4.1 Introduction

Neural translations are not neutral. On the contrary, as a new dilemma for neural machine translation as neural machine translation systems have learned to recognize patterns in lexical and semantic units in human languages (Reference Johnson, Schuster and LeJohnson et al., 2017; Reference Popel, Tomkova and TomekPope et al., 2020; Reference GrechishnikovaGrechishnikova, 2021) to translate more fluently, increasing cultural bias in the target language has emerged. Given that language use is heavily influenced by the culture of the host country and carries with it deeply ingrained perceptions, beliefs, and attitudes (Reference DownesDownes, 1998; Reference FishmanFishman, 2019; Reference Thomas and WareingThomas and Wareing, 1999; Reference MontgomeryMontgomery, 1995), increasingly fluent translations can increasingly convey those cultural aspects, and sometimes bring cultural biases along with them. In this respect, machine biases induced by translation are inevitable consequences of algorithms designed to achieve near-native level linguistic naturalness and communicative fluency in automatic translation outputs (Reference Weng, Yu, Wei and LuoWeng et al., 2020; Reference Feng, Xie and GuFeng et al., 2020; Reference Martindale, Carpuat, Duh and McNameeMartindale et al., 2019; Reference KoehnKoehn, 2020;Wu et al., 2016). And in fact, University of Cambridge researchers have indeed discovered gender bias in machine translations of English into German, Spanish, and Hebrew, chosen for their distinct linguistic and cultural properties (Reference Saunders and ByrneSaunders and Byrne, 2020). Their studies revealed that, in MT output, gender bias in particular was an inevitable consequence of language use in training datasets that included genres such as news reports and speeches. Similarly, several studies of machine translation quality assessment revealed widespread racial, as well as gender, bias (Reference Tomalin, Byrne and ConcannonTomalin et al., 2021; Reference Font and Costa-jussàFont and Costa-jussà, 2019; Salles et al., 2018; Reference BestBest, 2017). In the machine translations of job titles in the U.S. Bureau of Labor Statistics, Reference Prates, Avelar and LambPrates et al. (2019) showed a strong tendency toward male defaults in as many as twelve languages (Hungarian, Chinese, Japanese, Basque, Yoruba, Turkish, Malay, Armenian, Swahili, Estonian, Bengali, Finnish)Footnote ¹.

While commendable progress has been made in developing scalable approaches to reducing such bias in machine translations, the problem persists. Significant human effort is still required to revise MT training data, and the resulting MT datasets still do not address all forms of social discrimination inherited from target-language datasets. In our current study, we will speak of this issue as inevitably arising from the social and cultural constraints of artificial intelligence. Since human thoughts and behaviors do have social and cultural contexts, sexist, racial, class, and other types of bias are inevitable in MT output; and artificial intelligence, as our brainchild, will inevitably amplify these tendencies. Nevertheless, they are predictable and preventable. We argue that MT quality assessment should incorporate social, ethical, and cultural sensitivity, rather than focusing solely on linguistic accuracy and fluency. And specifically, for materials generated by neural translation, it is necessary to develop mechanisms to support decision-making concerning the trade-offs involving linguistic fluency and cultural biases. Special attention is needed in specialized domains. One such, multicultural mental healthcare, provides the focus of our study.

Globally, anxiety disorders are the largest burden on mental health (3.76 percent in 2017). Countries with the highest prevalence of anxiety disorders (5 percent–6 percent) are some of the most advanced economies in their regions (Argentina, Brazil, Chile, Uruguay) and worldwide (U.S., Canada, UK, Germany, Australia, Sweden, Spain, France, Italy, Norway, New Zealand, Denmark, Ireland), as well as a few countries in the Middle East and Africa (Algeria, Iran). In Asia, only a few countries ranked within this range (5–6 percent), even though anxiety disorders are traditionally prevalent in countries like Japan, South Korea, and, more recently, India and China. Developing Latin American countries, too, have a tradition of anxiety. However, we can ask whether mental disorders are openly treated in different countries, and whether differences in openness might affect these statistics. The use of medical and mental healthcare can be subject to discrimination and stigma. And in fact, we do find that, in some rapidly developing countries, mental health issues are underrepresented, even though their populations are exposed to environmental, social, and economic stressors. This underrepresentation might be a result of traditional cultural beliefs stigmatizing people with mental illnesses, and perhaps from a related lack of access to mental healthcare support.

