Mapping the Multilingual Margins: Intersectional Biases of Sentiment Analysis Systems in English, Spanish, and Arabic

Introduction

Large-scale transformer-based language models, such as BERT Devlin et al. (2018), are now the state-of-the-art for a myriad of tasks in natural language processing. However, these models are well-documented to perpetuate harmful social biases, specifically by regurgitating the social biases present in their training data which are scraped from the Internet without careful consideration Bender et al. (2021). While steps have been taken to “debias”, or remove, gender and other social biases from word embeddings Bolukbasi et al. (2016); Manzini et al. (2019), these methods have been demonstrated to be cosmetic Gonen and Goldberg (2019). Furthermore, these studies neglect to recognize both the impact of social biases on downstream task results as well as the complex and interconnected nature of social biases. In this paper, we detect and discuss unisectionalIn this paper, we refer to biases against a single social cleavage, such as racial bias or gender bias, as unisectional. and intersectional social biases in multilingual language models applied to downstream tasks using a novel statistical framework and novel multilingual datasets. Intersectionality is a framework introduced by Crenshaw (1990) to study how the composite identity of an individual across different social cleavages (e.g., race and gender) informs that individual’s social advantages and disadvantages. For example, individuals who identify with multiple disadvantaged social cleavages (e.g., Black women) face a greater and altered risk for discrimination and oppression than individuals with a subset of those identities (e.g., white women). This framework for understanding overlapping systems of discrimination has been explored in some studies of fairness in machine learning, including by Buolamwini and Gebru (2018) who show that face detection systems perform markedly worse for female users of color, compared to female users or users of color. Although work has begun to study intersectional social biases in natural language processing, to the best of our knowledge no work has explored fairness in an intersectional framework on downstream tasks (e.g. sentiment analysis). Social biases in downstream tasks expose users with multiple disadvantaged sensitive attributes to unknown but potentially harmful outcomes, especially when models trained on downstream tasks are used in real-world decision making, such as for screening résumes or predicting recidivism in criminal proceedings Bolukbasi et al. (2016); Angwin et al. (1999). In this work, we choose emotion regression as a downstream task because social biases are often realized through emotion recognition Elfenbein and Ambady (2002) and machine learning models have been shown to reflect gender bias in emotion recognition tasks Domnich and Anbarjafari (2021). For example, sentiment analysis and emotion regression may be used by companies to measure product engagement for different social groups. In addition, while some work has studied gender biases across different languages Zhou et al. (2019); Zhao et al. (2020), no work to our knowledge has studied racial, ethnic, and intersectional social biases across different languages. This lack of a multilingual analysis neglects non-English speaking users and their complex social environments. In this paper, we demonstrate the presence of gender, racial, ethnic, and intersectional social biases on five language models trained on an emotion regression task in English, Spanish, and Arabic. We do so by introducing novel supplementary test sets designed to measure social biases and a novel statistical framework for detecting the presence of unisectional and intersectional social biases in models trained on sentiment analysis tasks. Our contributions are summarized as:

Following Kiritchenko and Mohammad (2018), we introduce four supplementary test sets designed to detect social biases in language systems trained on sentiment analysis tasks in English, Spanish, and Arabic, which we make available for download.

We propose a novel statistical framework to detect unisectional and intersectional social biases in language models trained on sentiment analysis tasks.

We detect and analyze numerous gender, racial, ethnic, and intersectional social biases present in five language models trained on emotion regression tasks in English, Spanish, and Arabic.

