TriS: A Statistical Sentence Simplifier with Log-linear Models and Margin-based Discriminative Training

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 474–482, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP TriS: A Statistical Sentence Simplifier with Log-linear Models and Margin-based Discriminative Training Nguyen Bach, Qin Gao, Stephan Vogel and Alex Waibel Language Technologies Institute Carnegie Mellon University 5000 Forbes Avenue, Pittsburgh PA, 15213 {nbach, qing, stephan.vogel, ahw}@cs.cmu.edu Abstract We propose a statistical sentence simplifica- tion system with log-linear models. In contrast to state-of-the-art methods that drive sentence simplification process by hand-written linguis- tic rules, our method used a margin-based dis- criminative learning algorithm operates on a feature set. The feature set is defined on statis- tics of surface form as well as syntactic and de- pendency structures of the sentences. A stack decoding algorithm is used which allows us to efficiently generate and search simplification hypotheses. Experimental results show that the simplified text produced by the proposed system reduces 1.7 Flesch-Kincaid grade level when compared with the original text. We will show that a comparison of a state-of- the-art rule-based system (Heilman and Smith, 2010) to the proposed system demonstrates an improvement of 0.2, 0.6, and 4.5 points in ROUGE-2, ROUGE-4, and AveF10, respec- tively. 1 Introduction Complicated sentences impose difficulties on reading comprehension. For instance, a person in 5th grade can comprehend a comic book easily but will struggle to understand New York Times articles which require at least 12th grade average reading level (Flesch, 1981). Complicated sentences also challenge natural language processing applications including, but not limited to, text summarization, question answering, information extraction, and machine translation (Chandrasekar et al., 1996). An example of this is syntactic parsing in which long and complicated sentences will generate a large number of hypotheses and usually fail in disam- biguating the attachments. Therefore, it is desirable to pre-process complicated sentences and generate sim- pler counter parts. There are direct applications of sen- tence simplification. Dalemans et al. (2004) applied sentence simplification so that the automatically gen- erated closed caption can fit into limited display area. The Facilita system generates accessible content from Brazilian Portuguese web pages for low literacy readers using both summarization and simplification technolo- gies (Watanabe et al., 2009). This paper tackles sentence-level factual simplifica- tion (SLFS). The objective of SLFS is twofold. First, SLFS will process the syntactically complicated sen- tences. Second, while preserving the content meaning, SLFS outputs a sequence of simple sentences. SLFS is an instance of the broader spectrum of text-to-text gen- eration problems, which includes summarization, sen- tence compression, paraphrasing, and sentence fusion. Comparing to sentence compression, sentence simplifi- cation requires the conversion to be lossless in sense of semantics. It is also different from paraphrasing in that it generates multiple sentences instead of one sentence with different constructions. There are certain specific characteristics that compli- cate a sentence, which include length, syntactic struc- ture, syntactic and lexical ambiguity, and an abundance of complex words. As suggested by its objective, sen- tence simplification outputs “simple sentences”. Intu- itively, a simple sentence …

Content truncated to 500 words