We would expect such negative social sentiments or attitudes. Thus, to test our hypothesis, we gathered and generated translations into Chinese, Hindi, and Spanish of public health materials on anxiety disorders developed by health-promotion organizations in English-speaking countries. The back-translations from these three languages were paired with their original English health materials, and the distribution of negative emotion words in each pair was statistically analyzed (Figure 4.1).

Figure 4.1

Ratios of machine translations of statistically increased negative emotions

4.2 Data Collection

We developed a set of quality control criteria based on five considerations for searching English public health materials on anxiety disorders. In gathering our training data, we screened online mental health information on this topic according to these criteria. Our intent has been to ensure the usefulness of the machine learning classifiers we have developed for health organizations and their wide applicability in research and clinical settings for effective, positive cross-lingual health communication concerning mental health disorders among multicultural populations.

• Topic Relevance: Our study focuses on anxiety disorders, due to their high prevalence

• Information accessibility: The materials we selected were written in an accessible, familiar style. Materials of this type are more likely to be translated by Google into language that the general public, as opposed to health professionals, can understand. As the linguistic difficulty of English material increases, training data will be restricted to professional-oriented health resources. These will be less suitable for learning about the common language in a given country and the social attitudes toward mental disorders conveyed in its language.

• Information credibility: The English health materials selected were developed by national or charitable health-promotion organizations to ensure credibility (see Appendix 4).

• Understandability: Online health information materials may be intended for professionals or for patients. We chose English materials intended for the public, since their translations can be better understood by machine translation users from diverse cultural and linguistic backgrounds and thus can impact their opinions on mental health conditions. This criterion was significant because an important objective of our research was to develop machine learning classifiers that would improve translation quality – that is, that would help to produce less biased machine translations that could contribute to more positive understanding of anxiety disorders. With this goal in mind, we developed classifiers to process English health materials in an accessible and understandable manner.

• User relevance: While understudied, relevance to specific users is another key indicator of mental health resource quality. The causes, symptoms, and treatment of mental disorders vary considerably among people of varying demographic characteristics – people of different ages, genders, socioeconomic classes, and so on. We assume that health information can be significantly improved by tailoring it to specific user groups, and accordingly collected online English materials concerning anxiety disorders developed for children, teens, young adults, the elderly, men, women, and transgender people.

As part of the quality control process, we identified websites of national, charitable health-promotion organizations and selected original English health materials that met the above criteria. There were 557 original English health materials. There are three sets of natural language features annotated on the original English and back-translation health materials: multiple semantic categories using the university of Lancaster Semantic Annotation System (USAS); word frequency bands (WFB); and lexical dispersion rates (LDR), with the last two based on the British National Corpus. The total number of annotation classes was 153 including semantic classes (115), WFB (18), and LDR (20) (see Appendix 2).

4.3 Development of Machine Learning Classifiers

We collected 557 original English health materials regarding anxiety disorders that met all our search criteria. We generated their machine translations into Chinese, Hindi, and Spanish using the Google Translate API. We then compared the original English with their matching back-translations from the three languages and used the Linguistic Inquiry and Word Count System (LIWC) (University of Texas at Austin) to find the distribution of words expressing negative emotions in both sets of English materials, original and back-translation. A Wilkson signed-rank test found back-translations from the target languages showed a statistically significant increase (p*0.05) in expressions of negative emotions when compared to their original English texts.

