Unsupervised Syntactic Text Simplification with AMR

Abstract

Syntactic text simplification, the task of reducing the grammatical complexity of text while preserving the content, can be useful for non-native speakers, text summarization, and other downstream natural language processing tasks. Many traditional methods are rule-based and do not generalize, while methods that rely on modern large language models often are limited by prohibitive computational requirements or data privacy concerns. We present a text simplification pipeline based on Abstract Meaning Representation which can run on modest hardware, and report on intrinsic and extrinsic evaluations of its performance. We find that our method achieves comparable performance to GPT-3.5, at a fraction of the cost, and without any privacy concerns. Additionally, it outperforms a best in class rule based text simplifier. To see if our simplified text preserves the semantics of the original text, we evaluate our simplified text in two downstream tasks: relation extraction, and entity linking. We find that our syntactic simplification pipeline has limited or no impact on the performance of the methods we evaluate for these tasks, indicating that our pipeline preserves the information in the original text.