Exploring Cognitively Informed Sentence Simplification with Gaze-Guided Text Generation

Abstract

Automatic text simplification has mostly relied on human judgments when it comes to what is considered easy or difficult to read. Eye movements while reading can offer a more direct and objective signal of processing effort and reading ease. In this paper, we explore gaze-guided text generation (GGTG), an approach to control reading ease in generated texts, and assess its use for sentence simplification. GGTG employs a gaze model that is trained to predict eye-tracking measures such as reading times or regression rates, which are then used to rerank next-token probabilities generated by a language model. We evaluated the approach on an English sentence simplification benchmark and found gains in automatic evaluation metrics, although the simplification operations are mostly limited to the lexical level. Its modular nature also allows GGTG to be combined with other simplification techniques such as prompting or fine-tuning.

Resources

Details