The widespread rise of multimedia social platforms has diversified the ways in which people communicate and the content they share. Hate speech, as a threat to societal harmony, has also shifted its manifestation from a singular textual to a multimodal one. Previously, most methods for detecting hate speech were limited to the text modality, making it difficult to identify and classify newly emerging multimodal hate speech that combines text and images. This paper proposes a novel multi-modal hate speech detection model to respond to the above-mentioned needs for multi-modal hate speech detection. The proposed joint model can use moving windows to extract multi-level visual features and extract text features based on the RoBERTa pre-training model and introduces a multi-head self-attention mechanism in the later fusion process for image and text feature fusion. This article also conducted experiments on the multi-modal benchmark data set Hateful Memes. The model achieved an accuracy of 0.8780, precision of 0.9135, F1-Score of 0.8237, and AUCROC of 0.8532, defeating the SOTA multi-modal hate speech recognition model.
Research Article
Research on Multimodal Hate Speech Detection Based on Self-Attention Mechanism Feature Fusion
https://doi.org/10.21203/rs.3.rs-4836799/v1
This work is licensed under a CC BY 4.0 License
You are reading this latest preprint version
Hateful speech detecting
deep learning
multimodal data analysis
Modal fusion
Attention mechanism
With the growing role of social media and online communities in our daily life, monitoring and managing hate speech has become particularly important. Because it not only expands the channels for people to exchange information and facilitates the way people communicate, but also induces the breeding of hate speech due to the anonymity and freedom of online speech[1]. The United Nations defines hate speech as any kind of communication in speech, writing or behavior, that attacks or uses pejorative or discriminatory language with reference to a person or a group on basis of who they are, in other words, based on their religion, ethnicity, nationality, race, color, descent, gender or other identity facto[2]. To some extent, hate speech can lead to social conflicts and vicious events, and even endanger the harmonious development of society. For example, the second study on hate speech and discrimination in Costa Rica in 2022 found that hate speech caused division and a bad atmosphere in Costa Rica[3]. Therefore, the detection of hate speech has become a key social issue and is regarded as a very important task with potential significant benefits in the research field.
In multimodal scenarios, hate speech detection (HSD) is facing great challenges. Due to the huge amount of data involved, traditional manual detection methods are inefficient [4, 5]. With the development of natural language processing and machine learning technology, HSD has been able to automatically identify hate speech in text [6]. However, the development of multimedia makes people's communication mode not only limited to text, but also more willing to use multimodal superimposed information such as pictures, videos, or text as the carrier of communication. In this process, hate speech has become more difficult to detect due to the diversification of information carriers. As shown in Fig. 1, a certain image may not have hate factors by itself, but it can produce different emotional tendencies when paired with a certain sentence. The example in Fig. 1, after text is paired with images, can express gender discrimination, insult, pornography, and other hateful connotations. The text information in the first image reflects anti-lesbian remarks, the second image implies support for drugs, especially from the needle and other signs in the image, the text in the third image combined with the image implies racial discrimination, and the fourth image compares the hanging of the head to a swing, containing meanings such as violent death.
Multimodal hate speech detection (HSD) is defined as a technology that uses multimodal information including images and text to capture and distinguish hate speech [7]. Compared with a single text analysis method, this method can more accurately and efficiently realize the automation of hate speech detection.
This paper aims to explore the methods and applications of multimodal hate speech detection and construct a multimodal hate speech detection model based on image-text fusion to improve the accuracy and robustness of hate speech detection. In this paper, Hateful Memes, the currently typical multimodal hate speech dataset, will be used for the training and validation of the model.
The main contributions of our work are summarized as follows:
1) This paper proposes a new joint model for multi-modal hate speech detection, which uses a moving window for multi-level visual feature extraction for visual modalities and a RoBERTa pre-trained model for textual modalities. In addition, the structure of the model is relatively clear and highly interpretable.
2) This paper innovatively introduces a multi-head self-attention mechanism at the model fusion stage. This mechanism allows the model to dynamically adjust the weight of text features, effectively fusing image, and text information. Important text parts better match image features better, thereby improving the fusion effect.
3) This paper carries out a series of experiments on the characteristics, content, and modes of the benchmark dataset Hateful Memes for multimodal hate speech. We separate the text and image in the memes, removing the text to avoid interference with image feature extraction. Using innovative joint methods, this paper achieves better results on Hateful Memes than the SOTA model.
This section reviews the literature on the hate speech detection, considering both single text and multimodal aspects.
2.1 Text hate speech detection
here is a large body of relevant literature on hate speech detection. As early as the beginning of 2010, people had begun to study hate speech detection. Most of the existing research on hate speech detection focuses on hate speech content related to immigration or gender discrimination issues[8–12]. In addition, hate speech such as racial discrimination is also the focus of research[13–15]. Flor Miriam Plaza-del-Arco et al. compared the Spanish hate speech detection pre-training language model based on the Transformer mechanism[16].
