Disordered-DABS: A Benchmark for Dynamic Aspect-Based Summarization in Disordered Texts
This work addresses a gap in summarizing unstructured text for applications like social media analysis, though it is incremental as it adapts existing datasets.
The paper tackles the problem of aspect-based summarization in disordered texts, such as social media and customer feedback, by introducing Disordered-DABS, a benchmark that reveals unique challenges for current models, including GPT-3.5, with experiments showing it effectively highlights these difficulties.
Aspect-based summarization has seen significant advancements, especially in structured text. Yet, summarizing disordered, large-scale texts, like those found in social media and customer feedback, remains a significant challenge. Current research largely targets predefined aspects within structured texts, neglecting the complexities of dynamic and disordered environments. Addressing this gap, we introduce Disordered-DABS, a novel benchmark for dynamic aspect-based summarization tailored to unstructured text. Developed by adapting existing datasets for cost-efficiency and scalability, our comprehensive experiments and detailed human evaluations reveal that Disordered-DABS poses unique challenges to contemporary summarization models, including state-of-the-art language models such as GPT-3.5.