Unveiling the Segmentation Power of LLMs: Zero-Shot Invoice Item Description Analysis

Segmenting invoice item description into attributes that describe its features may be a newsworthy alternative for subsequent entity resolution. This paper presents a set of experiments to show the performance of seven LLMs, including Llama-3, Sabiá-2-Medium, Command R+, Claude 3 Opus, GPT-3.5, GPT-4, and Mixtral 8x22B, in segmenting text within Invoice items descriptions using zero-shot learning techniques. We have employed accuracy, precision, recall, and F1-score evaluation metrics to highlight the effectiveness of LLMs. The experiment involved segmentation preparation, model training, prompt optimization, attribute extraction, and output generation. The objective is to determine each model's precision in accurately identifying segmentation within invoice item descriptions.