Simplification Using Paraphrases and Context-Based Lexical Substitution

Proceedings of NAACL-HLT 2018, pages 207–217 New Orleans, Louisiana, June 1 - 6, 2018. c⃝2018 Association for Computational Linguistics Simplification Using Paraphrases and Context-based Lexical Substitution Reno Kriz∗, Eleni Miltsakaki†, Marianna Apidianaki∗△ and Chris Callison-Burch∗ ∗ Computer and Information Science Department, University of Pennsylvania △ LIMSI, CNRS, Universit´e Paris-Saclay, 91403 Orsay † Choosito, Inc. {rekriz,marapi,ccb}@seas.upenn.edu, eleni@choosito.com Abstract Lexical simplification involves identifying complex words or phrases that need to be sim- plified, and recommending simpler meaning- preserving substitutes that can be more eas- ily understood. We propose a complex word identification (CWI) model that exploits both lexical and contextual features, and a simpli- fication mechanism which relies on a word- embedding lexical substitution model to re- place the detected complex words with sim- pler paraphrases. We compare our CWI and lexical simplification models to several base- lines, and evaluate the performance of our sim- plification system against human judgments. The results show that our models are able to detect complex words with higher accuracy than other commonly used methods, and pro- pose good simplification substitutes in context. They also highlight the limited contribution of context features for CWI, which nonetheless improve simplification compared to context- unaware models. 1 Introduction Automated text simplification is the process that involves transforming a complex text into one with the same meaning, but can be more easily read and understood by a broader audience (Saggion, 2017). This process includes several subtasks such as complex word and sentence identification, lex- ical simplification, syntactic simplification, and sentence splitting. In this paper, we focus on lex- ical simplification, the task of replacing difficult words in a text with words that are easier to under- stand. Lexical simplification involves two main pro- cesses: identifying complex words within a text, and suggesting simpler paraphrases for these words that preserve their meaning in this context. To identify complex words, we train a model on data manually annotated for complexity. Unlike Figure 1: An example sentence with complex words identified by our classifier, and their substitutes pro- posed by the embedding-based substitution model. previous work, our classifier takes into account both lexical and context features. We extract can- didate substitutes for the identified complex words from SimplePPDB (Pavlick and Callison-Burch, 2016), a database of 4.5 million English simpli- fication rules linking English complex words to simpler paraphrases. We select the substitutes that best fit each context using a word embedding- based lexical substitution model (Melamud et al., 2015). An example sentence, along with the com- plex words identified by our model and the pro- posed replacements, is shown in Figure 1. We show that our complex word identification classi- fier and substitution model improve over several baselines which exploit other types of information and do not account for context. Our approach pro- poses highly accurate substitutes that are simpler than the target words and preserve the meaning of the corresponding sentences. 2 Related Work Prior approaches to text simplification have ad- dressed the task as a monolingual translation prob- lem (Zhu et al., …

Content truncated to 500 words