SimpleScience: Lexical Simplification of Scientific Terminology

Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 1066–1071, Austin, Texas, November 1-5, 2016. c⃝2016 Association for Computational Linguistics SimpleScience: Lexical Simplification of Scientific Terminology Yea-Seul Kim and Jessica Hullman University Washington Information School yeaseul1, jhullman@uw.edu Matthew Burgess University of Michigan Computer Science Department mattburg@umich.edu Eytan Adar University of Michigan School of Information eadar@umich.edu Abstract Lexical simplification of scientific terms rep- resents a unique challenge due to the lack of a standard parallel corpora and fast rate at which vocabulary shift along with research. We in- troduce SimpleScience, a lexical simplifica- tion approach for scientific terminology. We use word embeddings to extract simplification rules from a parallel corpora containing sci- entific publications and Wikipedia. To eval- uate our system we construct SimpleSciGold, a novel gold standard set for science-related simplifications. We find that our approach out- performs prior context-aware approaches at generating simplifications for scientific terms. 1 Introduction Lexical simplification, the process of reducing the complexity of words by replacing them with sim- pler substitutes (e.g., sodium in place of Na; insects in place of lepidopterans) can make scientific texts more accessible to general audiences. Human-in- the-loop interfaces present multiple possible simpli- fications to a reader (on demand) in place of jargon and give the reader familiar access points to under- standing jargon (Kim et al., 2015). Unfortunately, simplification techniques are not yet of high enough quality for fully automated scenarios. Currently lexical simplification pipelines for sci- entific texts are rare. The vast majority of prior methods assume a domain independent context, and rely on Wikipedia and Simple English Wikipedia, a subset of Wikipedia using simplified grammar and terminology, to learn simplifications (Biran et al., 2011; Paetzold and Specia, 2015), with translation- based approaches using an aligned version (Coster and Kauchak, 2011; Horn et al., 2014; Yatskar et al., 2010). However, learning simplifications from Wikipedia is not well suited to lexical sim- plification of scientific terms. Though generic or established terms may appear in Wikipedia, novel terms associated with new advances may not be re- flected. Wikipedia’s editing rules also favor gener- ality over specificity and eliminate redundancy, both of which are problematic in providing a rich train- ing set that matches simple and complex terms. Fur- ther, some approaches work by detecting all pairs of words in a corpus and filtering to isolate synonym or hypernym-relationship pairs using WordNet (Biran et al., 2011). Like Wikipedia, WordNet is a general purpose semantic database (Miller, 1995), and does not cover all branches of science nor integrate new terminology quickly. Word embeddings do not require the use of pre- built ontologies to identify associated terms like simplifications. Recent work indicates that they may improve results for simplification selection: deter- mining which simplifications for a given complex word can be used without altering the meaning of the text (Paetzold and Specia, 2015). Embeddings have also been explored to extract hypernym rela- tions from general corpora (Rei and Briscoe, 2014). However, word embeddings have not been used for generating …

Content truncated to 500 words