The original English texts were chosen since they yielded back-translations showing statistically increased sentiment negativity concerning anxiety disorders. Since our goal was to develop neural programs that could distinguish texts relatively likely to produce biased translations, we then manually developed training corpora as follows. We classified the original English texts into risky (1) versus safe (0) classes: risky English texts were associated with back-translations of increased negativity in at least one language of Chinese, Hindi, or Spanish; and safe texts were associated with back-translations in which negativity increase was statistically insignificant (p>=0.05) in all three test languages. Of the 557 texts collected to train and test machine learning classifiers, 428 texts were classified as safe (class 0) and 129 texts as risky (class 1).

Again, our goal was to distinguish texts that were safe, or unlikely to contain biased language, from those that were risky, or like to contain such language. We faced some analytical problems, however, in that (1) the languages we studied differed in their respective degrees of negativity and (2) our corpus contained many more safe than risky texts. The most negative translations were found in Chinese (13.26 percent), followed by Hindi (8.24 percent) and Spanish (9.68 percent); and within the three target languages, the ratio of English materials associated with increased negativity in machine translations and those without any negative machine translations was 3:10.

In other words, in statistical terms, our data was imbalanced – as would be the case, for example, if we attempted to distinguish legitimate credit card transactions from fraudulent ones, since the former will greatly outnumber the latter in any corpus. Fortunately, various techniques have been developed for handling such data imbalance. Synthetic Minority Oversampling Technique (SMOTE) (Reference Chawla, Bowyer, Hall and KegelmeyerChawla et al., 2002) in Python was applied to improve the balance between the two classes of machine translation output in terms of negative emotion words. We divided the whole dataset, after oversampling, into training (70 percent) and testing datasets, and performed five-fold cross-validation on the training dataset (see Table 4.1).

4.4 Feature Optimization

Figure 4.2

Recursive Feature Elimination with Automatic Feature Selection as the Base Estimator Cross-validation classification error (CVCE)

(a) automatic optimization of English lexical dispersion features (from 20 to 9 feature, CVCE= 0.333)

(b)

automatic optimization of English lexical frequency range features (from 18 to 7 features, CVCE= 0.333)

(c)

automatic optimization of English semantic features (from 115 to 66 features, CVCE= 0.260)

(d)

automatic optimization of all features (a, b, c) (from 153 to 119 features, CVCE=0.245)

4.5 Separate and Combined Feature Optimization

The dataset has 153 features, including 115 semantic classes, 18 WFB, and 20 LDR. By reducing high-dimensional feature sets, machine learning classifiers can be made more efficient and interpretable. Accordingly, support vector machine (SVM) methodology was used as the base estimator (RVM_SVM) in recursive feature elimination.

Figure 4.2 (a) shows the automatic optimization of English lexical dispersion rate features. After reduction of the LDR from 20 to 9, the cross-validation classification error reached its minimum (0.333). English lexis dispersion rates range from 0 to 1, with higher dispersion rates indicating wider distribution of the words across different textual genres, and thus indicating whether the relevant language is general or specialized. Both spoken and written dispersion rates were optimized: for very specialized words in spoken English (DiSp1:0.0–0.1, DiSp3:0.2–0.3); for general words in spoken English (DiSp6:0.5–0.6, DiSp9:0.8–0.9, DiSp10:0.9–1.0); and for medium-to-very-general words in written English (DiWr4:0.3–0.4, DiWr6:0.5–0.6, DiWr8:0.7–0.9, DiWr10:0.9).

