Exposure to sexist content has serious consequences for women's life and limits their freedom of speech. In this paper, we present a multilingual system based on pre-trained transformers and compare singletask to multi-task learning to identify sexism in social networks. Our methods have been evaluated in the framework of our participation in the EXIST shared task at IberLEF 2021 [1] obtaining promising results despite sharing parameters for both languages and tasks.
The development of web technologies has enabled the interaction between people
from many di erent countries and backgrounds. With more than 4 billion people
around the world now using social media each month [
The Oxford English Dictionary de nes sexism as \prejudice, stereotyping or
discrimination, typically against women, on the basis of sex". Therefore, sexism
is expressed in very di erent forms that do not always express hostility or hate.
Subtle forms of sexism can be as pernicious as other forms of sexism and a ect
women in many facets of their lives. According to [
Detecting sexist content is still a di cult task for social media platforms. For
instance, Amnesty International published a report [
In this paper, we describe our participation in the EXIST task at IberLEF 2021, a sexist language detection task in two di erent languages. The challenge was articulated in two di erent tasks: task 1 is a binary classi cation to determine whether a text is sexist or not, while task 2 is a ner-grained classi cation devoted to distinguishing di erent subtypes of sexism. We propose a multilingual system based on pre-trained transformers and experiment with single-task and multi-task approaches to jointly address the task of sexist language detection. We take advantage of the fact that both tasks are semantically connected to test whether both tasks can be simultaneously learned and one task can bene t the other using a multi-task framework. To the best of our knowledge, no previous work has employed this technique to identify sexism in social networks. Our single-model approach achieved competitive results, with a performance close to top-performing systems despite sharing parameters for both languages and tasks.
The rest of this paper is organized as follows: in section 2, we discuss related works. In section 3, we describe the classi cation system. Results and analysis are presented in section 4. Finally, the conclusions and future works are given in section 5. 2
The detection of hate speech and misogyny are tasks that are closely connected
and often confused with sexism detection [
Recently, the IberEval competition focused on the automatic identi cation
of misogyny in Twitter [
The appearance of multilingual transformers has shifted the trend in natural
language processing, with many positive experimental results for hate speech
detection. For instance, [
Multi-task learning (MTL) has proven successful in many Natural Language Processing (NLP) problems, as illustrated in the overview of [17]. In this paradigm, multiple tasks are simultaneously learned by a shared model o ering advantages like improved data e ciency, reduced over tting through shared representations, and fast learning by leveraging auxiliary information. Only a few studies are using MTL to detect hate speech language. [18] employed emotion detection as the auxiliary task to address the detection of abusive language. Another example can be found in [19], where a MTL approach was applied to detect hostile content.
Although multilingual and multi-task models have been tested as end-to-end solutions for several tasks related to hate speech, to the best of our knowledge, no previous work has explicitly used these techniques for the sexism detection task. 3
The shared task EXIST 2021 at IberLEF 2021 [20] asked participants to classify \tweets" and \gabs" in two di erent languages, English and Spanish. The objective of the shared task is to develop methodologies and classi cation systems to detect sexist messages according to the following two tasks: { Task 1: It is a binary classi cation task, where every system should determine whether a text or message is sexist or non-sexist. { Task 2: Once a message has been classi ed as sexist, the second task aims to categorize the message according to 5 types of sexism: Ideological and inequality, Misogyny and non-sexual-violence, Objecti cation, Sexual violence, Stereotyping and dominance.
Task 1 is evaluated in terms of accuracy, while for Task 2 the evaluation consists in the macro-average of the F1-scores on the 6 classes: Non-sexist, Ideological and inequality, Misogyny and non-sexual-violence, Objecti cation, Sexual violence, Stereotyping and dominance. Each participating team could submit a maximum of 6 runs, 3 runs for each task.
Two di erent datasets were shared during the challenge. In total, the organizers provided 6977 tweets for training and 4368 texts for testing composed by 3.386 tweets and 982 gabs. The organizers ensured class balancing according to task 1, while the distribution of data for task 2 was relatively unbalanced, re ecting a more natural distribution of sexist content. 4
In recent years, transformer-based language models like BERT [
For our work, we ne-tuned three di erent state-of-the-art multilingual transformer models: mBERT, XLM-RoBERTa [23], and XLM-Twitter [24]. mBERT shares the same training as single-language BERT but using a concatenated dataset of 104 languages, XLM-RoBERTa (XLM-R) was trained on data from 100 and XLM-Twitter (XML-T) makes start from XLM-R and continue pretraining on a large corpus of Twitter in 30 languages.
