Sentence Tagging as Metric Learning Using Data Augumentation with Large Language Model

Abstract

The task of tagging articles is one of the most fundamental tasks in natural language processing. Uzabase, Inc., which provides business information infrastructure, frequently faces the task of tagging economic articles. In particular, Flash Opinion service matches user questions with experts by tagging the questions to represent the field of expertise. There is a demand to reduce the workload of operators who perform the question tagging. If this tagging task is formulated as a conventional multi-label classification problem, the model would need to be retrained every time tags are added or removed. In the present study, we demonstrate a method that transforms the task into a problem of distance learning between tags and question texts by using data augmentation with tag names through large-scale language models. The proposed method was applied to an actual dataset obtained from operations and verified to provide better tag recommendation for operators than multi-label classification models.