Early hate speech detection was mainly based on text feature engineering, using techniques such as TF-IDF and n-gram to represent the textual characteristics of hate speech, and on this basis, combining machine learning algorithms such as support vector machines, naive Bayes, and logistic regression to perform classification and detection of hate speech text [17–19]. In addition, dictionary-based methods mainly use fixed or existing keywords (dictionaries) extracted from relevant hate texts for hate speech detection[20]. However, such methods are often limited in feature representation and model generalization. The rise of deep learning technology has brought new possibilities for feature representation and model generalization. Convolutional neural network (CNN) is also commonly used for text classification [21]. Björn Gambäck and Utpal Kumar Sikdar trained four convolutional neural network models respectively, and the word2vec embeddings performed the best with an F1-score of 78.3% [22]. Moreover, the recurrent neural network (RNN) and Transformer model can take context information into account when processing text data, to better capture semantics and context. Alshalan and others compared three network architectures of CNN, GRU and CNN + GRU, and the model effect of BERT was evaluated [23]. It was found that the CNN model had the best performance, with an F1 score of 0.79. Pinkesh Badjatiya and others compared the performance of three neural network architectures, CNN, LSTM, and Fast-Text, in the task of hate speech classification. The results showed that the model combining LSTM with Random Embedding and GBDT worked best[24].
2.2 Multimodal hate speech detection
Model fusion is a technique that combines multiple machine learning or deep learning models to improve the performance of the model. Such methods can be applied across various deep learning domains, including but not limited to computer vision, natural language processing, and speech recognition. The fusion of deep learning can solve the limitations of a single deep model to a certain extent and improve the generalization ability and performance of the model. Tan et al. used model fusion technology to improve the performance of convolutional neural networks [25]. Multimodal feature processing methods that exist tend to fall into two primary architectural categories: early fusion and late fusion. Early fusion refers to the integration that occurs after the features are extracted, with the key advantage being the ability to leverage correlations between multiple features from different modalities at an early stage. Late fusion, also known as decision-level fusion, refers to the process of integration that takes place after each modality has made its individual decision, whether it be classification or regression.
Sai et al. used BERT for extracting text features and Inception v3, Inception ResNet, ResNext for extracting image features. They applied various early fusion techniques, such as concatenation and product rule, as well as late fusion techniques including distribution summation, performance weighting, logarithmic opinion polling, and rules learned from training on probabilities, to effectively integrate visual and textual modalities [26]. Liang Yi et al. were the first to employ a multi-layer CNN and a simple splicing method to fuse multimodal features locally and globally. They input these two fusion features into different classifiers to obtain two probability distributions, and subsequently processed the information through a hybrid fusion mode of late fusion by proportionally adding the two probability distributions [27]. Chuanpeng Yang et al. proposed an extensible cross domain knowledge transfer (CDKT) framework, which can flexibly use mainstream visual language converters as the backbone [28]. Anusha Chhabra et al. introduced a "Multi-Scale Kernel Attractive Visual" (MSKAV) module, which can use an efficient multibranch structure to extract distinctive visual features and perform post fusion by splicing the extracted multimodal features [29]. Zheng Wenbo et al. embedded the text of multimodal knowledge map attention image into the doctor-patient dialogues, achieving promising results [30]. Suyash Sangwan et al. used RNN and VGG-16 to extract text and image features, and captured context information through a two-way gated loop unit to process multimodal data [31].
Research shows that although progress has been made in the detection of multimodal hate speech, there is still a need for further investigation to address existing problems and challenges. To our knowledge, previous studies have not employed the self-attention mechanism for the early fusion of image and text modal features, which introduces a novel aspect to the methodology discussed in Section 3. We believe that the application of self-attention mechanism in this scenario has obvious advantages. It can capture more profound correlations between modalities, which is particularly critical for understanding and analyzing hate speech that contains delicate emotions and implicit meanings. Compared with the traditional feature fusion technology, the self-attention mechanism provides a more flexible and dynamic way to dynamically weight different modalities’ information, to better reflect the complementarity and importance of different modalities.
In the proposed method, we combine text and image bimodal features through a self-attention mechanism for hate speech detection. This method makes full use of the complementary effects of text and images to improve the performance of the model on hate speech detection tasks to a certain extent. This section will introduce our text and image processing methods and the method of applying multi-head self-attention mechanism for feature fusion respectively.
