Data-Efficient Hate Speech Detection via Cross-Lingual Nearest Neighbor Retrieval with Limited Labeled Data

TL;DR

The paper tackles hate speech detection in languages with limited labeled data by introducing a cross-lingual nearest-neighbor retrieval framework. It embeds a large multilingual pool, retrieves the most relevant cross-language instances for each target-language example, and fine-tunes a multilingual LM on a combined dataset, achieving consistent gains over target-only training and often exceeding prior state-of-the-art. The approach is data-efficient, requiring as few as 20–200 retrieved instances to outperform baselines, with an average optimum near 2,000 retrieved samples across languages. It demonstrates strong scalability and adaptability to new languages and tasks, and investigates enhancements such as Maximum Marginal Relevance for diversity. These findings have practical implications for deploying hate speech detection models in low-resource settings with limited labeling budgets.

Abstract

Considering the importance of detecting hateful language, labeled hate speech data is expensive and time-consuming to collect, particularly for low-resource languages. Prior work has demonstrated the effectiveness of cross-lingual transfer learning and data augmentation in improving performance on tasks with limited labeled data. To develop an efficient and scalable cross-lingual transfer learning approach, we leverage nearest-neighbor retrieval to augment minimal labeled data in the target language, thereby enhancing detection performance. Specifically, we assume access to a small set of labeled training instances in the target language and use these to retrieve the most relevant labeled examples from a large multilingual hate speech detection pool. We evaluate our approach on eight languages and demonstrate that it consistently outperforms models trained solely on the target language data. Furthermore, in most cases, our method surpasses the current state-of-the-art. Notably, our approach is highly data-efficient, retrieving as small as 200 instances in some cases while maintaining superior performance. Moreover, it is scalable, as the retrieval pool can be easily expanded, and the method can be readily adapted to new languages and tasks. We also apply maximum marginal relevance to mitigate redundancy and filter out highly similar retrieved instances, resulting in improvements in some languages.

Figures (4)

• Figure 1: Overview of the proposed method. Given a small number of examples from a target language, we search in a large pool of multilingual data for closely related instances. We then combine the retrieved instances with the target language data and train a multilingual model on them for hate speech detection.
• Figure 2: Performance across different numbers of retrieved instances (10 to 100,000, log-scaled) for four languages. Each curve corresponds to selected training sizes. The AVG line shows the average over 12 training sizes.
• Figure 3: Sankey diagram of the distribution of the top four retrieved source tasks per target task.
• Figure 4: Sankey diagrams of the distribution of the top four retrieved source tasks (left), languages (middle), and labels (right) for each target task.