Neural Versus Non-Neural Text Simplification: A Case Study

1 Abstract We propose a modular rule-based system for Text Simplification and show that it out- performs the state-of-the-art neural-based simplification system in terms of simplic- ity. We compare the output of both systems to highlight the differences between the two approaches. Further, we present an ad- aptation of our system to handle domain- specific tasks, where we employ a hybrid approach of our rule-based system and phrase-based machine translation to sim- plify medical discharge summaries in a low-resource situation. We compile a small medical simplification dataset to evaluate our proposed solution. 1 Introduction Text Simplification is loosely defined as reducing the linguistic complexity of text, without changing its meaning, to suit a wider range of audience such as: non-native speakers, children, or people with language impairments. It is usually achieved by ap- plying rewrite rules to perform two types of opera- tions: (1) lexical simplification, where difficult words are substituted with more common alterna- tives; and (2) syntactic simplification, where com- plex sentence structures are split, reordered, or de- leted to produce simpler more readable structure. To implement those rewrite rules, researchers em- ploy various methods broadly categorized into two categories: rule-based methods and data-driven methods (Siddharthan, 2014). In rule-based methods, the rules are hand-crafted a priori then applied to new text at simplification time. Examples of such rules include dictionary- based lookups for lexical simplification (Kuro- hashi & Sakai, 1999) or rules aiming at sentence restructuring into more readable formats. (A. Sid- dharthan, 2002; Vickrey & Koller, 2008). In con- trast, data-driven methods frame the simplification process as a monolingual Machine Translation problem where the rewrite rules are learned from a parallel corpus of complex-simple sentences. This enables researchers to leverage the advances in Machine Translation to address the simplification task. Considering the breakthrough achieved by Neu- ral Machine Translation, we ask the questions: Could similar success be achieved in Text Simpli- fication by employing neural architectures? Would rule-based methods be more effective since simpli- fication is a fundamentally different task than translation? To answer these questions, first, we propose a non-neural general-purpose rule-based simplifica- tion system. We, then, show how it can be adapted to address domain-specific simplification tasks by leveraging a small parallel dataset from the target domain. Subsequently, we compare the output of our system with that of a recently proposed neural- based simplification system (Zhang & Lapata, 2017). In our study, we focus on two simplification domains: (1) general-purpose English, for which we run our tests using Wikipedia-based datasets; and (2) medical English, for which we compile a small medical parallel corpus of complex-simple pairs and use it to test our systems. We show that our rule-based system outperforms the neural sys- tem, in terms of simplicity, both qualitatively and quantitatively. Finally, we reflect on the output of both systems to pinpoint the shortcomings of each approach and encourage researchers to address them in future research. 2 Rule-Based Simplification System In this section, we describe our proposed rule- based simplification system. It comprises two …

Content truncated to 500 words