3.1 Text Modal Processing
RoBERTa, which stands for A Robustly Optimized BERT Pretraining Approach[32], is a natural language processing (NLP) pretraining model in the field of natural language processing (NLP) launched by Facebook AI in 2019. As compared to BERT (Bidirectional Encoder Representations from Transformers), RoBERTa utilizes a larger corpus of text data for pretraining. This enhances the model's ability to generalize. Moreover, RoBERTa incorporates a dynamic masking strategy that accommodates varying sentence lengths throughout the training process. This enables the model to capture a richer set of contextual information embedded within the text. In this experiment, RoBERTa is employed to transform hate speech texts, which are extracted via OCR technology, into word vectors that serve as the text feature inputs, as shown in Eq. (1), where \(\:T\) is the input text content and \(\:{V}_{txt}\in\:{\mathbb{R}}^{n}\) is the n-dimensional text feature vector.
$$\:\begin{array}{c}{V}_{txt}=RoBERTa\left(OCR\left(T\right)\right)\#\left(1\right)\end{array}$$
3.2 Image Model Processing
3.2.1 DeepFillV2
DeepFillV2 is a deep learning-based image inpainting technique aimed at repairing damaged, missing, or corrupted regions within an image by automatically filling in the missing parts [33]. Its applications span image inpainting, photo enhancement, and other related fields. In the context of hate speech memes, characteristics such as the text's location, length, and size can influence the model's feature extraction process for the image. This paper adopts the approach presented in [34]. Initially, EasyOCR recognition technology is utilized to detect and extract the text's location information, followed by the application of a segmentation mask. Subsequently, DeepFillV2 is employed to remove the text content from the meme image, resulting in clean, hateful memes without text.
3.2.2 Swin Transformer V2
Swin Transformer V2 is a deep learning model based on the self-attention mechanism launched by Microsoft Research Asia at the end of 2021. As the largest dense vision model to date, it boasts 3 billion parameters and can be trained on images with resolutions up to 1536×1536 pixels [35]. This model draws inspiration from the Vision Transformer's approach to processing images, utilizing a 'res post' normalization in place of the 'pre' normalization configuration. It has set new performance benchmarks in four key visual tasks: ImageNet-V2 image classification, COCO object detection, ADE20K semantic segmentation, and Kinetics-400 video action classification.
3.2.3 Image Feature Extraction
Given that an image may contain text and other elements that could influence the model's recognition capabilities, it is essential to initially eliminate the text from the image data. To begin with, an image is inputted, and the image size is standardized to 224 × 224 pixels, as shown in Equations (2).
$$\:\begin{array}{c}{I}_{std}=resize\left(I,224\times\:224\right)\#\left(2\right)\end{array}$$
Next, assess whether the image currently being processed contains text content. If text is present, proceed to input the image into the DeepFillV2 model for recognition and processing. Subsequently, remove the text-containing area from the image and perform the necessary repairs. Thereafter, utilize the Swin Transformer V2 model to extract features from the image that has been fixed. This process is illustrated in Equations (3), where \(\:{V}_{img}\in\:{\mathbb{R}}^{m}\) is the m-dimensional image feature vector.
$$\:\begin{array}{c}{V}_{img}=SwinTransformerV2\left({I}_{std}\right)\#\left(3\right)\end{array}$$
3.3 Model Fusion
The Multi-head Self-Attention mechanism [36] is widely utilized in the fields of natural language processing and neural networks as an attention mechanism. It enables a model to concurrently focus on information across various positions within the input sequence, thereby enhancing the capture of long-distance dependencies and associations. Facilitated by this mechanism, our model fusion strategy is depicted in Fig. 2.
The multi-head self-attention mechanism facilitates the fusion of text and image feature vectors. Specifically, the text and image features are first transformed into feature vectors, denoted as \(\:{V}_{img}^{{\prime\:}}\) and \(\:{V}_{txt}^{{\prime\:}}\) respectively, with matching dimensions through a dimensionality transformation. These vectors are then fed into the multi-head self-attention module, which calculates the self-attention weights that quantify the importance of each text position in relation to the image positions, as depicted in Eq. (4).
$$\:\begin{array}{c}A=MultiHeadSelfAttention\left({V}_{img}^{{\prime\:}},{V}_{txt}^{{\prime\:}}\right)\#\left(4\right)\end{array}$$
Subsequently, the text features are integrated with these weights through a dot product operation, as illustrated in Eq. (5), to derive \(\:{V}_{txt}^{weight}\). This step emphasizes the textual components that exhibit a high correlation with the image features. The aim is to accord greater focus to the text information pertinent to the image during the fusion of image and text features, thereby enhancing the model's classification performance.
$$\:\begin{array}{c}{V}_{txt}^{weight}=A\bullet\:{V}_{txt}^{{\prime\:}}\#\left(5\right)\end{array}$$
Finally, the obtained text and image feature vectors are combined by splicing to form a new vector, \(\:{V}_{fusion}\), representing the fused feature of both image and text. This image-text fusion feature vector is then passed through a fully connected layer and a softmax layer to reach the output layer of the model, yielding the final classification results.
