Evaluating Encoder and Decoder Robustness for Bangla Event Detection in Noisy Text

On June 29, 2026, researchers Tanvir Ahmed Sijan and colleagues published a study evaluating the robustness of event detection systems for the Bangla language, focusing on how these systems perform with noisy text. This research highlights the challenges faced in low-resource languages, where traditional models are often tested on clean, curated datasets.

Understanding the Need for Robust Event Detection

Event detection (ED) systems are crucial for processing and understanding real-time information, especially in a multilingual context. The study introduces a new Bangla news event ontology along with a benchmark containing 9,979 annotated sentences across 40 event subtypes. This benchmark includes clean news text, Automatic Speech Recognition (ASR) transcripts, and orthographically corrupted text, providing a comprehensive view of the challenges in real-world applications.

The researchers systematically assessed the performance of fine-tuned encoder-only models, such as BanglaBERT and XLM-R, against instruction-tuned decoder-only large language models (LLMs), including Llama 3 and Gemma 3. Their findings reveal significant differences in how these models handle various text conditions.

Key Findings on Model Performance

The research indicates a clear architectural trade-off: encoder models perform well on clean text but struggle significantly when faced with noise. In contrast, decoder-only LLMs demonstrate greater robustness, particularly when event triggers are corrupted. This finding is critical for future developments in event detection systems, particularly for languages like Bangla that may not have extensive resources.

• Encoder Models: Higher performance on clean text.
• Decoder Models: More robust under noisy conditions.

Moreover, the inclusion of embedding annotation guidelines during instruction tuning tends to elevate performance on noisy text but does not consistently mitigate performance degradation across different noise conditions. This inconsistency presents challenges for developers aiming to optimize these models for real-world applications.

The Impact of Model Scaling and Training Strategies

One of the noteworthy insights from the study is that scaling up models consistently enhances the robustness of decoder-only LLMs. The researchers advocate for a combined training approach that incorporates both clean and noisy data, which serves as an effective regularization strategy. This method particularly benefits encoder architectures, thereby narrowing the robustness gap between encoder and decoder models.

Such advancements in model training and evaluation are essential for improving the accuracy and reliability of event detection systems in low-resource languages like Bangla. As the field evolves, the findings from this study could inform future research and development efforts.