A Keyword-based Monolingual Sentence Aligner in Text Simplification

A Keyword-based Monolingual Sentence Aligner in Text Simplification 1Chung-Chi Huang 2Maxine Eskenazi 3Mei-Hua Chen 4Ping-Che Yang 1,2Language Technologies Institute, CMU, United States 3Hua Fan University, Taipei, Taiwan 4Institute for Information Industry, Taipei, Taiwan {1u901571,3chen.meihua}@gmail.com,2max+@cs.cmu.edu,4maciaclark@iii.org.tw Abstract We introduce a method for learning to align sentences in monolingual parallel articles for text simplification. In our approach, word keyness is integrated to prefer aligning essential words in sentences. The method involves estimating word keyness based on TF*IDF and semantic PageRank, and word nodes’ parts-of-speech and degrees of reference. At run-time, the keyword analyses are used as word weights in sentence similarity measure. And a global dynamic programming goes through sentence similarities further weighted by aligned content-word ratios and positions of aligned words to determine the optimal candidates of parallel sentences. We present a prototype sentence aligner, KEA, that applies the method to monolingual parallel articles. Evaluation shows that KEA pays more attention to key words during sentence aligning and outperforms the current state-of-the-art in alignment accuracy and f-measure. Our pilot study also indicates that language learners benefit from our sentence-aligned parallel articles in reading comprehension test. 1 Introduction Many articles are posted on the Web every day, and an increasing number of educational websites specifically provide articles for audiences with different needs. For example, NewsInLevels (www.newsinlevels.com) and BreakingNewsEnglish (www.breakingnewsenglish.com) select news articles and provide versions with different readability for language learners. Simple Wikipedia (simple.wikipedia.org) and EasierEnglishWiki (eewiki.newint.org) contain articles easier to read with simpler vocabulary and syntactic structure than English Wikipedia and New Internationalist for people with low literacy. And SoundReading (www.soundreading.com) even has audio recording for those with learning disabilities such as dyslexia. Language learning websites such as NewsInLevels and EasierEnglishWiki typically simplify original articles into easier ones and present the original and easier articles as pairs to non-native speakers, children, or lay people. However, language learners may want to compare the article pairs conveying the same information at sentence level, and most text simplification systems build on top of original and simplified sentence pairs. Unfortunately, current monolingual sentence alignment methods treat article sentences as bags of words, equally weight words, and align sentences with high word-overlap ratios. These article pairs could be sentence aligned more accurately if a system distinguished words of different importance and leveraged their importance levels in articles while aligning. Consider the original-simplified article pair in Figure 1. The best sentence alignment methods are probably not the ones with equal word weights (i.e., weights are the same with “the” and “gorilla” and the same with “everything” and “project”). A PACLIC 28 542 Copyright 2014 by Chung-Chi Huang, Maxine Eskenazi, Mei-Hua Chen, and Ping-Che Yang 28th Pacific Asia Conference on Language, Information and Computation pages 542–550 Figure 1. An example KEA sentence alignment for an article pair. good aligning approach might take into account the words’ significance in the pair. Intuitively, word significance can be evaluated by keyword extraction methods and by leveraging word significance, sentence aligners can be biased towards aligning sentences with more words that are more …

Content truncated to 500 words