Source-linked AI summary
Exploring Hate Speech Detection in Multimodal Publications
Raul Gomez, Jaume Gibert, Lluis Gomez, Dimosthenis Karatzas
TL;DR
The paper addresses hate speech detection in multimodal publications, where text and images jointly provide evidence and offensive language alone is insufficient. It constructs and annotates MMHS150K, evaluates textual, visual, and multimodal models, and analyzes task challenges. Images are useful, but the proposed multimodal models do not outperform textual models, while annotation subjectivity remains an important limitation.
Problem
Hate speech detection must distinguish attacks on communities from offensive language, including cases where the publication’s meaning depends on both text and image.
Method
The paper collects and annotates MMHS150K tweets, then evaluates textual, visual, and multimodal models using jointly analyzed tweet and image information.
Results
Images are useful for hate speech detection, but the proposed multimodal models do not outperform models using textual information alone.
Takeaways & Limitations
The work establishes multimodal hate speech detection as a research task and opens its dataset and challenges for further research.
Takeaways & Limitations
Annotation is subjective and depends on annotators’ convictions and sensitivity, with some workers labeling tweets by spotting slurs rather than interpreting multimodal context.
Abstract
from arXiv · showhide
In this work we target the problem of hate speech detection in multimodal publications formed by a text and an image. We gather and annotate a large scale dataset from Twitter, MMHS150K, and propose different models that jointly analyze textual and visual information for hate speech detection, comparing them with unimodal detection. We provide quantitative and qualitative results and analyze the challenges of the proposed task. We find that, even though images are useful for the hate speech detection task, current multimodal models cannot outperform models analyzing only text. We discuss why and open the field and the dataset for further research.
1. Introduction
The paper frames hate speech detection as distinguishing attacks on communities from merely offensive language, especially when meaning depends on both tweet text and image. It introduces a multimodal task and dataset while reporting that multimodal models do not outperform textual models.
- Problem definition: Hate speech targets individuals or communities because of shared characteristics, whereas offensive terms alone do not establish hate speech.Publications may attack communities without profanity, and some slurs can appear without malicious intent.
- Multimodal motivation: Multimodal tweets can convey hate speech only through the combination of text and image, making contextual interpretation necessary.Authors of hate speech may intentionally make text insufficient to determine the publication’s meaning.
- Contributions: The work proposes hate speech detection in multimodal publications and releases the large-scale MMHS150K dataset.The dataset is formed from tweets containing text and images.
- Contributions: The authors evaluate multimodal models against unimodal detection and study the task’s challenges for future research.The introduction identifies visual context as relevant while setting up comparisons between textual, visual, and multimodal approaches.
2. Related Work
Prior work extensively studies textual hate speech detection, but the field lacks consistent datasets and evaluation protocols. Visual and multimodal research exists but remains limited, often relying on pretrained visual features or multimodal alignment approaches from adjacent tasks.
- Textual hate speech detection: Textual hate speech research covers terminology, features, datasets, and classifiers, yet lacks a consistent dataset and evaluation protocol.Prior studies include manually annotated Twitter and forum datasets and comparisons among conventional classifiers.
- Textual hate speech detection: Existing textual datasets differ in domains, labels, and sizes, including hate-versus-non-hate, hate/offensive/neither, and racism or sexism categories.Examples range from 2,435 Refugees and Muslims tweets to 27,330 semi-supervised Twitter tweets.
- Visual information: Few studies exploit visual information in online harassment-related classification, and pretrained CNN features have not necessarily improved performance.Related work combines images with captions or comments, but visual processing may remain limited to extracted CNN features.
- Multimodal analysis: Multimodal visual-textual analysis commonly learns aligned feature spaces with CNNs and RNNs or merges modalities for decisions such as visual question answering.These adjacent approaches provide methodological context for combining image and text representations.
3. The MMHS150K dataset
MMHS150K is a large manually annotated multimodal hate-speech dataset created because existing datasets lacked visual information and a stable benchmark. Its construction emphasizes multimodal context, annotation disagreement, and the challenge of separating attacks on communities from offensive wording.
- Dataset construction: 150,000 tweets containing text and images form MMHS150K, created to address the lack of multimodal hate-speech datasets and a reference benchmark.The dataset was gathered from Twitter and made available online.
- Dataset construction: Tweets were filtered for Hatebase terms, excluded retweets and short or porn-related posts, and retained only when images were available.Tweets containing only textual images, such as screenshots, were removed to preserve visual and textual information.
- Annotation: Each tweet was shown with its text and image to three Mechanical Turk workers, who assigned one of six attack categories.The categories included no attack, racist, sexist, homophobic, religion-based, and other community attacks.
