TL;DR
This paper benchmarks various fine-tuned language models for detecting offensive and abusive language in Dutch tweets. It highlights the importance of out-of-corpus evaluation and introduces data cartography as a method to identify high-quality training data, finding it beneficial for specific tasks and dataset sizes.
What Problem Does This Paper Address?
Detecting offensive and abusive language on social media is crucial for creating safer online environments. Current methods rely on supervised learning, making the quality of training data paramount. However, standard evaluation methods using held-out test sets can be overly optimistic and fail to reflect real-world performance due to potential biases (annotator, topic, author) and differing data distributions. Evaluating models on out-of-corpus data or using functional tests can provide a more realistic assessment, but these methods are often resource-intensive. This paper addresses the need for robust benchmarking of Dutch offensive and abusive language detection models, exploring the effectiveness of different evaluation strategies and data selection techniques.
Methods
The authors fine-tune a Dutch BERT-based model (BERTje) using three experimental setups:
- Standard (std): Using the entire available training data from the DALC-v2.0 corpus.
- Data Cartography (amb-dim, amb-class): This technique, introduced by Swayamdipta et al. (2020), leverages training dynamics (model confidence and variability across epochs) to identify "ambiguous" training instances. They explore two variants:
amb-dim(selecting ambiguous instances per annotated dimension) andamb-class(selecting ambiguous instances per class). - Random Splits (rdm): Randomly sampled subsets of the training data, used as a control to assess the effectiveness of data cartography.
These fine-tuned models are then evaluated on three distinct test sets: the standard held-out test set from DALC-v2.0, a functional benchmark (HATECHEK-NL), and a dynamic test set (OP-NL).
Datasets
- DALC-v2.0: A Dutch Twitter corpus containing over 11,000 messages annotated for offensive and abusive language, with layers for explicitness and target. For this study, they focus on the explicitness layer, treating offensive and abusive language detection as both binary and multi-class classification tasks.
- HATECHEK-NL: An extension of the Multilingual Hate Check (MHC) benchmark, specifically for Dutch. It comprises 3,835 functional tests designed to evaluate hate speech detection models across various functionalities, including hate expressed via slurs, profanity, and threats.
- OP-NL: A dynamic test set of 1,500 tweets collected in March 2021, mentioning Dutch politicians. It is annotated for offensive language using the same definition as DALC-v2.0.
Key Insights and Ideas
- Out-of-Corpus Evaluation is Crucial: Models performing well on standard held-out test sets can show significantly different (and often worse) performance on out-of-corpus benchmarks like HATECHEK-NL and OP-NL, highlighting the limitations of traditional evaluation.
- Data Cartography's Nuanced Impact: While data cartography showed promise in previous work, its effectiveness on the smaller DALC-v2.0 dataset was mixed. It performed best when selecting ambiguous data per dimension (
amb-dim) and was particularly useful for the functional benchmark (HATECHEK-NL), suggesting it can help identify more robust training subsets. - Offensive vs. Abusive Language: Detecting offensive language in isolation generally yielded better results than jointly detecting offensive and abusive language. This suggests that abusive language is a more specific phenomenon, and models trained on it can be more precise, especially when compared to hate speech definitions.
- HATECHEK-NL Exposes Data Limitations: The functional tests in HATECHEK-NL revealed critical weaknesses in models trained on DALC-v2.0, particularly in identifying non-hateful instances. This suggests that while DALC-v2.0 is valuable, its annotations might not fully capture the nuances required for robust hate speech detection, especially for subtle or non-hateful cases.
- The "Ambiguous" Data Paradox: In some cases, the "ambiguous" data selected via cartography led to overgeneralization of positive classes, potentially at the expense of precision, questioning its universal benefit on smaller datasets.
Key Results
- DALC-v2.0 Performance: The standard training split (
std) generally achieved the best macro-F1 scores on the DALC-v2.0 held-out test set, with scores of 79.93 for offensive, 72.33 for abusive, and 58.90 for the joint classification. - HATECHEK-NL Performance: Models trained on offensive language performed better on HATECHEK-NL's hateful labels but struggled significantly with non-hateful functional tests. Abusive language models showed a more balanced behavior. The
amb-classsplit achieved the highest accuracy (64.52%) for offensive language detection. - OP-NL Performance: Models trained on offensive language and jointly performed best on OP-NL, with macro-F1 scores of 73.56 and 71.85 respectively for the
stdsplit. Abusive language models underperformed, which the authors interpret as a positive sign of specificity. - Data Cartography Effectiveness:
amb-dimconsistently outperformed other settings on HATECHEK-NL, suggesting its utility for out-of-domain generalization. However, on DALC-v2.0 itself,stdoften performed best. - Error Analysis: For the joint classification on DALC-v2.0, models struggled to distinguish between offensive and abusive classes. On HATECHEK-NL, models often failed on subtle non-hateful cases, indicating limitations in the training data's ability to generalize.
Limitations and Future Work
- Dataset Size: The authors suggest that DALC-v2.0 might be too small for data cartography to be fully effective in identifying truly robust training subsets.
- HATECHEK-NL Limitations: The paper notes that HATECHEK-NL inherits limitations from MHC, and some functional tests (e.g., reappropriation of slurs) could not be fully implemented due to a lack of evidence in Dutch.
- Future Work: The authors plan to extend DALC-v2.0 with more hate speech datasets and further validate HATECHEK-NL functionalities. They also aim to conduct focused interviews with Dutch organizations to improve the benchmark.
Summary updated