Parallel Sentence Compression

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 1503–1513, Osaka, Japan, December 11-17 2016. Parallel Sentence Compression Julia Ive1,2, Fran¸cois Yvon1 LIMSI, CNRS, Univ Paris-Sud, Universit´e Paris-Saclay, 91 403 Orsay, France,1 Cochrane France, INSERM U1153, 75181 Paris, France2 {firstname.lastname}@limsi.fr Abstract Sentence compression is a way to perform text simplification and is usually handled in a monolingual setting. In this paper, we study ways to extend sentence compression in a bilingual context, where the goal is to obtain parallel compressions of parallel sentences. This can be beneficial for a series of multilingual natural language processing (NLP) tasks. We compare two ways to take bilingual information into account when compressing parallel sentences. Their efficiency is contrasted on a parallel corpus of News articles. 1 Introduction Text simplification is a well studied application of Natural Language Processing (NLP) tech- niques. Its main goal is to reduce the complexity of a text without degrading the informational content. This task proves useful for a wide range of applications, be they human-oriented (e.g. text adaptation for language learning purposes, for people with reading disabilities etc. (Siddharthan, 2014; Klaper et al., 2013)) or machine-oriented, serving as a basis to improve the efficiency of other NLP downstream components (e.g., parsing (Jonnalagadda et al., 2009), semantic role labeling (Vickrey and Koller, 2008) etc.). Simplification can be performed at dif- ferent linguistic levels: lexical (Paetzold and Specia, 2016), syntactic (Siddharthan, 2011), or both (Paetzold and Specia, 2013). Sentence compression is a way to perform simplification at the level of sentences, by reducing the sentence length without sacrificing important information. Many works only consider purely syntactic simplifications, though lexical changes are also possible, especially in language learning scenarios (Cohn and Lapata, 2008; Napoles et al., 2011a). By and large, the motivations that have been put forward for monolingual sentence com- pression can be also used to motivate bilingual sentence compression, understood here as the generation of parallel compressions of parallel sentences. Bilingual Sentence Compression can be used, for instance, to produce simpler versions of a parallel text for learning purposes, or to generate summaries and subtitles in different languages, or even to build simplified parallel corpora for training a Machine Translation (MT) system. Parallel sentence compression can be approached in many ways: it is first possible to compress independently each side of a bitext using monolingual simplification, an approach which however runs the risk of breaking the parallelism of the resulting corpus. Compression could also be performed in a symmetric manner by generalizing monolingual algorithms to the case of parallel sentences. We focus here on an asymmetrical scenario, where the target compression is a translation of a previously simplified source sentence. In this context, our main contribution consists in studying various bitext compression method- ologies that should ensure the parallelism of the simplified bitext, which, to our knowl- edge, is the first attempt of this kind. Two compression methods are developed and compared in Section 2 : (1) a dynamic programming (DP) approach, considering the …

Content truncated to 500 words