In a comparable way, Figure 4.4 (b) shows the automatic optimization of English lexical frequency band features. When the number of features was reduced from 18 to 7, the minimal cross-validation classification error was obtained (0.333). In the British National Corpus (BNC), frequency bands refer to the ordinal ranges of word occurrence frequencies. For written materials, we listed nine frequency bands: FrWr1:0–500, FrWr2:500–1000, FrWr3:1000–1500, FrWr4:1500–2000, FrWr5:2000–2500, FrWr6:2500–3000, FrWr7:3000–3500, FrWr8:3500–4000, and FrWr9:4000–64420. The words that appear most frequently in the BNC corpus are those in the higher bands. For example, only 30 English words in the FrWr9 band occur more than 4,000 times in the entire database. Generally, the smaller the frequency bands, the less frequent or familiar the words are to the public. We also provided nine frequency bands for spoken materials: FrSp1:0–500, FrSp2:500–1000, FrSp3:1000–1500, FrSp4:1500–2000, FrSp5:2000–2500, FrSp6:2500–3000, FrSp7:3000–3500, FrSp8:3500–4000, and FrSp9:4000–57010. Again, higher frequencies indicate greater familiarity with words. We note that optimization of frequency band features reduced the original number of bands from 18 to 7: FrSp1:0–500, FrSp5:2000–2500, FrSp9:4000–57010, FrWr4:1500–2000, FrWr7:3000–3500, FrWr8:3500–4000, and FrWr9:4000–64420.

Finally, the automatic optimization of English semantic features is shown in Figure 4.4 (c). There were in total 115 semantic features covering as many as 21 semantic categories: general and abstract terms (A1-A15, 15 features); the body and the individual (B1-B5, 5 features); arts and crafts (C1); emotion (E1-E6, 6 features); food and farming (F1-F4, 4 features); government and public (G1-G3, 3 features); architecture, housing and the home (H1-H5, 5 features); money and commerce in industry (I1-I4, 4 features); entertainment, sports and games (K1-K6, 6 features); life and living things (L1-L3, 3 features); movement, location, travel and transport (M1-M8, 8 features); measurements (N1-N6, 6 features); substances, materials, objects and equipment (O1-O4, 4 features); education (P1), language communication (Q1-Q4, 4 features); social actions, states, processes (S1-S9, 9 features); time (T1-T4, 4 features); environment (W1-W5, 5 features); psychological actions, states and processes (X1-X9, 9 features); science and technology (Y1-Y2, 2 features); and names and grammar (Z0-Z9, Z99, 11 features).

The minimal classification error (0.260) was reached when the original semantic feature sets was reduced by almost half from 115 to 66: A12 (easy/difficult); A13 (degree, extent); A15 (safety/danger); A7 (probability); B1 (anatomy, physiology); B2 (health and disease); B3 (medicines, medical treatment); E2 (liking); E3 (calm/violent/angry); E4 (happiness, contentment); E5 (bravery, fear); E6 (worry, confidence); G2 (crime, law); I1 (money); I3 (employment); O4 (physical attributes); Q1 (linguistic actions, states, processes); S1 (social actions, states, processes); S2 (people); S8 (helping/hindering); S9 (religion); W1 (environment); W3 (geographical terms); X1 (psychological actions, states, processes); X3 (sensory); X4 (mental object); X5 (attention); X6 (Deciding); X7 (wanting, planning); X8 (trying); X9 (Ability); Z6 (negative); Z8 (pronouns); and so on. Figure 4.4 (d) shows the automatic optimization of the three sets of natural language features combined. The minimal classification error (0.245) was reached when the full feature set (153 features) was reduced to 119.

4.6 Classifier Training and Development

Relevance vector machine (RVM) methodology was used to develop Bayesian machine learning classifiers in Table 4.2. Different RVM models were compared using paired optimized and unoptimized feature sets, as well as their normalized versions, using three different techniques for feature normalization: min-maximal normalization (MMN), L2 normalization (L2 N), and Z-score normalization (ZSN). On the testing data, optimized feature sets of English LDR (Disp_9) achieved a higher area under the receiver operator characteristic (area under curve (AUC)=0.7023) than its matching unoptimized feature set (Disp_20) (AUC=0.7013). Feature normalization increased AUC of optimized and non-optimized feature sets to varying degrees. Optimization did not improve the performance of the feature set of WFB, since the AUC of Freq_7 on the testing data set (0.6626) was lower than Freq_18 (0.6784). By contrast, optimization did enhance the performance of RVMs using semantic features, as USAS_66 (0.7894) had a higher AUC than USAS_115 (0.773). With min-max normalization as the best technique, the AUC of the optimized model USAS_66 also increased.

