Sentence Simplification for Semantic Role Labelling and Information Extraction

Proceedings of Recent Advances in Natural Language Processing, pages 285–294, Varna, Bulgaria, Sep 2–4, 2019. https://doi.org/10.26615/978-954-452-056-4_033 285 Sentence Simplification for Semantic Role Labelling and Information Extraction Richard Evans Research Institute in Information and Language Processing University of Wolverhampton United Kingdom r.j.evans@wlv.ac.uk Constantin Or˘asan Research Institute in Information and Language Processing University of Wolverhampton United Kingdom c.orasan@wlv.ac.uk Abstract In this paper, we report on the extrin- sic evaluation of an automatic sentence simplification method with respect to two NLP tasks: semantic role labelling (SRL) and information extraction (IE). The paper begins with our observation of challenges in the intrinsic evaluation of sentence sim- plification systems, which motivates the use of extrinsic evaluation of these sys- tems with respect to other NLP tasks. We describe the two NLP systems and the test data used in the extrinsic evaluation, and present arguments and evidence motivat- ing the integration of a sentence simplifi- cation step as a means of improving the accuracy of these systems. Our evaluation reveals that their performance is improved by the simplification step: the SRL sys- tem is better able to assign semantic roles to the majority of the arguments of verbs and the IE system is better able to identify fillers for all IE template slots. 1 Introduction Sentence simplification is one aspect of text sim- plification, which is concerned with the conver- sion of texts into a more accessible form. In many cases, text simplification is performed to facili- tate subsequent human or machine text process- ing. This may include processing for human read- ing comprehension (Canning, 2002; Scarton et al., 2017; Or˘asan et al., 2018) or for NLP tasks such as dependency parsing (Jel´ınek, 2014), informa- tion extraction (Jonnalagadda et al., 2009; Evans, 2011; Peng et al., 2012), semantic role labelling (Vickrey and Koller, 2008), and multidocument summarisation (Blake et al., 2007; Siddharthan et al., 2004). In previous research, Caplan and Waters (1999) noted a correlation between sentence comprehen- sion difficulty for human readers and the numbers of propositions expressed in the sentences being read.1 Evans and Or˘asan (2019) presented an iter- ative rule-based approach to sentence simplifica- tion which is intended to reduce the per sentence propositional density of input texts by convert- ing sentences which contain compound clauses and complex NPs2 into sequences of simpler sen- tences. Evaluation of text simplification systems is dif- ficult, especially when such evaluations need to be conducted repeatedly for development purposes and cost is a critical factor. In general, the choice of evaluation method depends on the purpose of the simplification task. Various types of evaluation are currently used, but these are problematic. In previous work, evaluation of sentence simplifica- tion systems (including Evans and Or˘asan’s (2019) system, which is extrinsically evaluated in our cur- rent paper) has relied on one or more of three main approaches: the use of overlap metrics such as Levenshtein distance (Levenshtein, 1966), BLEU score (Papineni et al., 2002) and SARI (Xu et al., 2016) to compare system output with human sim- plified texts …

Content truncated to 500 words