Concept-Map-Based Multi-Document Summarization using Concept Coreference Resolution and Global Importance Optimization

Proceedings of the The 8th International Joint Conference on Natural Language Processing, pages 801–811, Taipei, Taiwan, November 27 – December 1, 2017 c⃝2017 AFNLP Concept-Map-Based Multi-Document Summarization using Concept Coreference Resolution and Global Importance Optimization Tobias Falke, Christian M. Meyer, Iryna Gurevych Research Training Group AIPHES and UKP Lab Department of Computer Science, Technische Universit¨at Darmstadt https://www.aiphes.tu-darmstadt.de Abstract Concept-map-based multi-document sum- marization is a variant of traditional sum- marization that produces structured sum- maries in the form of concept maps. In this work, we propose a new model1 for the task that addresses several issues in previous methods. It learns to identify and merge coreferent concepts to reduce redundancy, determines their importance with a strong supervised model and finds an optimal summary concept map via in- teger linear programming. It is also com- putationally more efficient than previous methods, allowing us to summarize larger document sets. We evaluate the model on two datasets, finding that it outperforms several approaches from previous work. 1 Introduction Concept-map-based multi-document summariza- tion (MDS) is a variant of traditional MDS that produces structured summaries in the form of a concept map instead of a coherent text (Falke and Gurevych, 2017a). A concept map, introduced by Novak and Gowin (1984), is a labeled graph show- ing concepts as nodes and relations between them as edges. As an example, consider a document col- lection discussing treatments for ADHD. A (very) small concept map would be Caffeine ADHD reduces in which Caffeine and ADHD are concepts, while reduces is a relation, forming the proposition “Caffeine – reduces – ADHD”. A summary in this form has interesting appli- cations, as it provides a concise overview of a 1Source code available at https://github.com/ UKPLab/ijcnlp2017-cmaps document collection, structures it across document boundaries and can be used as a table-of-contents to navigate in the collection. Several studies report successful applications of concept maps in this di- rection (Carvalho et al., 2001; Briggs et al., 2004; Richardson and Fox, 2005; Villalon, 2012; Valerio et al., 2012; Falke and Gurevych, 2017b). The task we consider in this work is defined as follows: Given a set of documents on a certain topic, extract a concept map that represents the most important content on that topic, satisfies a specified size limit and is connected. Although work dealing with the automatic ex- traction of concept maps from text exists (§2), current methods have several limitations. First, most approaches do not attempt to detect coref- erences between extracted concepts. For instance, if both ADHD symptoms and symptoms of ADHD are found, they treat them as separate concepts. In a concept map, such duplicate concepts are imme- diately visible to a user, waste valuable space and make it harder to look for relations of that concept, as they are spread among the duplicates. Second, previous work mostly focused on the extraction of concepts and relations, largely ig- noring the subsequent selection step necessary to produce a summary of manageable size. Existing studies suggested only a few unsupervised met- rics …

Content truncated to 500 words