Related Works

The presence and impact of harmful social biases in machine learning and natural language processing systems is pervasive and well-documented in popular word embedding methods Caliskan et al. (2017); Garg et al. (2018); Bolukbasi et al. (2016); Zhao et al. (2019) due to large amounts of human-produced training data that includes historical social biases. Notably, Caliskan et al. (2017) demonstrate such biases by introducing the Word Embedding Association Test (WEAT) which measures how similar socially sensitive sets of words (e.g., racial or gendered names) are to attributive sets of words (e.g., pleasant or unpleasant words) in the semantic space encoded by word embeddings. While Bolukbasi et al. (2016); Manzini et al. (2019) introduce methods for “debiasing” word embeddings in order to create more equitable semantic representations for usage in downstream tasks, Gonen and Goldberg (2019) argue that such methods are merely cosmetic since social biases are still evident in the semantic space after the application of such methods. Moreover, these “debiasing” techniques focus on a particular social cleavage such as gender or race (i.e., unisectional cleavages). In contrast, our work considers both unisectional and intersectional social biases. Recent studies have also begun to focus on social biases in transformer-based language models Kurita et al. (2019); Bender et al. (2021). In particular, Bender et al. (2021) discusses how increasingly large transformer-based language model in practice regurgitate their training data, resulting in such models perpetuating social biases and harming users. Therefore, in this work we consider both static word embedding techniques and transformer-based language models. Crenshaw (1990) introduces intersectionality as an analytical framework to study the complex character of the privilege and marginalization faced by an individual with a variety of identities across a set of social cleavages such as race and gender. A canonical usage of intersectionality is in service of studying the simultaneous racial and gender discrimination faced by Black women, which cannot be understood in its totality using racial or gendered frameworks independently; for one example, we point to the angry Black woman stereotype Collins (2004). As such, we argue that existing studies in fairness are limited in their ability both to uncover bias in and to “debias” language models without engaging with the intersectionality framework. Intersectional social biases have been documented in natural language processing models. Herbelot et al. (2012) first studied intersectional social bias by employing distributional semantics on a Wikipedia dataset while Tan and Celis (2019) studied intersectional social bias in contextualized word embeddings by using the WEAT on language referring to white men and Black women. Guo and Caliskan (2021) introduce tests that detect both known and emerging intersectional social biases in static word embeddings and extend the WEAT to contextualized word embeddings. Similarly, May et al. (2019) also extend the WEAT to a contextualized word embedding framework using sentence embeddings. However, these methods do not consider the effect of intersectional social biases on the results of downstream tasks, which is the focus of this work. Studies on non-English social biases in natural language processing are limited, with Zhou et al. (2019) extending the WEAT to study gender bias in Spanish and French and Zhao et al. (2020) examining gender bias in English, Spanish, German, and French on fastText embeddings Bojanowski et al. (2017). Notably, to the best of our knowledge there has been no work on studying intersectional social biases in languages other than English in natural language processing. While Herbelot et al. (2012) and Guo and Caliskan (2021) study the intersectional social biases faced by Asian and Mexican women respectively using natural language processing, both do so in English. In contrast, our work seeks to understand intersectional social biases in the languages that are used by the individuals and the communities that they help constitute. Most closely related to our work, Kiritchenko and Mohammad (2018) evaluate racial and gender bias in 219 sentiment analysis systems trained on datasets from and submitted to SemEval-2018 Task 1: Affect in Tweets Mohammad et al. (2018). Their work introduces the Equity Evaluation Corpus (EEC), a supplementary test set of 8,640 English sentences designed to extract gender and racial biases in sentiment analysis systems. Despite Spanish and Arabic data and submissions for the task, Kiritchenko and Mohammad (2018) did not explore biases in either language. Moreover, this study focused on submissions to the competition. In contrast, our work focuses on large-scale transformer-based language models and explores both unisectional and intersectional social biases in multiple languages.

Methods: Framework for Evaluating Intersectionality

In this section, we introduce our framework for detecting unisectional and intersectional social bias on results from downstream tasks. Given a model trained on emotion regression, we evaluate the model on a supplementary test set using our framework to measure social biases. First, we discuss our supplementary test sets composed of sentences corresponding to social cleavages (e.g., Black women, Black men, white women, and white men) (§3.1). We then use the results from each test set to run a Beta regression model Ferrari and Cribari-Neto (2004) where we fit coefficients for gender, racial, and intersectional social biases (§3.2). Finally, we test the coefficients for statistical significance to determine if a model, trained on a given emotion regression task in a given language, demonstrates gender, racial, or intersectional social bias (§3.3).

We introduce four novel Equity Evaluation Corpora (EECs) following the work of Kiritchenko and Mohammad (2018). An EEC is a set of carefully crafted simple sentences that differ only in their reference to different social cleavages as seen in Table 1. Therefore, differences in the predictions on a downstream task between sentences can be ascribed to language models learning those social biases. We use these corpora as supplementary test sets to measure unisectional and intersectional social biases of models trained on downstream tasks in English, Spanish, and Arabic.

