Go to ACL ARR 2025 July homepageAF-HPSum: Agentic Framework for Multi-Strategy Summarization of Hindi Podcasts
Abstract: Podcasts are lengthy audio conversations which require a significantly large context input for accurate summarization via computational models. Despite recent advancements in Large Language Models (LLMs), it is challenging to summarize a transcribed podcast conversation using LLMs, due to input context length, long-range dependencies, noisy data and attention mismatch. In this paper, we propose an agentic framework for LLM-based summarization of Hindi Podcasts (AF-HPSum) which leverages multiple strategies, including a rule-based deletion strategy for compressive summarization. Using multiple LLMs, both open-weighted and closed-source, we evaluate the performance of our framework and observe that an iterative strategy helps preserve long-range dependencies and produce relevant summaries. We also conducted a preliminary human evaluation, which elicits model selection and helps build a comprehensive pipeline for podcast summarization. Through parameter-efficient training of open-weighted models and our iterative approach, we achieved a significant performance improvement over closed-weight and larger models by a significant margin. We will release our framework codebase, prompts, data and output with this paper.
Paper Type: Long
Research Area: Summarization
Research Area Keywords: Hindi Summarization, Abstractive Summarization, Podcast Summarization, Summarization Framework, Large Language Models
Contribution Types: Approaches to low-resource settings, Approaches low compute settings-efficiency, Publicly available software and/or pre-trained models, Data resources
Languages Studied: Hindi
Submission Number: 651
Loading