In this paper, we present the contributions of the I2C-UHU team to the EXIST2024 Lab at CLEF 2024, focusing on the identification of sexism and the classification of source intent in social media texts. State-of-the-art transformer models are employed to address the complex and nuanced nature of sexist language. We adopt a two-fold approach: firstly, classifying tweets as sexist or non-sexist, and secondly, categorizing sexist tweets based on intent. Our innovative approach, employing Learning with Disagreement, incorporates diverse perspectives from multiple annotators, enhancing the robustness and accuracy of our models. We detail our data preprocessing, augmentation techniques, and hyperparameter optimization strategies. Our results in the competition demonstrated efectiveness, with our entries achieving positive rankings in the two tasks in which we participated. In Task 1, we secured the 10th position out of 70 participants on the hard labels leaderboard and the 13th position out of 40 for soft labels. In Task 2, we achieved the 11th position out of 46 participants for hard labels and the 17th position out of 35 in the best run for soft labels. Our findings provide a foundation for future research and practical applications in social media moderation and policy-making.
In the EXIST2024 Lab at CLEF 2024[
In the realm of detecting sexist tweets, researchers use various methodologies to navigate the complexities of language and intent. Binary classification models serve as a foundational tool, ofering a clear distinction between sexist and non-sexist content. However, the quest for a deeper understanding prompts the exploration of author intent, which requires delving into contextual cues and linguistic subtleties.
Task 1 of EXIST 2024 [
Task 2, however, takes a deeper dive into the realm of author intent, recognizing that the mere
presence of sexist language does not always imply malicious intent. To address this, researchers delve
into the intricate interplay between language, context, and underlying motives. Waseem and Hovy [
In sum, the exploration of related works underscores the multidimensional nature of detecting sexist tweets. While binary classification models provide a solid foundation, the pursuit of a more nuanced understanding necessitates the integration of author intent analysis and cutting-edge transformer models. These endeavors collectively advance our comprehension of sexism in online discourse and pave the way for more efective mitigation strategies.
In this section, the tasks in which participation was engaged and the datasets provided by the organizers are delineated.
Task 1 involves a binary classification problem where the objective is to determine whether a given tweet contains sexist expressions or behaviors. The classification is straightforward: each tweet is categorized as either sexist (“YES” ) or not sexist (“NO” ). Examples of sexist tweets include statements that directly express sexist sentiments, describe sexist situations, or criticize sexist behaviors. For instance, tweets that demean women’s capabilities, perpetuate stereotypes, or contain derogatory comments fall into the “YES” category. Conversely, tweets that do not exhibit these characteristics are labeled as “NO”.
Task 2 is a multi-class classification task aimed at understanding the intention behind sexist tweets. This task only applies to tweets already identified as sexist in Task 1. The intention of the tweet’s author is classified into one of three categories: • DIRECT: The tweet itself is overtly sexist. For example, a tweet stating, “A woman’s place is in the home,” directly conveys a sexist message. • REPORTED: The tweet reports or describes a sexist incident or situation. An example is, “Today,
I saw a man harass a woman on the subway.” • JUDGEMENTAL: The tweet condemns or criticizes sexist behaviors or situations. For instance, “It’s disgraceful how women are still paid less than men for the same work.”
Each of these categories provides insight into the various ways sexism can manifest and the diferent contexts in which it is discussed on social media.
The dataset provided by the organizers contains over 8000 labeled tweets in English and Spanish, with balanced language distribution. The training dataset has 6920 tweets and the development dataset 1038 tweets. Provided in JSON format, each tweet includes attributes such as “id_EXIST”, “lang”, “tweet”, “number_annotators”, and detailed annotator information (“annotators”, “gender_annotators”, “age_annotators”, “ethnicity_annotators”, “study_level_annotators”, “country_annotators” ). Labels are “labels_task1” for sexist content and “labels_task2” for author intent. The “split” attribute indicates the dataset subset and language. In Tables 1 and 2 examples of instances for Task 1 and Task 2 are described.
These instances were extracted from the file "training.json", which contains 6920 instances, of which 3660 are in Spanish and 3260 are in English. In the case of the file "dev.json", which contains 1038 total instances, the language distribution is 549 for Spanish and 489 for English.
In the training and development datasets, the distribution of ethnicities shows a predominant representation of the "White or Caucasian" group, followed by the "Hispanic or Latino" category. Additionally, regarding educational levels, the most common is "Bachelor’s degree," while the least represented are "Less than high school diploma" and "Doctorate." The class distributions for both the binary classification task (YES/NO) and the multiclass classification task (DIRECT, REPORTED, JUDGMENTAL) demonstrate substantial consistency between the training and validation datasets. The class distribution in the training and development datasets is depicted in Figure 1.
