mistake by translating the expression as “veined hand” and “calloused hand.” Although these two options do not change the meaning of the text, they have failed to capture the characterizations made in the source text. Concerning the translation method, all three have used the taboo-to-taboo transfer; however, ChatGPT and DeepSeek have employed lexical substitution, too

The above expression falls under the category of obscenity since it involves the use of both bodily violence and very obscene verbal insults. The human translation has captured both the physicality and the insulting nature of the expression by translating "يا مرة يا بنت الـ..." as "slut, bitch," thus capturing the offensive nature of the expression to the target reader. That is, the translator employs the domestication strategy. ChatGPT has maintained the literal sense of the expression but downgraded its insulting nature via literal translation and lexical compensation. DeepSeek has downgraded the insulting nature of the expression even more by using "You wretched woman" in place of the insult and "muttering" in place of "repeating" among other words that have made the translation inaccurate both lexically and pragmatically. Under MQM standards, these lexical and pragmatic changes explain why the two AI-based translations are inferior compared to the human translation. With regard to the translation technique, a human translator has used cultural adaptation (taboo-to-taboo), ChatGPT has depended on literal translation, while DeepSeek has used euphemization to make a pragmatic deviation from the source text and lose the feminist intensity of the source text.

4. Discussion and analysis

Figure 1 presents the overall MQM quality scores calculated from the complete corpus of 50 taboo expressions, revealing only minor quantitative differences among the three translation outputs. The human translation achieved the highest overall quality score (99.78%), followed by ChatGPT (98.97%) and DeepSeek (98.12%). Although the numerical differences are relatively small, the qualitative analysis demonstrates that the human translator more consistently preserved the pragmatic, cultural, and contextual meanings of taboo expressions, particularly in passages requiring culturally appropriate lexical choices and the preservation of El-Saadawi's feminist discourse. ChatGPT performed competitively, producing highly accurate lexical renderings in several examples, but occasionally introduced grammatical or semantic distortions that affected pragmatic equivalence. DeepSeek likewise generated fluent translations but showed a greater tendency toward lexical substitution and contextual inaccuracies, resulting in slightly lower MQM scores. Overall, the findings indicate that while contemporary AI systems can achieve translation quality approaching that of professional human translators, human expertise remains superior in preserving the cultural and pragmatic functions of taboo expressions in literary translation.

While ChatGPT and DeepSeek delivered highly accurate translations of the source texts into English, the qualitative study revealed the superiority of human translation in conveying pragmatic, cultural, and contextual features of taboo terms. The discrepancies between the three systems did not have much to do with the fluency of translation but had much to do with the choice of vocabulary, interpretation of context, and preservation of the power of culturally sensitive speech. ChatGPT often supplied highly accurate equivalents for words from the source texts and even outperformed human translation in some cases, but at the same time, it often distorted both grammar and semantics of the examples, affecting their pragmatic quality. DeepSeek usually supplied fluent translations but was prone to lexical substitutions, additions, and inaccurate contextual interpretation.

Further analysis of the translation strategies used shows several distinguishing features between the three translation systems. Firstly, the human translator always applied culturally appropriate equivalents where necessary without losing the pragmatic and ideological function of taboo words. Secondly, ChatGPT was based mainly on the use of direct transfer and managed to provide an accurate translation in most cases; however, there were some problems in conveying the interpersonal meaning of some sentences. Lastly, DeepSeek preferred using such strategies as lexical replacement and interpretation, which sometimes led to a loss of explicitness or the strength of the text.

Generally speaking, it has been found that modern-day AI technology is capable of achieving a level of quality of translation that is almost the same as that achieved by professionals. Nevertheless, some texts contain several culturally loaded and ideologically charged words and phrases, and they include the novel called Woman at Point Zero, which should be professionally translated since only professionals can translate it properly and interpret its content and pragmatics. Consequently, literary translation via AI technology is regarded as an additional option for professional translators.

5. Conclusions

The present research examined the efficiency of human translation, ChatGPT, and DeepSeek in translating taboo expressions from Arabic to English in Nawal El-Saadawi’s Woman at Point Zero by means of the MQM metrics system and Khoshsaligheh and Ameri’s (2014) classification of taboo translation strategies. The results indicate that, despite the high MQM scores of the translation tools under consideration, there were certain qualitative distinctions regarding their translation of culturally delicate expressions. The highest overall translation quality score was attained by the human translation (99.78%), proving its greater consistency in preserving the meaning of taboo expressions semantically, pragmatically, and culturally. Even though ChatGPT demonstrated comparable quality in performance (98.97%) and even sometimes offered more accurate lexemes than the human translation in question, it made some grammatical and semantic mistakes in several cases, which negatively influenced the pragmatic equivalence of the translation.

In addition, the qualitative analysis proved that human translator and ChatGPT generally employ a domestication strategy by making the text accessible and natural for the target readers. The three translations utilized different strategies, such as taboo-to-taboo translation, lexicon replacement, partial deletion, culture-specific transformation, euphemism, or mitigation strategy. The majority of translations maintain the implicature and the face-threatening acts in the target language. This is done as a result of the domestication strategy employed in the translation. This means that an effective translation of taboo terms cannot be simply based on the use of equivalent lexical items. Rather, the cultural and ideological roles played by the source text have to be considered. The human translation, in particular, was able to conserve better the discourse of El-Saadawi's narrative. Although there is some encouraging evidence regarding the success of modern AI programs, the results have indicated that AI must be seen as a helpful tool in translation rather than a substitute for literary translators while translating culturally and ideologically loaded texts. The involvement of humans is necessary to maintain contextuality, pragmatics, and cultural lexicons.

The research is limited to one literary piece, a human translation of it, and only two generative AI tools. As such, the results of the research cannot be generalised to all genres of literature and all AI algorithms. It would be possible in future research to study different literary texts, test other AI translation tools, analyze the translation of figurative language, and assess the efficiency of human post-editing of AI-generated translations of taboo language.