TL;DR
This paper introduces two novel datasets for detecting dehumanizing language, a critical but under-researched area in NLP. By creating these datasets and evaluating baseline models, the authors pave the way for more robust computational tools to identify and combat harmful speech.
What Problem Does This Paper Address?
Dehumanization, the act of portraying individuals or groups as less than human, is a precursor to severe mistreatment and violence. While the concept is well-understood in social sciences, its computational detection remains a significant challenge. Existing work often focuses on broader categories like hate speech, which is a surface phenomenon, rather than the underlying attitudes of dehumanization. A key hurdle is the lack of specialized datasets that capture the nuances of dehumanizing language, which can manifest in subtle ways, making automated detection difficult.
Methods
The authors employ a keyword-based approach to extract potential dehumanizing text from large corpora. They define dehumanization patterns based on established psychological frameworks, including:
- Negative evaluation of a target group: General disparagement.
- Denial of agency: Portraying a group as lacking free will or control.
- Moral disgust: Associating a group with revulsion or contamination.
- Animal metaphors: Comparing individuals to animals.
- Objectification: Treating individuals as inanimate objects.
They then use these extracted texts to train and evaluate models. A baseline model uses valence scores, while a more advanced approach fine-tunes a BERT-based model (HateBERT) on their newly created labeled dataset.
Datasets
The paper introduces two new datasets:
- Unlabelled Corpus: A large dataset (565,304 paragraphs) derived from movie/TV subtitles (OpenSubtitles) and political forum discussions (Common Crawl). This corpus is designed to provide a broad range of dehumanizing language examples.
- Labelled Data Set: A subset of the unlabelled data (918 samples), manually annotated by two independent annotators. This dataset serves as a benchmark for evaluating detection models. The annotation process involved defining criteria and providing example sentences, with annotators choosing between "Yes," "No," or "Not Sure."
Key Insights and Ideas
- Dehumanization vs. Hate Speech: The authors emphasize that dehumanization is an underlying attitude, distinct from the observable "hate speech." This distinction is crucial for developing more targeted detection methods.
- Keyword Extraction Limitations: While effective for initial data collection, keyword-based extraction risks missing relevant data and may not capture all nuances of dehumanization.
- Domain-Specific Knowledge is Key: Detecting dehumanization often requires understanding specific contexts and cultural references, posing a challenge for purely data-driven models. For instance, understanding political satire or historical events is vital for accurate interpretation.
- Animal Metaphors are Prevalent: The analysis of the labelled dataset revealed that animal metaphors are a frequently employed pattern in dehumanizing language.
Key Results
- Model Performance: The fine-tuned HateBERT model achieved an F1 score of 0.5941 on the labelled dataset, outperforming a baseline model (F1: 0.3643) and other pre-trained hate speech models on this specific task.
- Annotation Agreement: Inter-annotator agreement (Krippendorff's alpha) was moderate (around 0.5), indicating the inherent difficulty and subjectivity in labelling dehumanizing language.
- Pattern Distribution: The labelled data showed that animal metaphors (41.8%) and negative target evaluation (29.09%) were the most common patterns of dehumanization.
- Linguistic Associations: The study found significant similarities between sexual keywords and moral disgust terms, and between ethnic keywords and animal metaphors, suggesting linguistic connections in dehumanizing discourse.
Limitations and Future Work
- Data Bias: The datasets are heavily influenced by political discourse, which may limit their generalizability to other domains.
- Language Specificity: The datasets are exclusively in English, and the models trained on them may not perform well in other languages due to grammatical and idiomatic differences.
- Keyword Coverage: The keyword-based extraction method might not capture all relevant terms, potentially leading to an incomplete representation of dehumanizing language.
The authors suggest that future work should focus on developing models that can effectively handle domain-specific knowledge and context to improve detection accuracy. They also highlight the need for more comprehensive datasets that cover a wider range of dehumanizing patterns and linguistic contexts.
Summary updated