This paper conduct classification experiments using a multimodal hate speech dataset. In this section, we provide a brief overview of the experimental dataset, the various evaluation metrics used to assess the classification outcomes, the configuration of the experimental environment, and the settings for hyperparameters.
4.1 Datasets
Hateful memes dataset [37] is a multimodal hate speech dataset that was developed by Facebook AI for the Hateful Memes Challenge. Comprising 10,000 memes, this dataset includes content posted on public social media platforms in the United States. It encompasses five distinct categories: multimodal hate, unimodal hate, benign images, benign text, and non-hateful random content. Memes with hateful content constitute 37% of the total dataset.
4.2 Classification metrics
Referencing the evaluation indicators proposed by Chhabra et al. for HSD tasks [38], this study employs a set of five distinct performance metrics, including accuracy, recall, F1 Score, and AUCROC, to conduct a comprehensive evaluation of the model's performance. Table 1 presents the general formulation and the range of values for these performance metrics. An explanation of each performance metric is provided below.
| Metrics | Formula | Range |
|---|---|---|
| Accuracy | \(\:\frac{TP+TN}{TP+TN+FP+FN}\) | \(\:\left[\text{0,1}\right]\) |
| Precision | \(\:\frac{TP}{TP+FP}\) | \(\:\left[\text{0,1}\right]\) |
| Recall | \(\:\frac{TP}{TP+FN}\) | \(\:\left[\text{0,1}\right]\) |
| F1-Score | \(\:\frac{\left(2\times\:Recall\times\:Prcision\right)}{Recall+Precision}\) | \(\:\left[\text{0,1}\right]\) |
| AUCROC | \(\:\int\:ROC\) | \(\:\left[\text{0,1}\right]\) |
1) Accuracy: Accuracy is a fundamental performance metric that measures the proportion of samples correctly classified by the model. It offers an overall assessment of the model's performance.
2) Precision: Precision evaluates the model's accuracy in predicting the positive class, that is, the proportion of actual positive samples among all samples the model has predicted as positive.
3) Recall: Recall measures the model's ability to identify all positive class samples, which is the proportion of correctly predicted positive samples out of all actual positive samples.
4) F1-Score: The F1-Score is the harmonic mean of precision and recall. It provides a comprehensive performance metric that balances the trade-off between precision and recall.
5) AUC-ROC (Area Under the Receiver Operating Characteristic Curve): AUC-ROC is an indicator used to measure the performance of a binary classification model. It is based on the area under the curve that plots the true positive rate against the false positive rate, offering a way to evaluate the model's performance across various thresholds.
4.3 Experimental environment configuration
The configuration of our experimental environment is shown in Table 2.
| Name | Version |
|---|---|
| operating system | Ubuntu 20.04.3 |
| CPU | Intel(R) Xeon(R) Gold 6240R CPU @ 2.40GHz |
| GPU | NVIDIA A100 80GB |
| Pytorch | 1.13.1 + cu116 |
4.4 Hyper parameter specification
In our experiments, we set the patch size for the Swin Transformer to 4x4 pixels, the dimension of Embedding Layer to 128, the depth layer to [2, 2, 18, 2]. For the number of attention heads used in the multi-head attention mechanism, we set it to [4, 8, 16, 32], and the window size for moving is set to [8, 8]. To mitigate the overfitting problem that deep networks are prone to, the idea of random depth pruning is adopted, and the probability of closing each residual block is set to 0.5. In addition, we used AdamW optimizer with a learning rate of 1e-5 and set the batch size to 64. These parameters were tuned during the experiment, and we ran 30 epochs for training and selected the best model parameters based on the experimental results.
This section will present the accuracy, recall, F1 Score, and AUCROC of our proposed multimodal joint model on the dataset and demonstrate the effectiveness of this model by comparing its classification performance with those of single-mode models and the results from previous research.
5.1 Single-modality Comparison
By evaluating multiple models across different visual and textual modalities, our goal is to identify the top-performing models in each modality for subsequent combination efforts. Analyzing the performance metrics of various models presented in Table 3, we find that in the textual modality, XLNet performs notably in recall (0.6789). Meanwhile, RoBERTa stands out in the pure text hate speech classification task, achieving the highest scores for accuracy (0.7250), precision (0.6909), F1-Score (0.6247), and AUCROC (0.6879). In the visual modality, ResNet shows superior performance in accuracy (0.8610) and precision (0.9763), Swin Transformer has the best recall (0.6789), and the second-generation Swin Transformer excels in F1-Score (0.7846) and AUCROC (0.8234).
