Sentence Fusion via Dependency Graph Compression

Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pages 177–185, Honolulu, October 2008. c⃝2008 Association for Computational Linguistics Sentence Fusion via Dependency Graph Compression Katja Filippova and Michael Strube EML Research gGmbH Schloss-Wolfsbrunnenweg 33 69118 Heidelberg, Germany http://www.eml-research.de/nlp Abstract We present a novel unsupervised sentence fu- sion method which we apply to a corpus of bi- ographies in German. Given a group of related sentences, we align their dependency trees and build a dependency graph. Using integer lin- ear programming we compress this graph to a new tree, which we then linearize. We use GermaNet and Wikipedia for checking seman- tic compatibility of co-arguments. In an eval- uation with human judges our method out- performs the fusion approach of Barzilay & McKeown (2005) with respect to readability. 1 Introduction Automatic text summarization is a rapidly develop- ing field in computational linguistics. Summariza- tion systems can be classified as either extractive or abstractive ones (Sp¨arck Jones, 1999). To date, most systems are extractive: sentences are selected from one or several documents and then ordered. This method exhibits problems, because input sentences very often overlap and complement each other at the same time. As a result there is a trade-off between non-redundancy and completeness of the output. Al- though the need for abstractive approaches has been recognized before (e.g. McKeown et al. (1999)), so far almost all attempts to get closer to abstractive summarization using scalable, statistical techniques have been limited to sentence compression. The main reason why there is little progress on ab- stractive summarization is that this task seems to re- quire a conceptual representation of the text which is not yet available (see e.g. Hovy (2003, p.589)). Sen- tence fusion (Barzilay & McKeown, 2005), where a new sentence is generated from a group of related sentences and where complete semantic and con- ceptual representation is not required, can be seen as a middle-ground between extractive and abstrac- tive summarization. Our work regards a corpus of biographies in German where multiple documents about the same person should be merged into a sin- gle one. An example of a fused sentence (3) with the source sentences (1,2) is given below: (1) Bohr Bohr studierte studied atan der the Universit¨at University Kopenhagen Copenhagen und and erlangte got dort there seine his Doktorw¨urde. PhD ’Bohr studied at the University of Copenhagen and got his PhD there’ (2) Nach After dem the Abitur school studierte studied er he Physik physics und and Mathematik mathematics atan der the Universit¨at University Kopenhagen. Copenhagen ’After school he studied physics and mathemat- ics at the University of Copenhagen’ (3) Nach After dem the Abitur school studierte studied Bohr Bohr Physik physics und and Mathematik mathematics atan der the Universit¨at University Kopenhagen Copenhagen und and erlangte got dort there seine his Doktorw¨urde. PhD ’After school Bohr studied physics and mathe- matics at the University of Copenhagen and got his PhD there’ 177 Having both (1) and (2) in a summary would make it redundant. Selecting …

Content truncated to 500 words