While, in most cases, the multilingual models are trained and tested independently for each language and do not combine di erent languages in a single evaluation, our approach allows us to tackle the task for both languages at the same time and share the same model. 4.1
As the tasks are evaluated independently, we have explored transformer models for each task independently and will be referring to them as single-task models. Figure 1 shows the model architecture for this approach. On top of the transformer model, we added a linear layer to minimize loss function in our particular task. In particular, we used cross-entropy loss for both tasks, with two and six labels respectively: a binary problem for task1 and a multi-class classi cation with 5 types of sexism and non-sexist for task 2. 4.2
To exploit the fact that both tasks share the same data distribution and are semantically connected, we propose to learn a model jointly on both of them. Figure 1 shows the model architecture for this approach. Speci cally, we consider hard parameter sharing [17] among both tasks using a base model, followed by two linear layers for each classi cation task. As base model, we employed all transformer models previously described in this section (section 4).
We experimented with the inclusion of a learnable parameter to control the importance we place on each task in the multi-task learning framework. In particular, we compute loss with the following expression:
L =
LT ASK1 + (1 )LT ASK2
Where LT ASK1 and LT ASK2 are cross-entropy losses for each task. Since in our problem both tasks are equally important, we set an initial value of = 0.5. 4.3
Data Augmentation is a quite popular solution to improve systems generalization by generating slight variants of the given dataset and is extremely useful for small datasets [25].
For our approach, we did some experiments concatenating the MeTwo dataset
[
All the experiments were performed using Pytorch [26] and HuggingFace [27] Transformers library. As the implementation environment, we used a NVIDIA Tesla T4 GPU. Optimization was done using Adam [28] optimizer with an initial learning rate of 2 5 and a linear weight decay of 0.01 for training single-task and multi-task models. We trained all models with a batch size of 16 for 20 epochs with an early stopping of 8 epochs. We make our code publicly available at Github [29].
The only preprocessing step before feeding the input to the transformer tokenizers was converting to lowercase, replacing mentions, hashtags, and URLs with a keyword, and removing punctuation signs.
To evaluate our systems, we trained all models on 70% of the training data, and held out the remaining 30% for validation. 5.2
Here, we report the performance of the approaches described in the previous section. Table 1 summarizes the results obtained for di erent experiments in the validation set, they are reported in terms of accuracy and macro-f1. We observe that, among the individual transformer models, the best performance is obtained using XLM-T. It can be due to the fact that it is pre-trained using data from Twitter, the same datasource of our task.
In the case of multi-tasking approaches, classi ers perform well, having small di erences with respect to single-task systems for task 1 and outperforming them for task 2. Regarding data augmentation using the MeTwo corpus, we can observe that it generally improves results for task 2, which could suggest that adding instances from task 1 improves results in task 2. It also should be noted that the inclusion of a learnable parameter to control the importance we place on each task slightly improves results for task2.
We presented our three classi ers to the challenge using di erent approaches so that we could compare their performance in the test set. In particular, we sent a single-task classi er and two multi-task systems, using data augmentation and a parameter to control the importance of the tasks.
Table 2 illustrates the results obtained in the competition, where the results are reported in terms of accuracy for task 1 and macro-f1 for task 2. Regarding task 1, our single-task multilingual classi er performs quite well, achieving performances comparable to the winning teams. Similarly, our multi-task model performs fairly well, having a di erence of around 2% with respect to the best result.
For task 2, most participants achieved relatively low results, showing the di culty of this task. The multi-task approach yielded our best results and it stays in the top cluster of the competition (11 out of 63 runs). Unlike our experiments, using data augmentation did not perform well in the test set for task 2. This could be due to the inclusion of Gab in the test set, which is biased towards aggressive sexism.
Once the evaluation phase was over, organizers shared the labels for the test set in case participants wanted to perform further tests. We added two extra experiments to table 2 using two models we did not present to the competition. As we can see, we would have obtained slightly better results for task 2. As we observed in our experiments, multi-task approaches yield better results for task 2 than single-task models.
Task 1 Task 2 Accuracy Run Rank Macro-F1 Run Rank Rank-1 0,784 1 0,5787 1 Majority Class (baseline) 0,6845 66 0,4778 62 SVM TFIDF (baseline) 0,522 52 0,522 51 XLM-T-single-task (run 1) 0,772 7 0,544 15 XLM-T-multi-task-and-metwo-balanced (run 2) 0,7324 29 0,5246 22 XLM-R-multi-task-learnable-parameter (run 3) 0,7571 17 0,5509 11 XLM-T-multi-task 0,764 - 0,554 XLM-T-multi-task-learnable-parameter-concat-metwo 0,747 - 0,553 5.3
Although we achieve interesting results, all models are still making some mistakes. To understand better the source of the failures, we have performed a deep analysis of model errors. In particular, we further investigate the results of the single-task XLM-T model for each task.
