The pervasive presence of sexist content on social media platforms not only perpetuates harmful stereotypes but also fosters environments that can be exclusionary and hostile, especially towards women. Such content, which often targets people of a specific gender, i.e., sexist content, requests platforms to enhance their monitoring and policing eforts. Yet, policing such content is challenging for many reasons, including the volume of messages to check and the context of the content. Consequently, several studies have been conducted to automatically detect sexist language on social media, focusing on its identification and classification. However, variations in detection accuracy can depend on the diferences in architecture, training strategies, and data of existing models, [1, 2, 3], leading to potential variances in detection accuracy. This variability, further influenced by the types of messages and input prompts, motivates our exploration into the fusion of multiple Large Language Models (LLMs). As part of EXIST Task 1, which focuses on sexism identification in multilingual contexts, we introduce two novel approaches: the Dual-Transformer Fusion Network (DTFN) and the Multimodel Fusion Ensemble (MFE). These methods utilize fusion and ensemble learning techniques to enhance detection accuracy across multilingual datasets. Our extensive experimental evaluation during the EXIST 2024 competition demonstrates that these methodologies significantly outperform existing models, with MFE and DTFN ranking 1st and 2nd, respectively, in the English segment, and 4th and 13th in the combined English and Spanish segments of the oficial leaderboard.
The proliferation of social media platforms has fundamentally transformed how individuals communicate. However, these platforms have also become arenas for problematic interactions, including the dissemination of sexist content. This content not only perpetuates harmful stereotypes but also fosters an online environment that can be hostile and exclusionary, particularly towards women. The urgency to address this issue is underscored by the growing body of research indicating the growing exposure to sexist language and its profound impacts. Despite the clear need to mitigate this problem, the task of detecting sexist content online presents substantial challenges. Sexist language is not uniformly explicit; it often involves subtle cues and context-dependent expressions.
Addressing these challenges requires leveraging flexible computational methods and approaches
that can understand and interpret the complexities of language used in these settings. Large Language
Models (LLMs), which are pre-trained on vast corpora and fine-tuned for specific tasks, are promising
solutions. However, while individual LLMs ofer robust linguistic insights, they also have inherent
limitations when applied to specific tasks or domains such as detecting sexist or harmful content. Each
existing model may interpret nuances diferently based on its architecture, training strategy and data
[
This variability depends on the types of messages as well as on the input prompts, and it motivates
our exploration for fusing multiple LLMs. Our research, conducted as part of the EXIST 2024 (sEXism
Identification in Social Networks)[
In particular, the DTFN is a simple yet efective approach that integrates the outputs from two distinct
transformers, i.e., RoBERTa[
The efectiveness of these methods was evaluated in the EXIST 2024 competition - Task 1, where the MFE and DTFN approaches notably outperformed other methodologies, ranking 1st and 2nd in the English segment of the oficial leaderboard, and 4th and 13th in the combined English and Spanish languages, respectively.
The remainder of this paper is structured as follows: Section 2 provides an overview of related work in sexism detection and ensemble learning strategies. Section 3 briefly describes the datasets. Section 4 details our methodologies, including the DTFN and MFE techniques. Section 5 details Experimental Assessment and Section 6 presents the results and analysis, followed by our conclusions in Section 7.
Significant research has been conducted by researchers on the detection of hate speech, cyberbullying and ofensive language. However, despite the growing interest on the topic, the literature on sexism detection is still limited.
In recent studies large language models (LLMs) and Transformer-based architecutres have been used
for multi-modal detection of hate speech, sexism and ofensive language from the text, images, memes,
audio, and videos[
Moreover, the problem of sexism detection involves identifying harmful and biased language, often embedded within complex social contexts. Early approaches relied heavily on traditional machine learning techniques, such as SVMs and logistic regression, combined with manually crafted features Gaydhani et al. [18], Anistya and Setiawan [19]. However, these methods struggled with the subtleties of natural language and the contextual nature of sexism. Recent studies have leveraged the transformer models to address these challenges. In particular, [20] applied BERT to detect misogyny in social media. Similarly, Singh et al. [21] focused on the automatic detection of misogyny in multimodal online content by developing a large, annotated corpus of memes involving Hindi-English code-mixed language.