Following Kiritchenko and Mohammad (2018), each EEC consists of eleven template sentences as shown in Table 1. Each template includes a [person] tag which is instantiated using both given names representing gender-racial/ethnic cleavages (e.g. given names common for Black women, Black men, white women, and white men in the original EEC) Caliskan et al. (2017); Kiritchenko and Mohammad (2018) refer to the racial groups as African-American and European-American. For consistency and in accordance with style guides for the Associated Press and the New York Times, we refer to the groups as Black and white with intentional casing. and noun phrases representing gender cleavages (e.g. she/her, he/him, my mother, my brother). The first seven templates also include an emotion word, the first four of which are [emotion state word] tags, instantiated with words like angry and the last three are [emotion situation word] tags, instantiated with words like annoying. We contribute novel English, Spanish, and Arabic-language EECs that use the same sentence templates, noun phrases, and emotion words, but substitute Black and white names for Latino and Anglo names as well as Arab and Anglo names respectively. We introduce an English EEC and a Spanish EEC for Latino and Anglo names as well as an English EEC and an Arabic EEC for Arab and Anglo names, for a total of four novel EECs. The complete translated sentence templates, noun phrases, emotion words, and given names are available in the appendix and we make all four of our novel EECs available for download. The original EEC uses ten names for each gender-racial cleavage, selected from the list of names used in Caliskan et al. (2017), which in turn uses names from the first Implicit Association Test (IAT), a psychology study that measured implicit racial bias Greenwald et al. (1998). For example, given names include Ebony for Black women, Alonzo for Black men, Amanda for white women, and Adam for white men. The original EEC also uses five emotional state words and five emotional situation words sourced from Roget’s Thesaurus for each of the emotions studied. For example, furious and irritating for Anger, ecstatic and amazing for Joy, anxious and horrible for Fear, and miserable and gloomy for Sadness. Each of the sentence templates was instantiated with chosen examples to generate 8640 sentences. For names representing Latino women, Latino men, Anglo women, and Anglo men in the English and Spanish-language EECs we used the ten most popular given names for babies born in the United States during the 1990s according to the Social Security Administrationhttps://www.ssa.gov/oact/babynames/decades/names1990s.html. For the English and Arabic-language EECs, ten names are selected from Caliskan et al. (2017) for Anglo names of both genders. For male Arab names, ten names are selected from a study that employs the IAT to study attitudes towards Arab-Muslims Park et al. (2007). Since female Arab names were not available using this source, we use the top ten names for baby girls born in the Arab world according to the Arabic-language site BabyCenterhttps://arabia.babycenter.com/. All names are available in the appendix. For the Spanish and Arabic EECs, fluent native-speaker volunteers translated the original sentence templates, noun phrases, and emotion words. They then verified the generated sentences (i.e., using selected names and emotion words) for proper grammar and semantic meaning. Note that for the Arabic EEC, the authors transliterated names using English and Arabic Wikipedia pages of individuals with a given name. Due to fewer translated emotion words (e.g., two different English emotion words corresponded to the same word in the target language), each of the sentence templates were instantiated with chosen examples to generate 8640 sentences in English for both novel EECs, 8460 in Spanish, and 8040 in Arabic.

2 Regression on Intersectional Variables

In our model, we define $X_{1}$ to be an indicator function over sentences representing a minority group (e.g., Black people, women). For example, $X_{1i}=1$ for any sentence $i$ that refers to a Black person. As such, the corresponding coefficient $\beta_{1}$ describes the change in model prediction for sentences referring to an individual who identifies with that minority group, all else equal. For example, $\beta_{1}$ provides a measure of racial bias in the model. We define $X_{2}$ analogously for a second minority group. Therefore, the variable $X_{1}X_{2}=1$ if and only if a sentence refers to the intersectional identity (e.g., Black women) and thus $\beta_{3}$ is a measure of intersectional social bias.

3 Statistical Testing

After fitting the regression model, we test each regression coefficient for statistical significance. That is, we divide the coefficient by the standard error and then calculate the $p$-value for a two-sided $t$-test. If the coefficient for an independent variable (e.g., $X_{1}$) is statistically significant, we say that the model shows statistically significant social bias against the race and ethnicity, gender, or intersectionality identity corresponding to that variable. A positive coefficient for a variable implies that the emotion is exhibited more strongly by sentences representing the minority group that is coded by that variable.

Experiments

We experiment with five methods in this work. Our first three methods use pre-trained language models from Huggingface Wolf et al. (2019): BERT+ – for English we use BERT-base Devlin et al. (2018), for Spanish BETO Cañete et al. (2020), and for Arabic ArabicBERT Safaya et al. (2020), mBERT – multilingual BERT-base Devlin et al. (2018), XLM-RoBERTa – XLM-RoBERTa-base Conneau et al. (2019). For each language model, we fit a two-layer feed-forward neural network on the [CLS] (or equivalent) token embedding from the last layer of the model implemented in PyTorch Paszke et al. (2019), We do not fine-tune these models because we are interested in measuring the bias specifically encoded in the pre-trained publicly available model. Moreover, since the training datasets we use are small, fine-tuning has a high risk of causing overfitting. In addition, we also experiment with two methods using Scikit-learn Pedregosa et al. (2011): SVM-tfidf – an SVM trained on Tf-idf sentence representations, and fastText – fastText pre-trained multilingual word embeddings Bojanowski et al. (2017) average-pooled over the sentence and then passed to an MLP regressor.