Figure 2 displays the confusion matrix for tasks 1 and 2. Regarding task 1, the non-sexist class performs worse than the sexist one. For task 2, most errors come from the misogyny-non-sexual-violence and stereotyping-dominance. We attribute this to the heterogeneity of these classes thus many types of sexist attitudes could be part of them. For instance, the sentence \Some woman are so toxic they don't even know they are draining everyone around them in poison. If you lack self awareness you won't even notice how toxic you really are" and \They refuse to arrest the separatist who has broken the nose of a woman for removing ties" are both instances of the misogyny-non-sexual-violence class, but for di erent reasons. On the contrary, ideological-inequality is a more homogeneous group and the performance is better.
To analyze the reasons behind the errors of our model, we used the library transformers-interpret [30] to have more information about the importance of each token towards the predicted class. Figure 3 shows the word importance for some errors examples. In this gure, red means that the token is pushing towards the \incorrect" (and predicted) class, whereas green pulls towards the correct class. As we can see, it turns out that numerous spurious correlations are learned by our classi er: words such as \puta" or \all" trigger the sentence as sexist. Similarly, the existence of words such as \ill" or \vegan" pushes the prediction of the 4th sentence towards non-sexist. The last two examples are related to errors for task 2. Both are cases where the classi ers fail to detect the type of sexism because of the appearance of irrelevant terms like \straight" and \want". 6
In this paper, we have described a classi cation model for sexist language detection in a multilingual scenario. We also compared single-task to multi-task approaches and experimented with data augmentation techniques using a corpus from the same domain. The results obtained in the framework of the EXIST 2021 competition are promising since our single-model approach had close performance to top-performing systems despite sharing parameters for both languages and tasks. Furthermore, the results show how our model tted spurious correlations for certain terms that must be carefully analyzed with more experiments.
As future work, we plan to experiment with the inclusion of a ective lexicons to improve the automatic detection of sexism. It is also important to note that the strategy to construct the dataset is keyword-based, which can introduce natural biases towards certain sexist terms. Thus, bias mitigation techniques could be useful to improve performance.
This work was supported by the Spanish Ministry of Science and Innovation under Project Misinformation and Miscommunication in Social Media (PGC2018096212-B-C32). Speech and Language Technologies for Dravidian Languages (EACL 2021), pp. 270{276. (2021) 17. Crawshaw, M.: Multi-Task Learning with Deep Neural Networks: A Survey, arXiv (2020) 18. Rajamanickam, S., Mishra, P., Yannakoudakis, H., Shutova, E.: Joint Modelling of Emotion and Abusive Language Detection. In: Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL 2020), pp. 4270| 4279. (2020) 19. Kamal O., Kumar, A., Vaidhya, T.: Hostility Detection in Hindi Leveraging Pretrained Language Models. In: First International Workshop, Combating Online Hostile Posts in Regional Languages during Emergency Situation 2021, Collocated with AAAI 2021, (CONSTRAINT 2021), pp. 213{223. (2021) 20. Rodr guez-Sanchez, F., Carrillo-de-Albornoz, J., Plaza, L., Gonzalo, J., Rosso, P., Comet, M., Donoso, T.: Overview of EXIST 2021: sEXism Identi cation in Social neTworks. Procesamiento del Lenguaje Natural 67 (2021). 21. Liu, Y. et al.: RoBERTa: A Robustly Optimized BERT Pretraining Approach, arXiv (2019) 22. Pires, T., Schlinger, E., Garrette, D.: How Multilingual is Multilingual BERT?.
In: Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (ACL 2019), pp. 4996-{5001. (2019) 23. Conneau, A. et al.: Unsupervised Cross-lingual Representation Learning at Scale.
In: Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL 2020), pp. 8440|8451. (2020) 24. Barbieri F., Anke, L., E., Camacho-Collados, J.: XLM-T: A Multilingual Language
Model Toolkit for Twitter, arXiv (2021) 25. Feng, S., Y. et al.: A Survey of Data Augmentation Approaches for NLP, arXiv (2021) 26. Paszke, A. et al.: PyTorch: An imperative style, high-performance deep learning library. In: 33rd Conference on Neural Information Processing Systems (NeurIPS 2019), (2019) 27. Wolf, T. et al.: Transformers: State-of-the-Art Natural Language Processing. In: Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations (EMNLP 2020), pp. 38-{45. (2020) 28. Kingma, D., P., Ba, J.: Adam: A method for stochastic optimization. In: 3rd International Conference on Learning Representations (ICLR 2015), (2015) 29. Code, https://github.com/franciscorodriguez92/exist2021. Last accessed 25 May 2021 30. Transformers interpret library, https://github.com/cdpierse/transformersinterpret. Last accessed 25 May 2021