Ensemble approaches, which involves combining multiple models, have proven efective in various
NLP tasks, and the rationale is that diferent models can capture diferent aspects of the data based
on their architecture, training objectives and data; thus, their combination can mitigate individual
weaknesses. [22] provided a comprehensive overview of ensemble methods, emphasizing their potential
to enhance robustness and accuracy. In the context of text classification, Stacked Generalization [ 23]
and Bagging [24] are widely used ensemble techniques. More recent studies have focused on applying
these methods to deep learning models. For example, [
In this study, we utilized the EXIST 2024 Tweets Dataset, specifically tailored for Task 1, which involves sexism identification in tweets. This dataset is comprehensive, containing over 10,000 labeled tweets balanced between English and Spanish. The tweets are annotated for binary classification, where the task is to determine whether a tweet contains sexist expressions or behaviors, categorized as "YES" or "NO."
The dataset is split into three parts: training, development, and test sets. For our experiments, we used the training and development sets with hard labels (gold standard) for model training and validation. The detailed distribution of the dataset is described in Table 1.
We proceed by introducing the two methodologies designed to address the classification of online sexism. First, we first present a simple yet efective method, named Dual-Transformer Fusion Network (DTFN) (see Section 4.1), based on the fusion of vector representations generated by two diferent neural language models; we subsequently present a more complex and efective approach, Multimodel Fusion Ensemble (MFE) (Section 4.2), based on ensemble learning of several LLMs and of the aforementioned DTFN.
In our participation in the EXIST Task 1, we introduced a methodology named Dual-Transformer Fusion
Network (DTFN), which integrates two Transformer models known for their efectiveness in online
post analysis, namely RoBERTa-Large and DeBERTa-V3-Large [
Based on these observations, we design DTFN as a hybrid architecture that first processes input text —typically extracted from social media or other online platforms— through both models in parallel. Each model then independently analyzes the text and outputs dense representation from their respective last hidden layers, potentially encoding complementary aspects of the text’s semantic and contextual nuances.
Formally, let x denote the input text vector. Each Transformer model (RoBERTa and DeBERTa) processes x independently and outputs a representation h from its final hidden layer: Input Data DeBERTa-V3-Large
Fully-Connected Layer
Final Output
These output vectors, h and h, capture complementary linguistic features as determined by their distinct training paradigms and architectural innovations. Following the feature extraction, the outputs are concatenated to form a unified feature representation h: (2) (3) (4) h = [h; h] ^ = ((h))
This vector h is then passed through a fully connected linear layer to produce the final class prediction ^. The linear layer acts as a classifier, integrating the diverse features into a unified prediction: where denotes the sigmoid activation function, mapping the linear combination of features to a probability score indicating the final class predicted. The entire architecture is trained end-to-end with the objective of minimizing the binary cross-entropy loss ℒ:
ℒ(^, ) = − log(^) − (1 − ) log(1 − ^)
The linear layer, as the whole architecture, is trained end-to-end on the specific task of sexism detection. Figure 1 illustrates the overall pipeline of the DTFN, highlighting the flow from the input text to the final classification.
Based on the promising results of our preliminary study combining two Transformer architectures, we
devise a principled approach based on ensemble to dynamically combine multiple models with diferent
architectures and training strategies. This approach, which we named Multimodel Fusion Ensemble
(MFE), combines multiple models with distinct architectures and training strategies. Specifically, MFE
integrates outputs from four diferent Transformer-based models — RoBERTa-Large[
Each model in the ensemble was selected for its unique capabilities in processing and understanding
complex text structures, such as the dynamic masking strategy [
More formally, in our ensemble method the classification decision for each instance is derived through a majority voting mechanism among the outputs of the constituent models. Let = {1, 2, . . . , } represent the set of possible classes. For a given text instance , each model in the ensemble = {1, 2, . . . , } predicts a class . The ensemble prediction ^ is determined by: ^() = arg max ∑︁ 1(() = ) ∈ =1 (5) where 1 is the indicator function that equals 1 if the condition is true and 0 otherwise. This simple approach counts the votes for each class from all models and selects the class with the highest count. Majority Voting and Tie Handling In scenarios where the voting results in a tie, particularly when the ensemble is evenly split across classes, a predefined rule is applied to resolve the ambiguity. Considering the sensitivity and potential consequences of misclassifying sexist content, our tie-breaking strategy defaults to the "Yes" (Sexist) prediction. This decision was based on the task’s sensitivity and the potential social impact of underdetecting sexist content.