To efectively apply Learning with Disagreement techniques, it’s important to study how diferent annotator profiles are distributed across the labeled instances.
For training the dataset, there is an equal number of male and female annotators, with a total of 20760 annotators of each gender and a total of 41520. For the development dataset, the distribution is the same, 50% male (3114) and 50% female (3114), out of a total of 6228 annotators. The age distribution is also equitable across both datasets, with one third of annotators falling into each of the following age groups: 18-22 years, 23-45 years, and over 46 years.
In this section, the methodology used to develop the model submitted to the competition is described.
As previously described, our approaches are based on the use of transformer-based language models.
Given that the provided data is in both English and Spanish, four pre-trained models were chosen.
• XLM-RoBERTa Base: A pre-trained language model utilizing the RoBERTa architecture and
trained on multiple languages. It excels in eficiently and accurately understanding and generating
text in various languages[
The first step in developing classification tasks was to establish an initial benchmark or baseline. This baseline establishes a fundamental methodology that serves as a reference point for comparing more advanced models. It sets a performance threshold that other models must exceed in text classification for our approaches. Two baselines, Version A and Version B, were developed for addressing both Task 1 and Task 2.
MENTAL class. This pattern is consistent across other models, indicating dificulty in classifying all the labels together.
In Version B, the initial step involves classifying tweets into the two categories of Task 1 (YES and NO). Subsequently, tweets that are categorized as YES are further divided into the three distinct classes of Task 2 (DIRECT, REPORTED, and JUDGEMENTAL). The outcomes achieved with this "Baseline Version B" are detailed in Tables 4 and 5.
A schematic overview illustrating the distribution and creation of datasets employed for training the models is shown in Figure 2. These datasets are used in both Task 1 (annotated as v1.x-d) and Task 2 (annotated as v2.x-d). For example: • The v1.1 model is trained and validated with the v1.1-d dataset: – v1.1-d train (training data) – v1.1-d valid (validation data) – v1.3-d train (training data) – v1.3-d valid (validation data) • The v1.3 model is trained with the v1.3-d dataset: • The v2.1 model is trained with the v2.1-d dataset: – v2.1-d train (training data) – v2.1-d valid (validation data)
In the context of NLP, text preprocessing as data cleaning and normalization, are critical to ensuring texts are consistent and noise-free before use in machine learning models. Specifically, for cleaning tweets, the following techniques were employed: • Lowercase Conversion: Ensures uniform treatment of words, eliminating the distinction between "Cat" and "cat", simplifying the dataset and reducing the number of unique features. • Removal of Links: Eliminates web links present in tweets as they do not add semantic value and are often irrelevant to sentiment analysis or text meaning. • Removal of User Mentions: Removes mentions of other users and retweets, which usually do not provide relevant information for semantic analysis and can introduce noise. • Removal of Hashtags: Simplifies the text by removing hashtags, which may not be relevant for semantic analysis, focusing the analysis on words and phrases. • Removal of Emojis: Although emojis convey emotions or contexts, their interpretation can be complex in textual analysis. Initial attempts to translate emojis into words did not improve results, thus they were removed to reduce noise and simplify analysis.
An example of the data cleaning carried out is presented in Table 6. Collab betweet WeAreEqual X @TaravaNFT collab betweet weareequal x ? you already know ? YOU ALREADY KNOW IT. Join our Dis- it. join our discord on how to join our exclusive cord on how to join our exclusive Giveaway : giveaway : . https://t.co/x3stzfLLmh. #NFT #NFTGiveaway #art
Data augmentation is a crucial technique in natural language processing (NLP) to enhance the
performance of machine learning models by artificially expanding the dataset. Various strategies, including
back-translation, have been employed to improve model robustness and generalization. Recent studies
have demonstrated the efectiveness of data augmentation in text classification tasks, emphasizing
its importance in handling diverse linguistic patterns and enhancing model accuracy [
Oversampling addresses class imbalance [16] by generating syntactic and lexical variations through backtranslation, increasing dataset diversity without altering meaning [17]. Since the datasets are unbalanced, it is necessary to employ a balancing technique. In this case, the number of rows for the REPORTED and JUDGEMENTAL classes has been increased through backtranslation, while the original number of rows has been maintained for the DIRECT class. Using Helsinki-NLP/opus models from the OPUS project [18], tweets in Spanish are translated to English, then German, and back to Spanish. An example of data generation through backtranslation for a tweet in Spanish is shown in Table 7. Se supone q me tengo q avergonzar d ser mamá? ¿Debería avergonzarme de ser madre? Jajajajaajajaja naaaa
For tweets in English, they were translated from English to German, then from German to Spanish, and finally from Spanish back to English. An example of a newly generated instance is shown in Table 8. Easy to throw rocks and hide behind your gender Easy to throw stones and hide behind your sex or or sexual identity #onhere sexual identity #onhere
Hyperparameter tuning optimizes model performance by selecting optimal values for non-learned parameters. Optuna [19] helps define and iteratively optimize the hyperparameter search space. Ex
In reference to the metrics obtained after hyperparameter optimization and the application of the previously explained techniques, the results are explained in Tables 10 and 11.
