News Reporter: A Multi-lingual LLM Framework for Broadcast T.V News

TL;DR

This paper introduces a multi-lingual LLM framework tailored for broadcast TV news, leveraging a new dataset of QA pairs from news transcripts and a retrieval-augmented generation approach to enhance answer accuracy and verifiability.

Contribution

It presents a novel dataset of news transcript QA pairs and a fine-tuned LLM model with a retrieval-augmented method for improved news answer accuracy.

Findings

Model surpasses similar-sized base models on open benchmarks.

QA dataset improves LLM training for news applications.

RAG method enhances answer contextualization and verification.

Abstract

Large Language Models (LLMs) have fast become an essential tools to many conversational chatbots due to their ability to provide coherent answers for varied queries. Datasets used to train these LLMs are often a mix of generic and synthetic samples, thus lacking the verification needed to provide correct and verifiable answers for T.V. News. We collect and share a large collection of QA pairs extracted from transcripts of news recordings from various news-channels across the United States. Resultant QA pairs are then used to fine-tune an off-the-shelf LLM model. Our model surpasses base models of similar size on several open LLM benchmarks. We further integrate and propose a RAG method to improve contextualization of our answers and also point it to a verifiable news recording.