2 Tasks

We first train models on the emotion intensity regression tasks in English, Spanish, and Arabic from SemEval-2018 Task 1: Affect in Tweets (Sem2018-T1) Mohammad et al. (2018). Emotion intensity regression is defined as the intensity of a given emotion expressed by the author of a tweet and takes values in the range $$. We consider the following set of emotions: anger, fear, joy, and sadness. For each model and language combination, we report the performance using the official competition metric, Pearson Correlation Coefficient ($\rho$) as defined in Benesty et al. (2009), for each emotion in the emotion regression task.

Results and Discussion

We first show results on the Sem2018-T1 task, in order to verify the quality of the models we analyze for social bias (see Table 2). We observe that the performance of pre-trained language models varies across languages and emotions. BERT+, mBERT, and RoBERTa performed best on the English tasks, compared to Spanish and Arabic. Additionally, BERT+ had better performance than the multilingual models (e.g. mBERT and XLM-RoBERTa) across all languages and tasks, showing that language-specific models (e.g., BETO) can be superior to multilingual models. SVM-tfidf and fastText typically outperformed the multilingual models but were at-par or only slightly better than the language-specific models. This difference is likely due to the lack of fine-tuning performed on the transformer-based models. Our decision to not fine-tune does decrease performance on downstream tasks but is prudent given the risk of overfitting on a small training set and our interest in studying the social biases encoded in off-the-shelf pre-trained language models.

2 Evaluation using EECs

After training a model for a given emotion regression task in a language, we utilize the five EECs as supplementary test sets. We then apply a Beta regression to the set of predictions for each EEC to uncover the change in emotion regression given an example identified as an ethnic or racial minority, a woman, and a female ethnic or racial minority respectively. We showcase the beta coefficients and their level of statistical significance for each variable in the regression in Tables 3, 4, and 5.

3 Discussion

In this section, we discuss the unisectional and intersectional social biases that we do and do not detect, across our five models that we trained on emotion regression tasks and evaluated using the EECs and novel statistical framework. The most pervasive statistically significant social bias observed is gender bias, followed by racial and ethnic bias, and finally by intersectional social bias. Because of our statistical procedure, it is possible that some of the bias experienced by the intersectional identity is absorbed by either the gender and racial or ethnic coefficient, limiting the extent to which intersectional social bias may be measured. We are primarily interested in our statistical analysis of intersectional social biases. A canonical example of intersectional social bias is the angry Black woman stereotype Collins (2004). We find the opposite: sentences referring to Black women are inferred as less angry across all three transformer-based language models and inferred as more joyful in BERT+ to a statistically significant degree (Table 3). It is possible that this bias is captured by other coefficients. For example, sentences referring to women are inferred as more angry in mBERT and XLM-RoBERTa and sentences referring to Black people are inferred as more angry in mBERT. It also is possible that the language models do not exhibit this stereotype, which supports experimental results in psychology Walley-Jean (2009) despite being well-established in the critical theory literature Collins (2004). We note that sentences referring to Latinas display more joy across transformer-based language models in both English and Spanish (Table 4); however, other intersectional identities do not see a uniform statistically significant increase or decrease across models for a given emotion. We find evidence of racial biases in our experiments. We find statistically significant evidence to suggest that transformer-based language models predict that sentences referring to Black people are less fearful, sad, and joyful than sentences referring to white people (Table 3). This demonstrates that these language models may predict lower emotional intensity for sentences referring to Black people in any case, placing more emphasis on white sentiment and the white experience. We observe that ethnic biases are sometimes split by language. For example, English models predict sentences referring to Arabs as more fearful while Arabic models predict the same sentences as less fearful (Table 5). However, both languages predict those sentences as more sad. Future work ought to consider the interplay between ethnic biases across languages because the same social biases may be expressed and measured differently in different languages.We observe multiple gender biases across emotions and languages. In all Arabic models, sentences referring to women are predicted to be less angry than sentences referring to men (Table 5). Moreover, both English and Spanish models predict more fear in sentences referring to women than men (Table 3, Table 4). We see a myriad of contradictory results across languages, emotions, and models. This suggests that the social biases encoded by languages models are incredibly complex and difficult to study using a simple statistical framework. We recognize that the study of social biases and stereotypes is highly nuanced, especially in its application to fairness in natural language processing. Future analysis of these language models, their training data, and any downstream task data is necessary for the detection and comprehension of the impact of social biases in natural language processing. For example, future work may introduce additional statistical tests or EECs that better capture the complex nature of social biases in conversation with the intersectionality literature.

