TL;DR
This paper explores the detection of hate speech in cinema by introducing a novel dataset of 11,000 annotated movie subtitles. The researchers demonstrate that while models trained on social media can adapt to the cinematic domain, fine-tuning transformer-based models like BERT yields the best results, achieving a macro-F1 score of up to 89%.
What Problem Does This Paper Address?
While hate speech detection is a mature field in the context of social media platforms like Twitter and news comment sections, the world of cinema has been largely overlooked. Movies are a primary vehicle for cultural exchange, yet they often contain simulated hate speech and offensive language that can be difficult for automated systems to distinguish.
The core challenge lies in the nuance: distinguishing between hate speech (language intended to denigrate a group based on identity like race or religion) and offensive language (profanity or rudeness that isn't necessarily targeted). In movies, where dialogue is often stylized or context-dependent, this distinction becomes even more blurred. The authors aim to bridge this gap by creating a specialized dataset and testing how well existing models transfer from social media to the silver screen.
Methods
The researchers evaluated three distinct model architectures:
- Bag of Words (BoW): A classic baseline that treats text as an unordered collection of words.
- Bi-directional LSTM (Bi-LSTM): A recurrent neural network designed to capture sequential dependencies in text.
- BERT (Bidirectional Encoder Representations from Transformers): A state-of-the-art transformer model that excels at understanding context.
They employed two primary strategies for transfer learning:
- Domain Adaptation: Training models on social media datasets and testing them directly on movie subtitles.
- Fine-tuning: Taking models pre-trained on social media and further training them on a portion of the movie subtitle dataset to specialize them for cinematic dialogue.
To ensure high-quality data, they used Amazon Mechanical Turk (MTurk) for annotation, employing a rigorous pilot study to select the most accurate workers.
Datasets
The study utilized three main data sources:
- Twitter Dataset: ~24,500 tweets labeled as hate, offensive, or normal.
- Fox News Dataset: ~1,500 comments from news threads, primarily distinguishing between hate and normal speech.
- Movie Subtitle Dataset (New): The authors introduced a new benchmark consisting of 10,688 utterances from six movies.
The movies were strategically chosen based on IMDB tags to ensure a mix of content:
- Racism/Hate-speech tags: Django Unchained, BlacKkKlansman, and American History X.
- Friendship tags (as a control): South Park: Bigger, Longer & Uncut and The Wolf of Wall Street.
- General Drama: Pulp Fiction.
Key Insights and Ideas
- The "Offensive" Dominance: Models trained on Twitter often over-predict the "offensive" class because social media data is heavily skewed toward profanity. In contrast, 84% of the movie dataset is "normal" speech, creating a significant domain mismatch.
- Sarcasm is the Enemy: A qualitative analysis revealed that the most difficult hate speech to detect is sarcastic or indirect. For example, a line like "It’s just so beautiful. We’re cleansing this country..." contains no explicit slurs but is clearly hateful in context.
- Contextual Subjectivity: Annotators sometimes disagreed on labels based on their interpretation of a scene. The authors noted that terms like "Black Power" were labeled "normal" by humans in specific contexts but flagged as "hate" by models, highlighting the need for deeper contextual understanding.
- Genre Trends: Interestingly, movies tagged with "friendship" actually showed a higher percentage of offensive language (profanity) even as the percentage of actual hate speech decreased.
Key Results
BERT was the clear winner across all experiments.
- Domain Adaptation: When trained on Twitter and tested on movies, BERT achieved a macro-F1 score of 77%.
- Fine-tuning: The best performance came from fine-tuning the Fox News-trained BERT model on the movie data, which boosted the macro-F1 score from 64% to 89%.
- Error Analysis: The majority of misclassifications (69%) occurred between "offensive" and "normal" classes. However, 18% of errors involved confusing "hate" with "normal" speech, which the authors identify as the most critical area for improvement.
Limitations and Future Work
The authors acknowledge several hurdles:
- Text-Only Limitations: Subtitles lack the emotional cues of audio and the visual context of video. Future work should explore multimodal models that process sound and imagery alongside text.
- Subjectivity: Hate speech is inherently subjective, evidenced by the 81% agreement rate among human annotators.
- Data Sparsity: Even in "hateful" movies, actual hate speech is relatively rare (3% of the dataset), making it difficult for models to learn robust patterns without data augmentation.
- Target Identification: The current study focuses on detection; the next step is identifying the specific groups being targeted by the hate speech.
Summary updated