Automatic identification of cross-document structural relationships

Abstract

Analysis on inter-document relationship is one of the important studies in multi document analysis. In this paper, we will focus on some special properties that multi document articles hold, specifically news articles. Information across news articles reporting on the same story are often related. Cross-document Structure Theory (CST) gives the relationship between pairs of sentences from different documents. For example, two sentences might have relationships such as identical, overlapping or contradicting. Our aim here is to automatically identify some of these CST relationships. We applied the well known machine learning technique, SVMs for this purpose and obtained some comparable results.

Graph-Based Complex Representation in Inter-Sentence Relation Recognition in Polish Texts

“…The best result was achieved with CBR based on the cosine similarity measure. It expressed improved results than in K u m a r et al [12]: Identity 0.966, Subsumption 0.803, Description 0.786, and 0.722 for Overlap.…”

“…The features of A l e i x o and P a r d o [1] were expanded with: 1) cosine similarity of word vectors, 2) intersection of common words measured with the Jaccard Index, 3) an indicator of longer sentence (1 if S1 was longer, 0 if equal, -1 if S1 was shorter), 4) and uni-directional word coverage ratio (S1 → S2 and S2 → S1). K u m a r et al [12] followed Z a h r i and F u k u m o t o [37], but restricted the set of relations further down to four: Identity, Subsumption, Overlap and Elaboration.…”

Graph-Based Approach to Recognizing CST Relations in Polish Texts

“…In our work we proposed a method for the recognition of the full set of 17 CST relations, in contrast to the limited of subsets used in literature, e.g. in (Kumar et al, 2012a). Our method outperforms also the state of the art algorithm when compared on a corpus of the similar origin and content.…”

“…The features of (Zahri and Fukumoto, 2011) were extended with the Jaccard based similarity of noun phrases and verb phrases. CBR based on the cosine similarity measure expressed improved results than in (Kumar et al, 2012a): Identity 0.966, Subsumption 0.803, Description 0.786, and 0.722 for Overlap.…”

“…The features of (Aleixo and Pardo, 2008) were expanded with: (i) cosine similarity of word vectors, (ii) intersection of common words measured with the Jaccard Index, (iii) an indicator of longer sentence and (iv) one-sided word coverage ratio. Kumar et al (2012a) restricted the set of relations further down to four: Identity, Subsumption, Overlap and Elaboration. Four features were used: (i) tf-idf based cosine sentence similarity, (ii) words coverage ratio, (iii) sentence length difference and (iv) the indicator of longer sentence.…”