A New Annotated Portuguese/Spanish Corpus for the Multi-Sentence Compression Task

A New Annotated Portuguese/Spanish Corpus for the Multi-Sentence Compression Task Elvys Linhares Pontes1, Juan-Manuel Torres-Moreno1,2, St´ephane Huet1, Andr´ea Carneiro Linhares3 1CERI/LIA, Universit´e d’Avignon et des Pays de Vaucluse, Avignon, France 2 ´Ecole Polytechnique de Montr´eal, Montr´eal, Canada 3Universidade Federal do Cear´a, Sobral-CE, Brasil {elvys.linhares-pontes, juan-manuel.torres, stephane.huet}@univ-avignon.fr andrea.linhares@ufc.br Abstract Multi-sentence compression aims to generate a short and informative compression from several source sentences that deal with the same topic. In this work, we present a new corpus for the Multi-Sentence Compression (MSC) task in Portuguese and Spanish. We also provide on this corpus a comparison of two state-of-the-art MSC systems. Keywords: Annotated Corpus, Multi-Sentence Compression, Multilingual Corpus. 1. Introduction Among the various applications of Natural Language Pro- cessing, Automatic Text Summarization (ATS) aims at summarizing one or more texts automatically. Summariza- tion systems identify relevant data and create a summary from key information. The (Multi-)Sentence Compression task can be seen as a subproblem of ATS with the objective to generate a shorter, informative and correct sentence from source sentence(s). In many cases, state-of-the-art NLP systems are evaluated with experiments restrained to the English language, in part because there are a lot of available English resources for most NLP tasks. As regards Multi-Sentence Compression (MSC), the available resources are unfortunately limited; to our knowledge, only one dataset is freely available and it is confined to the French language (Boudin and Morin, 2013). In this work, we present a new annotated corpus in the Por- tuguese and Spanish languages for the MSC task. Using this corpus, we evaluate two state-of-the-art systems and show that the use of several languages leads to more miti- gated results on the superiority of one system than the use of the French corpus alone. The remainder of this paper is organized as follows. In Sec- tion 2, we characterize MSC with respect to related tasks from the perspective of the available corpora. Section 3 describes the creation and the features of our corpus. In Section 4 we analyze the results achieved by state-of-the- art methods using our dataset. Finally, conclusions are set out in Section 5. 2. Related Work Sentence Compression (SC) aims at producing a reduced grammatically correct sentence from a source sentence. SC can be used in the context of the abstractive summarization of documents, the generation of article titles or the simpli- fication of complex sentences, using diverse methods (opti- mization, syntactic structure, deletion of words and/or gen- eration of sentences). The corpora for SC can be divided in two categories: deletion-based and summarization-based SC. In the case of SC by deletion of words, sentences are compressed by removing irrelevant words (Filippova et al., 2015; Ive and Yvon, 2016). Knight and Marcu (2002) de- veloped a SC corpus by aligning abstracts and sentences extracted from the Ziff-Davis corpus, which is a collec- tion of newspaper articles announcing computer products. Clarke and Lapata (2008) provided two manually created two-reference corpora for deletion-based compression. Fil- ippova and Altun (2013), and Filippova et al. (2015) ex- tracted and …

Content truncated to 500 words