Training was conducted using the Trainer class from Hugging Face, incorporating optimized hyperparameters. The adamw _torch[20] optimizer was employed for updating model weights, with evaluations conducted at the end of each epoch and models saved periodically. The best model, determined by the F1 metric, was loaded. Training was halted using the EarlyStoppingCallback if no improvements were observed. These strategies were then tested on the structured dev dataset. The RTX 4070 graphics card was utilized for its high performance and capability to manage intensive processing tasks, thereby ensuring eficient and speedy development and execution of complex models.
To train the final models that will generate predictions on the test data provided by the competition for Task 1, we selected the two best-performing models based on their metrics during the training process.
To train the v1.1 model, data from the v1.1-d dataset was used. Each tweet in this dataset is labeled by six annotators in both the training and validation sets. To obtain the majority label, following the competition guidelines to obtain the gold label, the votes were averaged, selecting the labels that received two or more votes from among the six possible annotators. In case of a tie, the instance in question was completely excluded. A multilingual model, XLM-RoBERTa-Base, was trained to handle both English and Spanish instances simultaneously. Figure 3 shows this process.
Baseline
Data augmentation + Hyperparameters Baseline
Data augmentation + Hyperparameters
Subsequently, this model was used to predict the labels of the data in the oficial competition test set. The results are presented indicating the majority predicted label for each instance of the test set, followed by the score_label, which represents the similarity score assigned by the classifier to the majority predicted label on a scale of 0 to 1.
To obtain the hard label, the majority predicted label was selected. Regarding the soft label, since it is a binary classifier (YES or NO), the score_label value was assigned to the majority class in each case, and the value of the minority class was calculated as 1 minus the score_label. It is important to note that the sum of the label values in the soft results should not exceed 1. The results model’s evaluation is shown in the Table 12
The training process for model v1.2 is almost identical to the previously described process, but with some key diferences regarding the models used and the workflow structure. For training, two datasets were used: v1.3-d for English instances and v1.4-d for Spanish instances. As for the models, DeBERTa v3 Base was used for English and RoBERTa Base Bne for Spanish. The figure 4 shows the process.
The workflow began with the separate training of the two models: the DeBERTa v3 Base model was used for the English instances of dataset v1.3-d, and the RoBERTa Base Bne model was employed for the Spanish instances of dataset v1.4-d. The models’ evaluation results are shown in Table 13
Model Version v2.1 was designed to address the second task of the competition, which focuses on classifying the intentionality of tweets previously categorized as sexist by model version v1.2 (Source Intention in Tweets). This task follows the initial classification of sexist messages and seeks to categorize such messages according to the author’s intent, thus providing insights into the role of social media in issuing and spreading sexist messages. In this task, a classification between three classes DIRECT, REPORTED, and JUDGEMENTAL is proposed.
The training data comes from dataset version v2.1-d, containing only instances of the three classes, excluding instances categorized as NO, thus avoiding introducing noise in the training data and refining the model’s accuracy. Only hard labels were generated for the final predictions, as the model does not return the score label of predicted classes as minority. The Figure 5 shows the process. Obtained results are shown in the Table 15
The next model applies Learning with Disagreement because it considers and leverages the diferences in opinion among multiple human annotators when labeling the training data. This approach captures a greater diversity of perspectives, which is especially useful in subjective or complex tasks where there may be significant disagreement about the correct labels.
This method improves the model’s predictions by integrating multiple viewpoints, creating a more robust and representative training dataset. Additionally, the soft labels resulting from this process enable the model to capture the uncertainty and variability inherent in human annotations, leading to better generalization and performance in real-world situations where data may not be clear or fully defined. The Figure 6 shows the process. Obtained results are shown in the Table 16.