As part of the EXIST 2024, we conducted a thorough experimental evaluation of the presented methodologies addressing, in particular, Task 1. In our experimental assessment, we initially evaluated the performance of individual models to establish a baseline. Then, we analysed the results yielded by the proposed Dual-Transformer Fusion Network (DTFN) and Multimodel Fusion Ensemble (MFE).
We proceed by first introducing the baselines, hyperparameters, and evaluation metrics adopted. We conclude by discussing the results on the EXIST dataset and the oficial leaderboard, along with quantitative analyses of the ensemble mechanism.
In the following, we briefly describe the baselines evaluated:
• RoBERTa-Large [
We adopted their pre-trained versions, available through the HuggingFace library1.
For each model used in our experiments, we identified a set of optimal hyperparameters through preliminary testing. The selected hyperparameters include the number of training epochs, learning rate ( ), batch size, and weight decay ( ). Table 2 presents the optimal hyperparameters for each model. In the evaluation of Task 1 for EXIST 2024, the oficial metrics used are ICM-Hard, ICM-Hard Norm, and F1. The ICM metric, proposed by [25], is based on information theory and measures the similarity between system classifications and gold standard labels. The organizers have also provided a normalized version, ICM-Hard Norm, to account for dataset imbalances, ensuring fair comparisons across diferent test conditions. For this shared task, higher values of the ICM and ICM-Hard Norm metrics indicate a stronger alignment between system outputs and the ground truth, with higher values considered better.
The evaluation of the baseline models shows diferences in performance across RoBERTa-Large, DeBERTa-V3-Large, Mistral-7b, and our proposed Dual-Transformer Fusion Network (DTFN), as reported in Table 3.
RoBERTa-Large achieved an F1 score of 0.864 and an ICM score of 0.592, demonstrating its robustness in handling the task. DeBERTa-V3-Large marginally outperformed RoBERTa-Large. Mistral-7b, on the other hand, yielded a slightly lower F1 score and the lowest ICM score among the baselines. This indicates that despite the higher number of parameters, Mistral-7b might not be as well-suited to the specific task of sexist identification compared to the other models evaluated.– Our proposed model, the Dual-Transformer Fusion Network (DTFN), slightly surpassed the other baseline models with an F1 score of 0.868 and showed a significant improvement with an ICM score of 0.606. The higher performance of DTFN highlights the eficacy of our dual-transformer architecture in improving classification accuracy. ID
Models
F1 Score ICM Score
Baseline Models To systematically understand the voting results, we analysed the number of times each combination of models agreed on the sexist (Yes) and non-sexist (No) classes. This is done to determine whether there is a dominant model (or combination) in the ensemble. The combinations and their respective agreement counts are detailed in Table 4. The table provides detailed insights into the agreement and non-conformity of various ensemble model combinations on the development set.
Among the combinations, the ensemble of RoBERTa-Large, DeBERTa-V3-Large, Mistral-7b, and DTFN shows the highest majority agreement. Conversely, the combination of RoBERTa-Large, DeBERTaV3-Large, and Mistral-7b without DTFN showed the lowest majority agreement. This indicates that the addition of DTFN significantly boosts the ensemble’s agreement, particularly in identifying sexist content. Out of all instances, we observed a total of 58 ties where the aforementioned tie-breaking rule was applied.
Additionally, we reported the isolation frequency, i.e., how often a single model’s prediction difered from the majority vote within the ensemble, reflecting the model’s conformity with others. RoBERTaLarge had the highest isolation frequency, which could be due to the fact that it takes context into account. DeBERTa-V3-Large showed a lower isolation frequency, while Mistral-7b frequently disagreed with the ensemble on non-sexist classifications. DTFN had the lowest isolation frequency, suggesting it is the most conforming model within the ensemble, underscoring the DTFN’s role in enhancing ensemble cohesion.
In this section, we present the results of our participation in Task 1 of the EXIST 2024 challenge, where our team, EquityExplorers, submitted two runs: EquityExplorer-1 using the DTFN technique and EquityExplorer-2 employing the Multimodel Fusion Ensemble (MFE) approach.
The MFE and the DTFN demonstrated notable performance by ranking 1 and 2 for the Task 1 in the English segment, respectively. The efectiveness of the MFE, evidenced by its ICM-Hard score of 0.6178 and ICM-Hard Norm of 0.8153, coupled with an F1 score of 0.7610, demonstrated its robust capability to discern nuances of sexist content in English tweets efectively. This ensemble approach, by combining diferent strategies and model outputs, has proven to be particularly efective in improving accuracy and reliability over individual models, including the Dual-Transformer Fusion Network.
