Cluster-based Web Summarization

International Joint Conference on Natural Language Processing, pages 1124–1128, Nagoya, Japan, 14-18 October 2013. Cluster-based Web Summarization Yves Petinot and Kathleen McKeown and Kapil Thadani Department of Computer Science Columbia University 503 Computer Science Building 1214 Amsterdam Avenue New York, New York, 10027 {ypetinot|kathy|kapil}@cs.columbia.edu Abstract We propose a novel approach to abstrac- tive Web summarization based on the ob- servation that summaries for similar URLs tend to be similar in both content and structure. We leverage existing URL clus- ters and construct per-cluster word graphs that combine known summaries while ab- stracting out URL-specific attributes. The resulting topology, conditioned on URL features, allows us to cast the summariza- tion problem as a structured learning task using a lowest cost path search as the de- coding step. Early experimental results on a large number of URL clusters show that this approach is able to outperform previ- ously proposed Web summarizers. 1 Introduction Abstract Web summaries, which describe the top- ics and functionalities of Web pages at an abstract level, play an essential part in the discovery of new sites and services on the Web. Such summaries are intrinsically difficult to obtain using the content of Web pages. As such, the most successful meth- ods for generating them are effectively extractive and based on the identification of likely abstractive content from linking pages. These URL-centric techniques however require a significant amount of redundancy in linking content (Delort et al., 2003). In this paper, we propose a summary-centric ap- proach to Web summarization based on the ob- servation that summaries for similar URLs exhibit both similar content and structure. This simi- larity is apparent when analyzing summaries from a single ODP category, examples of which are shown in Table 1. We can see there that sum- maries of semantically related URLs tend to share common concepts and differ mostly at the level of target-specific attributes. Given a previously un- seen URL U, such a cluster could thus be used as a source of potentially relevant terms for that URL’s summary. In particular, these relevant terms in- clude abstract terms, which may not otherwise be observed in the input data. We propose a graph- based summarization framework that can leverage this phenomenon. 2 Proposed Framework Given a reference cluster C, we represent the space for summary generation as a graph GC = (VC, EC). This graph is obtained by fusing train- ing summaries in C into a word graph. Each sum- mary gi is mapped to a path between the shared source and sink nodes. Each word gj i is thus mapped to a node Nk and each pair of neighboring words (gj i , gj+1 i ) to a directed edge (Nk, Nl). No- tably, we add nodes as needed to guarantee that in- dividual summaries are cycle-free. Figure 1 shows a simple summary graph. 2.1 Node Alignment During the graph construction, nodes from distinct paths are iteratively combined to elicit the struc- tural and content commonalities between sum- maries in the reference …

Content truncated to 500 words