The training flow of the model shown in the image can be explained in detail, focusing on how the disagreement among annotators is handled and how soft labels are generated. Here’s the step-by-step explanation: 1. Training data comes from six groups of annotators diferentiated by gender and age: ["F 18-22", "F 23-45", "F 46+", "M 46+", "M 23-45", "M 18-22"]. Each group of annotators has provided labels for the training data. 2. Six datasets (v2.2.1-d, v2.2.2-d, v2.2.3-d, v2.2.4-d, v2.2.5-d, and v2.2.6-d) are used to train six instances of the XLM-RoBERTa Base model. Each dataset corresponds to the annotations of one of the six mentioned groups. 3. The six trained models are combined using an ensemble method. This process integrates the outputs of the diferent models to produce a more robust final prediction. The ensemble calculates a weighted average (sum) of the predictions of the six models. 4. To generate the soft labels, the proportion of annotators who voted for each label is taken into account. For example, if 2 out of 6 annotators labeled a data point as "DIRECT", the soft label for "DIRECT" would be 2/6 = 0.33333. This process is repeated for the other labels, "REPORTED" and "JUDGEMENTAL".
In the previous task (Task 1), the data was classified into the classes "YES" and "NO". If a data point was classified as "YES" with a probability of 0.80, this value is used to adjust the soft labels of Task 2. For example, if the soft label for "DIRECT" is 0.33333, the adjusted value would be 0.33333 * 0.80 = 0.26666. This adjustment is performed for all sub-classes of "YES" ("DIRECT", "REPORTED", and "JUDGEMENTAL").
This process must be done for the YES label when it is the majority class in Task 1, as well as to predict the percentage of this when it is the minority class in Task 1. In conclusion, the extremely low probability of the diferent YES classes in the instances that have been classified by the models of version 1 as NO is also being calculated.
Finally, Model Version 2.3 follows the same guidelines as Version 2.2, explained above, but the training data comes from three groups of annotators diferentiated by gender and age: ["F 18-22", "F 23-45", "F 46"]. Only female groups have been selected to train the models that will compose the ensemble. The Figure 7 shows the process. Obtained results are shown in the Table 18.
4.6.1. Task 1 This section provides a detailed analysis of errors made by the models in Task 1: Sexism Identification in Tweets, focusing on classification discrepancies between YES and NO classes. By scrutinizing misclassifications, patterns and insights into challenges faced by the models are aimed to be identified. Additionally, potential strategies to improve classification performance, especially for the minority class (YES), are explored. Examples are presented in Table 18. Woman driving beside me a few minutes ago holding her phone to her ear with her shoulder, while holding a mug of cofee. Baby on
Boardsticker on both rear windows.
Por qué todos los hombres cuando su novia o esposa está embarazada andan más de culeros que de costumbre NO YES
YES NO 4.6.2. Task 2 This section analyzes errors encountered by models in Task 2: Source Intention in Tweets, focusing on classification accuracy across DIRECT, REPORTED, and JUDGEMENTAL categories. Through examination of misclassifications, factors influencing performance across these categories are aimed to be understood, and refinements to improve the model’s ability to discern nuanced intentions in sexist tweets are discussed. Examples are provided in Table 19, and confusion matrices in Figure 8 depict prediction distributions for Task 2 models. Lo irónico es que en su mayoría sean hombres quienes apoyan la criminalización de las mujeres frente al aborto. Claro, a las mujeres hay que castigarlas, juzgarlas y señalarlas siempre, como si no fuera
suficiente tener que cargar con el peso de una violación.
En total delirio esta tipa quiere legalizar el terrorismo. ¿Y esta escoria quiere definir los destinos de Chile? Permitirlo es de anti chilenos.
If you don’t vote, you ARE the problem. #VoteBlueIn2022 #WomensRights #GunControl #bookban #CivilRights #VotingRights REPORTED
DIRECT DIRECT
NO
NO REPORTED
The analysis of errors in Task 1 and Task 2 uncovers various reasons for misclassifications. Many tweets feature nuanced language or context, challenging for models to interpret. For example, a tweet warning about sympathetic individuals may discuss predatory behavior broadly, misinterpreted by the model as sexist content. Tweets often employ sarcasm, idiomatic expressions, or ambiguous wording, leading to misclassification. A tweet about a woman multitasking while driving may be misconstrued as a gender stereotype critique rather than a comment on unsafe driving practices. Multilingual or culturally referential tweets add complexity. A Spanish tweet discussing men’s behavior could be viewed contextually as commentary on male behavior patterns rather than explicit sexism.