Based on the overall performance of each model, we select RoBERTa and Swin Transformer, which demonstrate excellence in their respective domains, for model fusion. This selection ensures that the combined model has the potential to achieve optimal performance.
|
Modal |
Model |
Accuracy |
Precision |
Recall |
F1-Score |
AUCROC |
|---|---|---|---|---|---|---|
|
Text |
XLNet |
0.6900 |
0.5785 |
0.6789 |
0.5802 |
0.6815 |
|
ALBERT |
0.7020 |
0.6331 |
0.5132 |
0.5669 |
0.6654 |
|
|
BERT |
0.6990 |
0.6082 |
0.5842 |
0.5960 |
0.6768 |
|
|
RoBERTa |
0.7250 |
0.6909 |
0.5237 |
0.6247 |
0.6879 |
|
|
Image |
ViT |
0.8481 |
0.9713 |
0.6236 |
0.7596 |
0.8061 |
|
ResNet |
0.8610 |
0.9763 |
0.6500 |
0.7804 |
0.8202 |
|
|
Swin Transformer |
0.8520 |
0.9084 |
0.6789 |
0.7771 |
0.8185 |
|
|
Swin Transformer V2 |
0.8600 |
0.9444 |
0.6711 |
0.7846 |
0.8234 |
5.2 Baseline Comparison
Our research provides a detailed comparison of the performance of the proposed model with current leading methods on the Hateful Memes dataset. Visual BERT COCO [37] is based on the BERT model structure, utilizes a stacked Transformer as an encoder for coding, and undergoes pretraining on the COCO dataset. The DisMultiHate [39] model uses BERT and Faster R-CNN to extract text and image features and performs model fusion by concatenation. MSKAV [29]incorporates multiscale convolutional kernels, a multi-branch architecture, and visual attention mechanisms for image feature extraction, and uses DiscolBERT, which is based on knowledge distillation, for text feature extraction. The feature fusion in this model is carried out using a serial approach.
|
Model |
Accuracy |
Precision |
Recall |
F1-Score |
AUCROC |
|---|---|---|---|---|---|
|
Visual BERT COCO |
0.6947 |
- |
- |
- |
0.7544 |
|
DisMultiHate |
0.7580 |
- |
- |
- |
0.8280 |
|
MSKAV |
0.8750 |
0.8333 |
- |
0.6950 |
0.8383 |
|
Our Model |
0.8780 |
0.9135 |
0.7500 |
0.8237 |
0.8532 |
Table 4 presents the results indicating that our proposed multimodal hate speech joint model, which integrates multi-level visual features using moving windows and employs text feature extraction based on the RoBERTa pretraining model through a multi-head self-attention mechanism, has achieved notable performance improvements on the Hateful Memes dataset. Specifically, there was an enhancement of 0.3% in accuracy and 8.02% in precision. Moreover, our model showed a significant increase of 12.87% in the F1-Score metric, suggesting that our approach has bettered the balance between accuracy and recall for the task of detecting hateful memes. In addition, the AUCROC metric saw an improvement of 1.49%, demonstrating our model's enhanced capability to distinguish between positive and negative samples in the classification of hateful memes.
Table 5 provides an example of the model's judgment results concerning the presence of hate factors in the test set data. The text within the images is accurately recognized and extracted, and the areas originally covered by the text are effectively eliminated from the images. The classification results demonstrate that the model has learned to discern the implicit associations between text and image. Particularly with the image captioned 'how to fix a noisy dishwasher', the model exhibits advanced cognitive capabilities in understanding the relationship between text and image content. It identifies the discrepancy between the potentially misleading text and the innocuous nature of the image itself, indicating that the model transcends mere surface-level recognition to interpret the underlying emotions and intentions.
|
Image |
Fixed Image |
Text |
Result |
|---|---|---|---|
|
|
|
happy pride month let's go beat up lesbians |
hate |
|
|
|
how to fix a noisy dishwasher |
hate |
|
|
|
organic vegetable |
no hate |
5.3 Ablation Trials
This section presents the results of ablation experiments conducted to assess the effectiveness of the proposed model on the Hateful Memes dataset. Within the framework of the proposed multimodal joint model, three distinct ablation studies were carried out. The first study modeled and analyzed the textual component exclusively. The second investigation was dedicated to confirming the model's efficacy with respect to the image component alone. The third experiment entailed the removal of the self-attention mechanism used for feature fusion. The outcomes of these ablation experiments are detailed in Table 6. Based on the observations from the ablation experiments, the following conclusions can be drawn:
1) Our proposed multimodal hate speech detection joint model performs best on the Hateful Memes dataset.
2) Removing any modality leads to a deterioration in the model's predictive performance.