Ethical Considerations and Limitations

Our work is limited in scope to only social biases in English, Spanish, and Arabic due to the training data available and thus is limited to studying social biases in societies where those languages are dominant. In addition, our statistical framework formalizes intersectional social bias across strictly defined gender-racial cleavages. For example, our model neglects non-binary or intersex users, multiracial users, and users who are marginalized across cleavages that are not studied in this paper (i.e. users with disabilities). Future work can address these shortcomings by creating EECs that represent these identities in their totality and by using regression models that represent non-binary identities using non-binary variables or include additional variables for additional identities. Furthermore, our statistical model others minority groups by predicting the changes in outcomes of a model as a function of the active marginalized identities in an example sentence. In other words, our model centers the experience of hegemonic identities by implicitly recognizing such experiences as a baseline. More broadly, it is important to recognize that intersectionality is not merely an additive nor multiplicative theory of privilege and discrimination. Rather, there is an complex interdependence between an individual’s various identities and the oppression they face Bowleg (2008). Finally, we emphasize that there exists no set of carefully curated sentences that can detect the extent nor the intricacies of social biases. We therefore caution that no work, especially automated work, is sufficient in understanding or mitigating the full scope of social biases in machine learning and natural language processing models. This is especially true for intersectional social biases, where marginalization and discrimination takes places within and across gender, sexual, racial, ethnic, religious, and other cleavages in concert.

Conclusion

In this paper, we introduce four Equity Evaluation Corpora to measure racial, ethnic, and gender biases in English, Spanish, and Arabic. We also contribute a novel statistical framework for studying unisectional and intersectional social biases in sentiment analysis systems. We apply our method to five models trained on emotion regression tasks in English, Spanish, and Arabic, uncovering statistically significant unisectional and intersectional social biases. Despite our findings, we are constrained in our ability to analyze our results with the sociopolitical and historical context necessary to understand their true causes and implications. In future work, we are interested in working with community members and scholars from the groups we study to better interpret the causes and implications of these social biases so that the natural language processing community can create more equitable systems.

Acknowledgements

We are grateful to Max Helman for his helpful comments and conversations. Alejandra Quintana Arocho, Catherine Rose Chrin, Maria Chrin, Rafael Diloné, Peter Gado, Astrid Liden, Bettina Oberto, Hasanian Rahi, Russel Rahi, Raya Tarawneh, and two anonymous volunteers provided outstanding translation work. This work is supported in part by the National Science Foundation Graduate Research Fellowship under Grant No. DGE-1644869. Any opinion, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation.

References

Appendix A Appendix

The names used in the original English EEC can be found in Table 6. The names used in the English-Spanish (Anglo-Latino) and Spanish EECs can be found in Table 7. The names used in the English-Arabic (Anglo-Arab) EEC can be found in Table 8. The names in the Arabic EEC (in Arabic text) can be found in Table 9.

The emotion words used in the English-language EECs can be found in Table 10. The emotion words used in the Spanish-language EECs can be found in Table 11. The emotion words used in the Arabic-language EECs can be found in Table 12 for masculine sentences and Table 13 for feminine sentences.

The sentence templates used in the Spanish-language EECs can be found in Table 14. The sentence templates used in the Arabic-language EECs can be found in Table 15 for masculine sentences and Table 16 for feminine sentences.

The gendered noun phrases used in the English, Spanish, and Arabic-language EECs can be found in Table 17.

A.2 Instructions to Original Translators

Translators were recruited at universities and are all university students. All translators are at least 18 and are fluent native speakers of the languages for which they translated. Each translator received an ID number to anonymize their work. Dear translator, Thank you for your help with our project. Your contribution is helping us conduct one of the first multilingual and intersectional bias analysis studies for natural language processing, a subset of artificial intelligence and linguistics. Natural language processing is responsible for tasks such as auto-completion, spell-check, spam detection, and searches on sites like Google. You and your work will be acknowledged in our final report. In the following document are the instructions for translations. First, answer the survey questions.