The results of the separate optimizations of the English feature sets are shown in Table 4.2. Table 4.3 shows the results of combining the three feature sets. Although the optimized full feature set (F119) (AUC=0.778) did not achieve a higher AUC than the unoptimized full feature set (F153) (AUC=0.830), feature normalization significantly increased the AUC of classifier F119. The most effective normalization technique was min-max normalization, which increased the AUC of classifier F119 from 0.778 to 0.896, very similar to that of classifier F153 after the same normalization process (0.897).

Pairwise comparisons were conducted of any two optimized feature sets to determine the best combination of features. The results show that the combination of optimized dispersion rates (Disp_9) and optimized semantic features (USAS_66) yielded the highest AUC on the testing data: F75 (AUC=0.919, sensitivity=0.803, specificity=0.896, accuracy=0.848), followed by the combination of all three optimized feature sets: F82 (AUC=0.906, sensitivity=0.841, specificity=0.832, accuracy=0.837). F75 thus emerged as the best model.

4.7 Statistical Refinement of the Optimized Classifier

In order to further improve the performance of the optimized classifier F75, we performed statistical analyses of the dispersion rate features and semantic features in the two sets of English mental health materials: labeled as 0, indicating no back-translation associated with statistically increased negative emotions, and labeled as 1, indicating back-translations with strong negative connotations in one or more of the three languages – Chinese, Hindi, and Spanish (Table 4.4). Appendix 3 shows the results of the Mann Whitney U test between the two sets of original English texts. As compared to “safe” original English materials, five features yielded statistically different distributions with respect to their respective probabilities of being translated into Chinese, Hindi, or Spanish with strong negativity: DiSp9:0.8–0.9 (p<0.001), DiSp10:0.9–1.0 (p<0.001), DiWr6:0.5–0.6 (p=0.045), DiWr8:0.7–0.9 (p<0.001), and DiWr10:0.9–1.0 (p<0.001). The automatically selected features of LDR were reduced from 9 to 5. Similarly, the number of semantic features was reduced from 66 in the automatic feature selection (RVM_SVM) to 59. The classifier was subsequently fine-tuned by comparing four combinations of optimized LDR with optimized semantic features (USAS_59). (For details on the different dispersion rates used, see Appendix 2.) With ZSN, F (best 64) emerged as the best-performing classifier (AUC=0.893, accuracy=0.852, sensitivity=0.864, specificity=0.84).

RVM classifiers with different feature sets are compared in Tables 4.5 and 4.6. The comparison was to determine whether the sensitivity and specificity of the best-performing model were significantly higher than those of other classifiers. To control for any false discovery rate, we applied the Benjamini–Hochberg correction procedure.

With respect to sensitivity, the results show that F (best 64) yielded statistically higher sensitivity than the other seven high-performing classifiers selected from the 72 classifiers we developed. There was no statistically significant difference between F (best 64) and F119 (jointly optimized features and normalized using min- max optimization). However, F (best 64) was much less complex with only 64 features.

With respect to specificity, F (best 64) gave statistically greater specificity than five high-performing classifiers (F63, F153, F68, F119, and F82), while F (best 64) gave statistically similar specificity to classifiers F64 and F65. The specificity of F (best 64) was statistically lower than that of F75 (MMN) and F71 (ZSM), but the sensitivity of F (best 64) was statistically higher than F75 (MMN) and F71 (ZSM).

As the primary aim of our study is to detect English texts that are more likely to be translated with strong negative connotations in the target languages, model sensitivity is more important than specificity. Therefore, F (best 64) was chosen as the best-performing classifier.

4.8 Model Stability

On the testing data set, Figure 4.3 shows how AUC varies when the size of the training data was adjusted from 150 to 550 on 100 intervals. The RVMs show themselves unlikely to have overfitting issues, unlike other classifiers such as extreme gradient boosting trees, random forests, and neural networks that require hyperparameter tuning. The RVM classifiers all demonstrated stability and scalability, as their performance (AUC) increased gradually as we increased the training dataset size. F (best 64) outperformed other classifiers when the size of training data exceeded that of testing data. Figure 4.4 shows the mean AUC of RVM classifiers on test data and Table 4.7 shows the paired sample t test of the AUC of these classifiers. Even though F (best 64) employed the smallest number of features, its mean AUC was comparable to that of other high-dimensional classifiers.