3) Using multi-head self-attention mechanism for fusing textual and visual features effectively enhances the overall performance of the model.
|
Model |
Accuracy |
Precision |
Recall |
F1-Score |
AUCROC |
|---|---|---|---|---|---|
|
RoBERTa |
0.7250 |
0.6909 |
0.5237 |
0.5802 |
0.6815 |
|
Swin Transformer V2 |
0.8600 |
0.9444 |
0.6711 |
0.7846 |
0.8234 |
|
Model/no attention |
0.8620 |
0.8104 |
0.8477 |
0.7763 |
0.8454 |
|
Our Model |
0.8780 |
0.9135 |
0.7500 |
0.8237 |
0.8532 |
The results of the ablation experiment compellingly demonstrate that our multimodal hate speech joint model comprehends content from a broader perspective compared to single-modal models. In cases involving ambiguity or wordplay, the model can more precisely discern the true intent and emotional tone of the speech by integrating visual and linguistic cues. Moreover, when encountering noise or misleading information in one modality, the multimodal model leverages information from another modality for correction, thereby enhancing its robustness. The self-attention mechanism employed by the model effectively captures long-distance dependencies and dynamically adjusts the attention weights for different modal information. This integration of multimodal information is more effective, rendering the fusion process more agile and efficient.
5.4 Engineering Applications
In this section, we discuss the potential value and practical feasibility of the proposed multimodal hate speech detection joint model. Initially, our model demonstrates strong performance in processing the Hateful Memes dataset, achieving significant improvements over current leading methods. This underscores its capability to accurately recognize and classify texts and images with hate factors, providing robust support for its application across social media platforms, news websites, online communities, and similar contexts. Secondly, by comparing single-modal and multimodal results and the outcomes of ablation experiments, we have confirmed the model's advantages in the integrated use of textual and visual information. This is particularly significant for scenarios requiring the synthesis of diverse data sources for decision-making, such as social media content moderation, news report authenticity verification, and online community management [40]. I Furthermore, in application contexts like the medical field [41], where understanding model decision-making is critical, especially in sensitive domains where explanations for model decisions are necessary, our model employs a self-attention mechanism for feature fusion. This not only enhances the model's overall performance but also augments its interpretability.
Ultimately, our model exhibits a notable degree of flexibility and adaptability for engineering applications. By fine-tuning the model parameters further, we can more precisely capture domain-specific features and patterns throughout the iterative testing process, thus enhancing its application efficacy in targeted fields. In practical scenarios, integrating domain experts' knowledge and feedback, coupled with thorough analytical insights, can lead to a more refined model functionality, and expanded application scope. Overall, our multimodal hate speech detection joint model presents extensive potential for application in engineering contexts, particularly in managing the intricacies of social media and online content. It holds promise in contributing to the creation of a more secure and inclusive digital society.
This article presents a novel multimodal hate speech detection model that integrates RoBERTa for processing the text modality and Swin Transformer V2 for extracting image features using a multi-head self-attention mechanism. The model underwent training and testing for generalization on a public multimodal hate speech dataset. The experimental outcomes demonstrate that the proposed architecture surpasses baseline models, yielding favorable results across metrics such as accuracy, precision, recall, F1-Score, and AUC-ROC. Additionally, we performed ablation experiments on the dataset, which confirmed that the proposed multimodal joint model significantly enhances the detection performance of multimodal hate speech. In summary, our research introduces a novel multimodal approach to hate speech detection, utilizing attention mechanisms to fuse textual and visual information and thereby boost model performance. The insights from this study hold significant implications for the deployment of multimodal hate speech detection systems and the moderation of online social environments across social media and content platforms.
Although our proposed model has yielded promising results in experimental settings, we acknowledge its limitations, particularly regarding interpretability. The current model's decision-making process may lack transparency for users, constraining its effectiveness and trustworthiness in certain application scenarios. Consequently, a key direction for future research is to investigate and develop advanced interpretive AI technologies designed to enhance model transparency and interpretability. Doing so will not only enable users to comprehend the model's decision-making logic more effectively, but also broaden its acceptance and application scope.
Author Contribution
Mao wrote the main manuscript text and Shi implemented the image model processing and related experiments. Li prepared figures 1-2 and table 1-5. All authors reviewed the manuscript.