The patterns observed in the Spanish and Both (combining results from English and Spanish) evaluations align with these findings. Although the performance gap in the Spanish evaluation is wider, it highlights the robustness of MFE in a diferent linguistic environment. Notably, DTFN ranks 25ℎ out of 66 in Spanish, suggesting that while it is efective, it might not fully adapt to diferent languages as eficiently as MFE. The aggregated results for both languages demonstrate the consistent advantage of using an ensemble approach, with MFE achieving the 4ℎ rank out of 70, compared to the 13ℎ rank for DTFN. In conclusion, the oficial leaderboard results validate the proposed approach and highlight the significant improvements achieved through the Multimodel Fusion Ensemble (MFE). The consistent outperformance of MFE across various metrics and datasets underscores the potential of ensemble methods involving neural language models.
In this work, we introduced the Dual-Transformer Fusion Network (DTFN) and the Multimodel Fusion Ensemble (MFE) for identifying sexist content across multiple languages within the context of the EXIST 2024 competition. A thorough evaluation on the development and test sets highlighted the superior performance of the MFE, particularly ranking highly in both the English (1) and combined language (4ℎ) categories. This performance, paired with a comparative analysis against baseline models, allowed for a detailed assessment of the relative improvements ofered by the DTFN and MFE approaches. It demonstrated the benefits of integrating diverse transformer models into an ensemble framework to leverage the characteristics of each neural language model, thereby achieving higher accuracy and reliability in detecting complex linguistic patterns associated with sexism. (Eds.), Findings of the Association for Computational Linguistics: EACL 2023, Association for Computational Linguistics, Dubrovnik, Croatia, 2023, pp. 1079–1091. [15] A. Irfan, D. Azeem, S. Narejo, N. Kumar, Multi-modal hate speech recognition through machine learning, in: 2024 IEEE 1st Karachi Section Humanitarian Technology Conference (KHI-HTC), 2024, pp. 1–6. doi:10.1109/KHI-HTC60760.2024.10482031. [16] J. Devlin, M.-W. Chang, K. Lee, K. Toutanova, Bert: Pre-training of deep bidirectional transformers for language understanding, 2019. arXiv:1810.04805. [17] A. Vetagiri, P. Pakray, A. Das, A deep dive into automated sexism detection using fine-tuned deep learning and large language models, Available at SSRN 4791798 (2024). [18] A. Gaydhani, V. Doma, S. Kendre, L. Bhagwat, Detecting hate speech and ofensive language on twitter using machine learning: An n-gram and {TFIDF} based approach, CoRR abs/1809.08651 (2018). URL: http://arxiv.org/abs/1809.08651. arXiv:1809.08651. [19] F. Anistya, E. B. Setiawan, Hate speech detection on twitter in indonesia with feature expansion using glove, Jurnal RESTI (Rekayasa Sistem dan Teknologi Informasi) 5 (2021) 1044 – 1051. URL: http: //www.jurnal.iaii.or.id/index.php/RESTI/article/view/3521. doi:10.29207/resti.v5i6.3521. [20] A. Rahali, M. A. Akhloufi, A.-M. Therien-Daniel, E. Brassard-Gourdeau, Automatic misogyny detection in social media platforms using attention-based bidirectional-lstm*, in: 2021 IEEE International Conference on Systems, Man, and Cybernetics (SMC), 2021, pp. 2706–2711. doi:10. 1109/SMC52423.2021.9659158. [21] A. Singh, D. Sharma, V. K. Singh, Mimic: Misogyny identification in multimodal internet content in hindi-english code-mixed language, ACM Trans. Asian Low-Resour. Lang. Inf. Process. (2024).
URL: https://doi.org/10.1145/3656169. doi:10.1145/3656169, just Accepted. [22] T. G. Dietterich, Ensemble methods in machine learning, in: International workshop on multiple classifier systems, Springer, 2000, pp. 1–15. [23] D. H. Wolpert, Stacked generalization, Neural Networks 5 (1992) 241–259. [24] L. Breiman, Bagging predictors, Machine Learning 24 (1996) 123–140. URL: https://api.
semanticscholar.org/CorpusID:47328136. [25] E. Amigo, A. Delgado, Evaluating extreme hierarchical multi-label classification, in: S. Muresan, P. Nakov, A. Villavicencio (Eds.), Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Association for Computational Linguistics, Dublin, Ireland, 2022, pp. 5809–5819.