In Task 1, the best-performing strategy was a combination of models for diferent languages: RoBERTa Base BNE was used for classifying Spanish tweets, and DeBERTa v3 Base was employed for English tweets. This dual-model approach significantly outperformed other strategies, emphasizing the efectiveness of leveraging specialized models for each language. Following this, the multilingual model XLM RoBERTa Base also showed strong performance, though it was slightly behind the combined approach. In Task 1, Model v1.1 produced the run I2C-UHU_1, while v1.2 produced I2C-UHU_2. The oficial results for Task 1 are shown in the tables 20 and 21.
In Task 2, the best results were achieved using the Learning with Disagreement method with six groups of annotators (three male and three female). This approach outperformed the run that applied Learning with Disagreement with only three groups of female annotators. This finding suggests that having a more diverse set of annotators can enhance the model’s performance by providing a broader range of perspectives, which likely leads to better generalization and robustness in the model’s predictions.For Task 2, v2.1 generated the run I2C-UHU_1, v2.2 produced I2C-UHU_2, and v2.3 resulted in I2C-UHU_3. The oficial results for Task 2 are shown in the tables 22 and 23.
In this paper, the efectiveness of advanced transformer models in addressing the identification of sexism and the classification of source intent in social media texts has been demonstrated. The approach employed, which integrates Learning with Disagreement, facilitates the incorporation of diverse annotator perspectives, thereby enhancing the robustness and accuracy of the models. The methodology, consisting of classifying tweets as sexist or non-sexist and subsequently categorizing the intent of sexist tweets, has shown significant improvements in understanding and detecting nuanced sexist content. The results of the EXIST 2024 Leaderboard for Task 1 and Task 2 provide valuable insights into efective strategies for multilingual tweet classification and the impact of annotator diversity. For Task 1, superior performance was observed with the combination of language-specific models (RoBERTa Base BNE for Spanish and DeBERTa v3 Base for English), indicating the benefit of using specialized models tailored to individual languages. Meanwhile, Task 2 results indicated that Learning with Disagreement, utilizing a diverse set of annotators (both male and female), led to better outcomes compared to using only female annotators. This underscores the importance of diversity in annotation to capture a wider array of linguistic nuances and biases, thus improving the overall performance of the model. Future work will focus on refining the models by incorporating additional data sources and exploring more sophisticated ensemble methods. Additionally, eforts will be made to extend the research to other forms of harmful online content, applying the insights gained from this study to broader applications in social media moderation and policy-making. The insights derived from this research provide a valuable foundation for the development of more efective strategies to combat online sexism and other forms of digital harm.
This paper is part of the I+D+i Project titled “Conspiracy Theories and hate speech online: Comparison of patterns in narratives and social networks about COVID-19, immigrants, refugees and LGBTI people [NON-CONSPIRA-HATE!]”, PID2021-123983OB-I00, funded by MCIN/AEI/10.13039/501100011033/ and by “ERDF/EU”. [13] Author(s), Backtranslate what you are saying and i will tell who you are, Journal Name (2024). [14] Author(s), Data augmentation using back-translation for context-aware neural machine translation,
Journal Name (2019). [15] Author(s), Back-translation-style data augmentation for end-to-end asr, Journal Name (2018). [16] S. Kobayashi, Contextual augmentation: Data augmentation by words with paradigmatic relations, in: Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), 2018, pp. 452–457. [17] S. Y. Feng, V. Gangal, J. Wei, S. Chandar, S. Vosoughi, T. Mitamura, E. Hovy, A survey of data augmentation approaches for nlp, 2021. arXiv:2105.03075. [18] M. Aulamo, J. Tiedemann, The OPUS resource repository: An open package for creating parallel corpora and machine translation services, in: M. Hartmann, B. Plank (Eds.), Proceedings of the 22nd Nordic Conference on Computational Linguistics, Linköping University Electronic Press, Turku, Finland, 2019, pp. 389–394. URL: https://aclanthology.org/W19-6146. [19] M. Junczys-Dowmunt, R. Grundkiewicz, T. Dwojak, H. Hoang, K. Heafield, T. Neckermann, F. Seide, U. Germann, A. Aji, N. Bogoychev, A. Martins, A. Birch, Marian: Fast neural machine translation in c++, 2018, pp. 116–121. doi:10.18653/v1/P18-4020. [20] I. Loshchilov, F. Hutter, Decoupled weight decay regularization, 2019. arXiv:1711.05101.