- Annotation: Annotation is subjective and depends on workers’ convictions and sensitivity, while stronger attacks are expected to receive cleaner labels.The authors excluded responses made in under three seconds because multimodal context requires more time to assess.
- Annotation: 112,845 tweets were labeled not hate and 36,978 hate, with balanced validation and test sets of 5,000 and 10,000 tweets.The hate labels included racist, sexist, homophobic, religion-based, and other hate categories, but experiments used only the hate/not-hate split.
- Annotation: Using annotator-derived hate scores instead of binary labels produced no significant differences from the reported experimental results.The raw annotations were nevertheless published for further research.
- Dataset significance: MMHS150K is described as the biggest hate-speech dataset to date and the first multimodal hate-speech dataset known to the authors.A central challenge is distinguishing tweets that use the same offensive keywords but do or do not attack a community.
4.1. Unimodal Treatment
The unimodal treatment uses image processing and recurrent text processing to represent multimodal tweet content. Image features come from a fine-tuned Inception v3, while text and image-embedded text are encoded with an LSTM.
- Image processing: Images are resized with a 500-pixel shortest side, then augmented by random 299 × 299 crops and mirroring during training.
- Image processing: A CNN based on ImageNet-pre-trained Google Inception v3 extracts image features and is fine-tuned for hate-speech detection.Fine-tuning adapts its weights to features useful when combined with textual information.
- Text processing: A single-layer LSTM with a 150-dimensional hidden state classifies hate versus not hate from 100-dimensional GloVe word representations.The embeddings were pretrained on two billion tweets to cover Twitter slang and related vocabulary.
- Text processing: Text detected inside images is extracted with Google Vision API and supplied separately from tweet text to preserve their distinct information.The image text is encoded by the LSTM using the hidden state after its last word.
4.2. Multimodal Architectures
The multimodal models combine tweet text, image text, and visual features through concatenation, spatial fusion, or textual kernels. These CNN+RNN architectures are trained to test whether multimodal context improves hate-speech classification.
- Architecture overview: FCM, SCM, and TKM are CNN+RNN models with three inputs: the tweet image, tweet text, and text extracted from the image.
- Feature Concatenation Model: FCM concatenates a 2048-dimensional visual vector with two 150-dimensional LSTM states, producing a 2348-feature representation for fully connected classification layers.
- Spatial Concatenation Model: SCM preserves spatial structure by combining the 8 × 8 × 2048 visual feature map with text vectors at each spatial location.The resulting feature map is processed by two Inception-E blocks before dropout, pooling, and classification.
- Textual Kernels Model: TKM learns text-dependent kernels from tweet text and image text, then convolves them with the visual feature map to model cross-modal interactions.The architecture uses Kt = 10 tweet-text kernels and Kit = 5 image-text kernels.
- Training and analysis: Multimodal models use cross-entropy loss with softmax activations and ADAM, while input ablations test how each modality contributes to classification.
5. Results
Images provide useful signals for hate speech detection, but the proposed multimodal models do not outperform textual models and tend to rely primarily on text. The task remains difficult because annotations are subjective, multimodal relations are complex, and multimodal hate examples are limited.
- Evaluation: The evaluation reports F-score, AUC, and mean accuracy across proposed models, LSTM, a prior method, and random scores under different input settings.Inputs include tweet text, image text, and image.
- Challenges: Subjective judgments and annotator discrepancies make optimal evaluation scores virtually impossible and affect complex image-based and multimodal detection more strongly.A system with relatively low metric scores may still be useful when it fires on publications most annotators agree are hateful.
- Image-only results: Image-only FCM achieves decent results and can distinguish offensive from non-offensive uses of slurs in community contexts.Top-scored hate images include demeaning nudity, racist caricatures, and memes, while top-scored non-hate images include minority portraits.
- Multimodal results: Despite image usefulness, the proposed multimodal models do not improve detection over textual models.Different architectures and training strategies were tested, but the models produced results similar to textual models.
- Multimodal results: The multimodal models eventually use almost only the text input for prediction, including models such as TKM that perform well in other tasks.The authors therefore analyze why these models do not perform well for this task and dataset.
- Challenges: Complex and diverse multimodal relations, background knowledge, and a small multimodal-hate subset make the required relations difficult for neural networks to learn.Although the dataset contains 150K tweets, the multimodal hate subset remains too small for the complex relations needed to identify multimodal hate.
6. Conclusions
The work introduces MMHS150K for multimodal hate-speech detection and finds that textual models outperform multimodal models despite images being useful.
- MMHS150K is presented as the biggest available hate-speech dataset and the first composed of multimodal image-and-text tweets.
- The authors train textual, visual, and multimodal models on MMHS150K.
- Despite images being useful for hate-speech detection, multimodal models do not outperform textual models.
- The work analyzes challenges of the task and dataset and argues for advancing research because most social-media content is multimodal.