IIIT at SemEval-2016 Task 11: Complex Word Identification using Nearest Centroid Classification

Proceedings of SemEval-2016, pages 1017–1021, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics IIIT at SemEval-2016 Task 11: Complex Word Identification using Nearest Centroid Classification Ashish Palakurthi and Radhika Mamidi Kohli Center on Intelligent Systems IIIT-Hyderabad, India. ashish.palakurthi@research.iiit.ac.in radhika.mamidi@iiit.ac.in Abstract This paper describes the system that was sub- mitted to SemEval2016 Task 11: Complex Word Identification. It presents a preliminary investigation into exploring word difficulty for non-native English speakers. We developed two systems using Nearest Centroid Classifi- cation technique to distinguish complex words from simple words. Optimized over G-score, the presented solution obtained a G-score of 0.67, while the winner achieved a G-score of 0.77 and the average G-score of all the submit- ted systems in the task was 0.56. 1 Introduction Lexical Simplification aims at improving the read- ability and comprehensibility of text by transform- ing complex text into simple text. Lexical Simpli- fication (Specia et al., 2012; Belder et al., 2010; Horn et al., 2014) is the process of replacing a word in a given context with its simplest substitute to enhance the readability of the text. The process should make sure that while replacing words with other variants, the meaning of the text is preserved. Lexical Simplification (Siddharthan, 2014) is use- ful to a wide variety of target audience like people with aphasia, children and also non-native speakers. Complex Word Identification (Shardlow 2013; Paet- zold 2015) is considered to be the first step in the pipeline of Lexical Simplification. The overall per- formance of a Lexical Simplification system is thus crucially dependent upon Complex Word Identifica- tion. The problem of Complex Word Identification is relatively new in the field of Natural Language Processing. However, a few approaches have been previously proposed for this task. The simplicity score (Bott et al., 2012) of a word is computed by integrating both, frequency and length of a word. They consider a threshold value and simplify words only if the word’s frequency is lower than the fixed threshold. Matthew Shardlow (2013) explores the frequency thresholding to differentiate between sim- ple and complex words by experimenting with each threshold value on a particular corpus. However, this approach is not practically convincing. The same author also frames the problem as a machine learn- ing classification problem by designing a few fea- tures. We approach the problem at hand on similar lines. The Complex Word Identification (CWI) task is framed as a binary classification problem. Given a word in a sentence, the task is to predict whether the word is simple or complex. A word is tagged with 0 if it is simple and 1 if the word is found to be complex. 1 c(w) = � 1 if w ∈ C 0 if w ∈ S C is the set of complex words, S is the set of sim- ple words and c(w) represents the class of the word. Here is an example of a sentence taken from the training dataset provided by the organizers. ⋄ A frenulum is a …

Content truncated to 500 words