Figure 4.3

AUCs of RVMs on testing data using different training dataset sizes

(150, 250, 350, 450, 550).

Figure 4.4

Mean AUC of RVMs on testing data using different training dataset size (150, 250, 350, 450, 550).

To review, then, we have succeeded in developing a high-performing relevance vector machine (RVM) classifier to predict the likelihood of a certain English health text being translated by Google as having statistically increased negative connotations when compared to the original English text.

Among the three languages we studied, the negative emotions and attitudes toward mental health – specifically, toward anxiety disorders – introduced by automatic translation were widespread. To understand the reasons, we carefully read the original English health texts in our database and their corresponding back-translations from the target languages. We found some illuminating examples, collected in Appendix 1. In some cases, mental health disorders have been translated by Google into Chinese as mental health diseases; people with anxiety disorders are describe in translation as mental illness patients; in Spanish, mental health conditions are translated as mental illnesses; in Hindi, shyness is translated as general shame, and panic is translated as nervousness. In Chinese, there is a subtle difference in word connotation: neural verbs such as “have a mental disorder” are replaced by “suffer from a mental disease.” As discussed above, neural machine translation emphasizes naturalness and fluency of translations, as compared with the more literal translations characteristic of statistical machine translation. The overall translation does improve significantly in terms of readability, fluency, and grammar; however, the accuracy of local translations may be compromised at the lexical and lexico-grammatical levels.

As Way noted in the following text:

As we have seen, a direct result of the target-language-oriented approach to neural MT in mental health translation has been unintentional increased negativity and discrimination in the translation output, even though such connotations were absent in the original English mental health materials. And as discussed, we have developed machine learning classifiers mitigate this undesirable effect by detecting English mental health information that might lead to biased translation in the three languages.

Our RVM classifier reached its statistically highest sensitivity (mean=0.864, 95 percent C.I.: 0.805, 0.922) and specificity (mean=0.832, 95 percent C.I.: 0.766, 0.898) when the probability threshold was set at 0.5. However, the default threshold of 0.5 can be adjusted further to fine-tune the classifier. When the threshold of the classifier was increased, sensitivity decreased, while specificity increased; conversely, when the threshold was decreased, sensitivity increased, and specificity decreased. Thus, one can select the best sensitivity and specificity pairing according to the practical circumstances.

For example, high-sensitivity classifiers can be useful for screening purposes – to identify mental health materials in English which cannot be adequately translated by full, unverified automatic machine translation, due to their heightened likelihood of biased or discriminatory translations. By contrast, a low-sensitivity classifier is relatively unlikely to identify potentially problematic original English mental health information – that is, to screen out any materials that are not safe and suitable for neural machine translation. Thus, any social biases or discrimination against mental disorders still present in the target language would unfortunately be reinforced in machine-translated mental health resources, even if indirectly or unintentionally – clearly not the intent of global mental health promotion.

As for classifiers with high specificity, they are more suitable for identifying original English mental health texts which are suitable for neural machine translation, at least for the three languages we studied, Chinese, Hindi, and Spanish. On the other hand, when a classifier with a low specificity is used, there is a raised likelihood of false positive predictions: that is, even safe and suitable original English mental health information may be erroneously considered unsuitable for neural MT. While subsequent human post-evaluation could correct such inaccurate predictions, logistical burdens and staff costs would increase. This extra effort might well be impractical or prohibitive in low-resource healthcare service scenarios, often subject to tight budget constraints or lacking bilingual workers with sufficient knowledge of the relevant languages.