- Gomez R, Gibert J, Gomez L, Karatzas D (2020) Exploring Hate Speech Detection in Multimodal Publications. In: 2020 IEEE Winter Conference on Applications of Computer Vision (WACV), pp. 1459–1467. IEEE, Snowmass Village, CO, USA https://doi.org/10.1109/WACV45572.2020.9093414
- Guterres A (2019) United Nations strategy and plan of action on hate speech https://www.un.org/en/hate-speech/un-strategy-and-plan-of-action-on-hate-speech
- Mora D (2022) New un study finds that hate speech in costa rica grew by 71 https://unsdg.un.org/latest/stories/new-un-study-finds-hate-speech-costa-rica-grew-71
- Yadav A, Chandel S, Chatufale S, Bandhakavi A (2023) Lahm: Large annotated dataset for multi-domain and multilingual hate speech identification. arXiv preprint arXiv :230400913
- Leite JA, Scarton C, Silva DF (2023) Noisy Self-Training with Data Augmentations for Offensive and Hate Speech Detection Tasks. arXiv
- Tontodimamma A, Nissi E, Sarra A, Fontanella L (2021) Thirty years of research into hate speech: topics of interest and their evolution. Scientometrics 126:157–179. https://doi.org/10.1007/s11192-020-03737-6
- Wang J, Wang S, Lin M, Xu Z, Guo W (2023) Learning speaker-independent multimodal representation for sentiment analysis. Inf Sci 628:208–225. https://doi.org/10.1016/j.ins.2023.01.116
- Capozzi Arthur T, Mirko L, Valerio B, Fabio P, Sanguinetti M, Cristina B, Viviana P, Giancarlo R, Cataldo M, Marco P et al (2019) : Computational linguistics against hate: Hate speech detection and visualization on social media in the contro l’odio project. In: CEUR Workshop Proceedings, vol. 2481 CEUR-WS
- Basile V, Bosco C, Fersini E, Nozza D, Patti V, Pardo FMR, Rosso P, Sanguinetti M (2019) Semeval-2019 task 5: Multilingual detection of hate speech against immigrants and women in twitter. In: Proceedings of the 13th International Workshop on Semantic Evaluation, pp. 54–63
- Florio K, Basile V, Lai M, Patti V (2019) Leveraging Hate Speech Detection to Investigate Immigration-related Phenomena in Italy. In: 2019 8th International Conference on Affective Computing and Intelligent Interaction Workshops and Demos (ACIIW), pp. 1–7. IEEE, Cambridge, United Kingdom https://doi.org/10.1109/ACIIW.2019.8925079
- S¸ahi H, Kılı¸c Y, Saˇglam RB (2018) Automated detection of hate speech towards woman on twitter. In: 2018 3rd International Conference on Computer Science and Engineering (UBMK), pp. 533–536 https://doi.org/10.1109/UBMK.2018.8566304
- Ahluwalia R, Soni H, Callow E, Nascimento A, De Cock M (2018) Detecting hate speech against women in english tweets. EVALITA Evaluation of NLP and Speech Tools for Italian. 12:194. https://doi.org/10.4000/books.aaccademia.4421
- Mozafari M, Farahbakhsh R, Crespi N (2020) Hate speech detection and racial bias mitigation in social media based on BERT model. PLoS ONE 15(8):0237861. https://doi.org/10.1371/journal.pone.0237861
- Gamb¨ack B, Sikdar UK (2017) Using convolutional neural networks to classify hate speech. In: Waseem, Z., Chung, W.H.K., Hovy, D., Tetreault, J. (eds.) Proceedings of the First Workshop on Abusive Language Online, pp. 85–90. Association for Computational Linguistics, Vancouver, BC, Canada https://doi.org/10.18653/v1/W17-3013. https://aclanthology.org/W17-3013
- Uyheng J, Bellutta D, Carley KM (2022) Bots Amplify and Redirect Hate Speech in Online Discourse About Racism During the COVID-19 Pandemic. Social Media + Soc 8(3):205630512211047. https://doi.org/10.1177/20563051221104749
- Plaza-del-Arco FM, Molina-Gonz´alez MD, Ure˜na-L´opez LA, Mart´ın Valdivia MT (2021) Comparing pre-trained language models for Spanish hate speech detection. Expert Syst Appl 166:114120. https://doi.org/10.1016/j.eswa.2020.114120
- Djuric N, Zhou J, Morris R, Grbovic M, Radosavljevic V, Bhamidipati N (2015) Hate Speech Detection with Comment Embeddings. In: Proceedings of the 24th International Conference on World Wide Web, pp. 29–30. ACM, Florence Italy https://doi.org/10.1145/2740908.2742760
- Saha P, Mathew B, Goyal P, Mukherjee A (2018) Hateminers: Detecting hate speech against women. arXiv preprint arXiv :181206700
- Akuma S, Lubem T, Adom IT (2022) Comparing bag of words and tf-idf with different models for hate speech detection from live tweets. Int J Inform Technol 14(7):3629–3635. https://doi.org/10.1007/s41870-022-01096-4
- Alkomah F, Ma X (2022) A Literature Review of Textual Hate Speech Detection Methods and Datasets. Information 13(6):273. https://doi.org/10.3390/info13060273
