Bangla Event Detection: LLMs beat encoders on noisy text

A new study from Bangladeshi researchers tackles a critical blind spot in event detection (ED): real-world noise. Typical ED systems are tested only on clean, curated text, ignoring the messy data encountered in practice—especially for low-resource languages like Bangla. The team introduces a generalized Bangla news event ontology and a benchmark of 9,979 manually annotated sentences spanning 40 event subtypes. The dataset includes three conditions: clean news text, real-world Automatic Speech Recognition (ASR) transcripts, and orthographically corrupted text (a common issue in user-generated content).

They systematically evaluated fine-tuned encoder-only models (BanglaBERT and XLM-R) alongside instruction-tuned decoder-only LLMs (Llama 3 and Gemma 3). The results reveal a stark architectural trade-off: encoders achieve higher performance on clean text but suffer severe degradation under noise—especially when event triggers are corrupted. Decoder-only LLMs, in contrast, proved far more robust. Additional experiments showed that embedding annotation guidelines during instruction tuning raised the performance baseline on noisy text but inconsistently reduced the degradation gap. Scaling decoder-only models consistently improved robustness, while combining clean and noisy training data served as an effective regularization that disproportionately helped encoders, significantly narrowing the robustness gap.