4.9 Conclusion

In comparison with earlier statistical machine learning models, current neural machine translation technologies exhibit greater linguistic fluency. However, as noted, while improving linguistic fluency, neural machine translation also learns, inevitably if unconsciously, to reflect the sentiments, attitudes, and biases of the target cultures, societies, and communities. Since the design favors the most natural sequence of translated words and phrases in the target language – its natural lexical and syntactic patterns – its results inevitably convey the social and cultural connotations of the cultures from which the languages sprang.

In many countries and cultures, mental disorders are still stigmatized and subject to discrimination. In associated languages, this deeply rooted sentiment comes to be reflected in conventional lexical and semantic units. Consequently, the neural machine translation of mental health information too often entails the transmission of negative social sentiments in the output, even when social prejudice against mental disorders is actually absent from the original English materials. Our study has examined this understudied tendency in neural machine translation. We argue that this examination is appropriate and necessary as human communication technologies move rapidly toward more human-centric AI. Accordingly, we have developed Bayesian machine learning classifiers to assist with the probabilistic detection and prediction of socially biased neural machine translation outputs, using computational modeling and pairwise comparisons of original English and back-translations of neural machine translation outputs on anxiety disorders in Chinese, Hindi, and Spanish.

Via Google’s Translate API, we collected and compared original English documents on anxiety disorders from U.S., UK, Canadian, and Australian health authorities and their back-translations from Chinese, Hindi, and Spanish. Through automatic, statistically informed feature optimization, RVM classifiers were developed. These models provided informative probabilistic predictions of the likelihood that an English text on anxiety disorders would be translated by Google into one of the three languages with subtle but strong negative connotations – again, because neural machine translation favors natural language translations.

The best-performing RVM (RVM_ best 64) contained 64 English linguistic features: 59 (semantic features) and 5 (LDR: DiSp9:0.8–0.9, DiSp10:0.9–1.0, DiWr8:0.7–0.9, DiWr9:0.8–0.9, DiWr10:0.9–1.0). This result suggests that, in spoken and written English words belonging to certain semantic classes, words with high dispersion of meaning are relatively likely to produce negative neural machine translation results for anxiety disorders. The best-performing RVM (both optimized and normalized) achieved a mean AUC of receiver operator characteristic (0.893), accuracy (0.892), sensitivity (0.864), and specificity (0.84). Its sensitivity (SE) and specificity (SP) were statistically higher than those of unoptimized, normalized classifiers RVM_153 (min-max normalized MMN) (SE: p=0.0098, SP: p=0.007); of automatically optimized and normalized classifiers: RVM_82 (MMN) (SE: p=0.0095, SP: p=0.0083), RVM_75 (SE: p=0.0078, SP: p=0.0121); and of an automatically optimized and statistically refined classifier: RVM_71 (MMN) (SE: p=0.0103, SP; p=0.0095); RVM_68 (MMN) (SE: p=0.0095, SP: p=0.0080); RMV_63 (ZSM) (SE: p=0.0095, SP: p=0.0077). The stability of RVM_ best 64 appeared in its mean AUC (0.82, SD=0.05) when the training data sizes were reduced from 600 to 150.

Content analysis indicates that negative neural machine translations of anxiety disorders were primarily associated with increased linguistic fluency and communicative naturalness in the target Chinese, Hindi, and Spanish texts. However, once again, these stylistically enhanced phrasal patterns reflect persistent social attitudes toward mental health disorders in the relevant languages and cultures. And again, this bias is to be expected: as in any form of artificial intelligence, neural machine translation is designed to accommodate and understand human wants, needs, and thinking patterns.

This study demonstrates that phrasal patterning in target cultures does indeed yield increased negativity toward mental disorders as a consequence of greater translation naturalness. In compensation, however, we demonstrate that machine learning tools for the promotion of mental health translation can indeed detect instances of the automatic generation and dissemination of negative, discriminatory translation. In this way, neural tools can also promote positive and supportive social understanding and acceptance of mental disorders. This study confirms that, while neural machine translation technology is inevitably and increasingly culturally skewed, it can nevertheless be harnessed to foster more tolerant global health cultures.