Tree transformers for neural text simplification

Abstract

Text simplification deals with the problem of making new sentences easier to read and understand, while retaining the main idea. There are five major approaches for neural text simplification, including monolingual translation approach and editing approach. A prominent issue, however, is that they all fail to utilize the rich syntactic and semantic structure embedded in the sentence. We queried the literature, and classified simplification operations according to lexical simplification, syntactic simplification and semantic simplification. We proposed structural simplification, which involves the latter two. There are three reasons for that: first, structural simplification is less studied; second, it is harder to implement; third, it is important in reducing reading comprehension difficulty. To address structural simplification, we proposed to utilize tree transformer, which is able to induce a constituency tree structure from the input sentence. We demonstrate that the tree transformer is comparable, in terms of SARI score, to three strong baselines. In addition, the tree transformer surpasses transformer baseline in terms of overall correctness, and is more superior in terms of performing more structural simplification operations, including syntactic rewrites, semantic rewrites, and sub-sentence deletions.

Date of Award	2022
Original language	English
Awarding Institution	The Hong Kong University of Science and Technology
Supervisor	Dit Yan YEUNG (Supervisor) & Michelle YIK (Supervisor)