Exploring the subtopic-based relationship map strategy for multi-document summarization

Abstract

Neste artigo, foi adaptada e explorada uma estratégia de geração de sumários multidocumento com base na abordagem de mapas de relacionamento, a qual representa textos como grafos (mapas) de segmentos inter-relacionados e aplica técnicas de percurso em grafo para produzir os sumários. Em particular, utilizou-se o Caminho Denso-Segmentado, um método sofisticado que tenta representar em um sumário os principais subtópicos dos textos de origem, mantendo a informatividade do sumário. Além disso, também foram investigadas algumas técnicas de segmentação e agrupamento de subtópicos bem conhecidas, a fim de selecionar corretamente as informações mais relevantes para compor o sumário final. Mostra-se que este método de sumarização baseado em subtópicos supera outros métodos de sumarização multidocumento e atinge resultados do estado da arte, competindo com os métodos de sumarização profundos mais sofisticados da área. Explorando mapas de relacionamento com base em subtópicos para sumarização multidocumento Abstract: In this paper we adapt and explore a strategy for generating multidocument summaries based on the relationship map approach, which represent texts as graphs (maps of interrelated segments and apply traversing techniques for producing the summaries. In particular, we work on the Segmented Bushy Path, a sophisticated method which tries to represent in a summary the main subtopics from the source texts while keeping its informativeness. In addition, we also investigate some well-known subtopic segmentation and clustering techniques in order to correctly select the most relevant information to compose the final summary. We show that this subtopic-based summarization method outperforms other methods for multi-document summarization and achieves state of the art results, competing with the most sophisticated deep summarization methods in the area.

Update summarization: building from scratch for Portuguese and comparing to English

“…As discussed in [16,18], to do summarization, or, in more general terms, to produce any natural language processing (NLP) application, requires dealing with linguistic knowledge of varied levels, including, e.g., processing words and their morphology, syntax, semantics, and discourse. Discourse, in particular, has traditionally been investigated for Portuguese summarization (see, e.g., [13,14,19,20]), including discourse relations (as predicted by the rhetorical structure theory-RST [21] and cross-document structure theory-CST [22]) and topics/subtopics. Subtopics are of special interest in this work and are introduced in more details in "The new methods" section when we present one of the methods that we test.…”

“…We performed experiments over two distinct datasets for the US task, the CSTNews-Update and the TAC 2009 datasets. The first one, the CSTNews-Update, is a distinct arrangement of the CSTNews corpus [28,29], which has been used for many investigations on Automatic Summarization for Portuguese (see, e.g., [2][3][4][5][6][7][8][9][10][11][12][13][14]). We propose such arrangement here to be a reference corpus to train and test US methods for Portuguese.…”

“…Most of them are for the English language. For Portuguese, there are many investigations on the traditional summarization tasks (see, e.g., [2][3][4][5][6][7][8][9][10][11][12][13][14]), but, to the best of our knowledge, there is only one previous effort in the US field, which is a preliminary work conducted by the authors of this paper [15].…”

“…Since subtopics have recently shown to be very useful in summarization (see, e.g., [13,14]), producing better results, we have opted to explore this specific linguistic knowledge for US. We have included subtopics into the KLSum strategy, which is used by several summarization systems, as already commented in the previous section.…”

An investigation of linguistic problems in automatic multi-document summaries / Uma investigação de problemas linguísticos em sumários automáticos multidocumento

“…It is a consensus that a good summary should contain the most relevant information in the texts, and the area has achieved significant progress in producing summaries that are more informative. The progress is the result of both linguistically poor and rich summarization methods, such as the empirical/statistical approaches (see, e.g., ANDO et al, 2000;CARBONELL et al, 1997;HAGHIGHI;VANDERWENDE, 2009;MIHALCEA;TARAU, 2005;RIBALDO et al, 2016) and the deep ones (CARDOSO; PARDO, 2010;RADEV, 1995;RADEV, 2000;ZHANG et al, 2002).…”