SimpleText Best of Labs in CLEF-2022: Simplify Text Generation with Prompt Engineering

Abstract

This paper reports our approach to the SimpleText@CLEF-2022. For the task 1: what is in (or out)?, we designed a two-stage filtering scheme that utilizes the traditional keyword finding approach TF-IDF score to find the important documents in the first stage and the important sentences in the second stage. The result is comparable to manual run and ranked first in task 1. For the Task 3: Rewrite this!, our system adopts the T5 generation model to rewrite the original sentences. We fine-tuned the model to generate simplified sentence. The result ranked second in task 3. The simplified sentence generated by T5 model cannot fully express the meaning of the original sentence, in a following further experiments, we adopted the GPT3.5 and GPT4 models to generate simple text, and they give better results according to in our evaluation metrics based on readability and vocabulary simplicity.