- MacAvaney S, Yao H-R, Yang E, Russell K, Goharian N, Frieder O (2019) Hate speech detection: Challenges and solutions. PLoS ONE 14(8):0221152. https://doi.org/10.1371/journal.pone.0221152
- Gamb¨ack B, Sikdar UK (2017) Using Convolutional Neural Networks to Classify Hate-Speech. In: Proceedings of the First Workshop on Abusive Language Online, pp. 85–90. Association for Computational Linguistics, Vancouver, BC, Canada https://doi.org/10.18653/v1/W17-3013
- Alshalan R, Al-Khalifa H (2020) A deep learning approach for automatic hate speech detection in the Saudi Twittersphere. Appl Sci 10(23):8614. https://doi.org/10.3390/app10238614
- Badjatiya P, Gupta S, Gupta M, Varma V (2017) Deep learning for hate speech detection in tweets. In: Proceedings of the 26th International Conference on World Wide Web Companion - WWW ’17 Companion, pp. 759–760 https://doi.org/10.1145/3041021.3054223
- Tan M, Le Q (2019) Efficientnet: Rethinking model scaling for convolutional neural networks. In: International Conference on Machine Learning, pp. 6105–6114 PMLR
- Sai S, Srivastava ND, Sharma Y (2022) Explorative application of fusion techniques for multimodal hate speech detection. SN Comput Sci 3(2):122. https://doi.org/10.1007/s42979-021-01007-7
- Liang Yi T, Tuerdi T (2023) Multimodal fake information detection using multi-layer cnn feature fusion and multi-classifier ensemble prediction. Comput Eng Science/Jisuanji Gongcheng yu Kexue 45(6)
- Yang C, Zhu F, Liu G, Han J, Hu S (2022) Multimodal Hate Speech Detection via Cross-Domain Knowledge Transfer. In: Proceedings of the 30th ACM International Conference on Multimedia, pp. 4505–4514. ACM, Lisboa Portugal https://doi.org/10.1145/3503161.3548255
- Chhabra A, Vishwakarma DK (2023) Multimodal hate speech detection via multiscale visual kernels and knowledge distillation architecture. Eng Appl Artif Intell 126:106991. https://doi.org/10.1016/j.engappai.2023.106991
- Zheng W, Yan L, Gou C, Zhang Z-C, Zhang J, Hu J, Wang M (2021) Pay attention to doctor–patient dialogues: Multi-modal knowledge graph attention image-text embedding for COVID-19 diagnosis. Inform Fusion 75:168–185. https://doi.org/10.1016/j.inffus.2021.05.015
- Sangwan S, Akhtar MS, Behera P, Ekbal A (2020) I didn’t mean what I wrote! Exploring Multimodality for Sarcasm Detection. In: 2020 International Joint Conference on Neural Networks (IJCNN), pp. 1–8. IEEE, Glasgow, United Kingdom https://doi.org/10.1109/IJCNN48605.2020.9206905
- Liu Y, Ott M, Goyal N, Du J, Joshi M, Chen D, Levy O, Lewis M, Zettlemoyer L, Stoyanov V (2019) RoBERTa: A Robustly Optimized BERT Pretraining Approach. arXiv
- Yu J, Lin Z, Yang J, Shen X, Lu X, Huang TS (2019) Free-form image inpainting 17with gated convolution. In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 4471–4480 https://doi.org/10.1109/ICCV. 2019.00457
- Zhu R (2020) Enhance Multimodal Transformer With External Label And In-Domain Pretrain: Hateful Meme Challenge Winning Solution. arXiv
- Liu Z, Lin Y, Cao Y, Hu H, Wei Y, Zhang Z, Lin S, Guo B (2021) Swin Transformer: Hierarchical Vision Transformer using Shifted Windows. In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pp. 9992–10002. IEEE, Montreal, QC, Canada https://doi.org/10.1109/ICCV48922.2021.00986
- Vaswani A, Shazeer N, Parmar N, Uszkoreit J, Jones L, Gomez AN, Kaiser L, Polosukhin I (2017) Attention is all you need. Adv Neural Inf Process Syst 30
- Kiela D, Firooz H, Mohan A, Goswami V, Singh A, Ringshia P, Testuggine D (2020) The hateful memes challenge: Detecting hate speech in multimodal memes. Adv Neural Inf Process Syst 33:2611–2624
- Chhabra A, Vishwakarma DK (2023) A literature survey on multimodal and multilingual automatic hate speech identification. Multimedia Syst 29(3):1203–1230
- Lee RK-W, Cao R, Fan Z, Jiang J, Chong W-H (2021) Disentangling hate in online memes. In: Proceedings of the 29th ACM International Conference on Multimedia, pp. 5138–5147
- Mena P (2020) Cleaning up social media: The effect of warning labels on likelihood of sharing false news on facebook. Policy internet 12(2):165–183
- Stiglic G, Kocbek P, Fijacko N, Zitnik M, Verbert K, Cilar L (2020) Interpretability of machine learning-based prediction models in healthcare. WIREs Data Min Knowl Discov 10(5):1379. https://doi.org/10.1002/widm.1379
No competing interests reported.