IBC2024 Tech Papers: Multi-Label Indexing Technology for News with AI-based Text Processing

Abstract

Broadcast media organisations produce many news scripts every day for dissemination as content. Such text data is often reused in the process of producing TV programmes and web news. To efficiently utilise this much data, it is necessary to accurately attach metadata such as labels that indicate the content of the text. However, manually assigning labels takes an enormous amount of time and effort. With the aim of reducing costs, we have developed a system that automatically labels news articles. A major challenge in the multi-label text classification task in the news domain is known as ‘imbalanced learning.’ We proposed a novel loss function that utilises some weights and a label-smoothing technique to suppress label imbalance. Experimental results show that our method out performs baselines. We introduce a prototype system based on our method as a test bed for content creation and discuss some of the results that it achieves.

Introduction

Much text data is utilised in the process of producing TV programmes and web news. To create media content efficiently, it is necessary to attach accurate metadata, such as labels that indicate the content, to large amounts of text. Metadata attached to text can enable producers to efficiently retrieve and use past material in the creation of new content and enable viewers to easily access articles that they want.