No Train, No Pain? Assessing the Ability of LLMs for Text Classification with no Finetuning
Richard Fechner, Jens Dörpinghaus · Annals of Computer Science and Information Systems · 2024
Modern SotA Text Classification algorithms depend heavily on well annotated and diverse data capturing the intricacies of the unknown data distribution.What options do we have when labeled data is sparse or annotation is expensive and time consuming?With the advent of strong LLM backbones, we have another option at our disposal: Text Classification by making use of the reasoning ability and the strong general prior of contemporary foundation models.In this work we assess the ability of cutting edge LLMs for Text Classification and find that for the right combination of backbone and prompt strategy we're able to near-rival trained baselines for the advanced task of mapping job-postings to a taxonomy of industrial sectors without any finetuning.All our code is made publicly available at our github repository 1 .