For each sentence, translate the template or individual word. We provide space for the female singular, female plural, male singular and female plural. If your language does not have separate masculine and feminine forms for any of the sentences, please include the singular and plural version in the first two boxes and if your does not have separate singular and plural forms, please include the singular versions for each gendered form as appropriate. If your language has additional cases, such as neutral, please make another column and note it for us (e.g. neuter in German). For the last ten, only give translations for the sentences as they are written. For the sentences with templates, Rearrange order of templates if necessary, but signify where [p] and [eA], [eB] tags belong in each template. For example, the [p] tag denotes person, e.g. she/her, this woman, my sister; the [eA] tag denotes emotional state words, e.g. angry, happy; and the [eB] tag denotes emotional event words, e.g. annoying, funny. For the emotion vocabulary, there are four categories: anger (red), fear (green), joy (yellow) and sadness (blue). If the English words do not correspond well, feel free to write the most approximate set of words for your language in any order. Let us know if there are intricacies in spelling due to, for example, consonants and vowels (e.g. a/an in English or le l’ in French). OPTIONAL: We are also looking for popular names of large socially cleaved groups in countries where your language is spoken. For example, in English, this includes male, female, Black and white names (5 for each combination of race and gender). If you are familiar with social cleavages or popular names in those cleavages in countries where your language is spoken, please note it. Sentence Templates:

found himself/herself in a/an [eB] situation

told us all about the recent [eB] events

goes to the school in our neighborhood

Words: angry, annoyed, enraged, furious, irritated, annoying, displeasing, irritating, outrageous, vexing, anxious, discouraged,fearful, scared, terrified, dreadful, horrible, shocking, terrifying, threatening, ecstatic, excited, glad, happy, relieved, amazing, funny, great, hilarious, wonderful, depressed, devastated, disappointed, miserable, sad, depressing, gloomy, grim, heartbreaking, serious, she/her, this woman, this girl, my sister, my daughter, my wife, my girlfriend, my mother, my aunt, my mom, he/him, this man, this boy, my brother, my son, my husband, my boyfriend, my father, my uncle, my dad Sentences:

He found himself in a displeasing situation

My boyfriend told us all about the recent dreadful events

This man goes to the school in our neighborhood

This woman found herself in a vexing situation

She told us all about the recent wonderful events

The conversation with my uncle was gloomy

A.3 Instructions to Checking Translators

Dear translator, Thank you for your help with our project. Your contribution is helping us conduct one of the first multilingual and intersectional bias analysis studies for natural language processing, a subset of artificial intelligence and linguistics. Natural language processing is responsible for tasks such as auto-completion, spell-check, spam detection, and searches on sites like Google. You and your work will be acknowledged in our final report. In the following document are the instructions for translations. First, answer the survey questions. Second, go through the sentences provided. For each sentence, indicate if the sentence is grammatically and semantically incorrect in the D column. You do not need to mark the cell if the sentence is correct. If it is incorrect, write the correct translation. If multiple consecutive sentences are incorrect in the same fashion: indicate the correct translation for the first sentence, note the error, and note the ID numbers for the sentences that are incorrect in that fashion. Ignore the lines that are blacked out. Here are some points to keep in mind: 1. Is the sentence grammatically correct? For example: does the sentence use the correct gendered language? Is the tense correct? 2. Is the meaning of the sentence the same as the English sentence listed next to it? It is okay if it is not the exact same as how you would translate it as long as the emotional word is similar. Informed Consent Form Benefits: Although it may not directly benefit you, this study may benefit society by improving our understanding of intersectional biases in natural language processing models across different languages. Risks: There are no known risks from participation. The broader work deals with sensitive topics in race and gender studies. Voluntary participation: You may stop participating at any time without penalty by not submitting the translations. We may end your participation or not use your work if you do not have adequate knowledge of the language. Confidentiality: No identifying information will be kept about you except for the translations you submit to us. No information will be shared about your work except an acknowledgement in the paper. Questions/concerns: You may e-mail questions to ac4443@columbia.edu. Submitting translations to António Câmara at ac4443@columbia.edu indicates that you understand the information in this consent form. You have not waived any legal rights you otherwise would have as a participant in a research study. I have read the above purpose of the study, and understand my role in participating in the research. I volunteer to take part in this research. I have had a chance to ask questions. If I have questions later, about the research, I can ask the investigator listed above. I understand that I may refuse to participate or withdraw from participation at any time. The investigator may withdraw me at his/her professional discretion. I certify that I am 18 years of age or older and freely give my consent to participate in this study.