First Story Detection using Entities and Relations

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 3237–3244, Osaka, Japan, December 11-17 2016. First Story Detection using Entities and Relations Nikolaos Panagiotou National and Kapodistrian University of Athens npanagio@di.uoa.gr Cem Akkaya Yahoo! Research, Sunnyvale, CA cakkaya@yahoo-inc.com Kostas Tsioutsiouliklis Yahoo! Research, Sunnyvale, CA kostas@yahoo-inc.com Vana Kalogeraki Athens Univerisity of Economics and Business vana@aueb.gr Dimitrios Gunopulos National and Kapodistrian University of Athens dg@di.uoa.gr Abstract News portals, such as Yahoo News or Google News, collect large amounts of documents from a variety of sources on a daily basis. Only a small portion of these documents can be selected and displayed on the homepage. Thus, there is a strong preference for major, recent events. In this work, we propose a scalable and accurate First Story Detection (FSD) pipeline that identifies fresh news. In comparison to other FSD systems, our method relies on relation extraction methods exploiting entities and their relations. We evaluate our pipeline using two distinct datasets from Yahoo News and Google News. Experimental results demonstrate that our method improves over the state-of-the-art systems on both datasets with constant space and time requirements. 1 Introduction First Story Detection (FSD) is the task of detecting the first document about a new event given a stream of docu- ments (Allan et al., 1998; Allan et al., 2000a). The task is also known as New Event Detection (NED). The problem appears in several real-world applications where news stories are accumulated and presented to users in near real- time. A FSD system should be accurate, scalable, and process a stream of articles in a single pass. The output of such a system is very valuable to news portals such as Yahoo News, Google News, since the rapid detection of a new event is crucial for the service reputation. FSD is a very challenging, if not impossible, task for human operators, since it requires inspecting millions of documents per day. As a result, accurate, automatic solutions are very desirable. The majority of the FSD systems in the literature attempt to classify a document as a first story if the document differs significantly from those published before and thus may describe a new event. This is accomplished usually in two steps. In the first step, the nearest neighbor of a new document in the previous document stream is identified. In the second step, the similarity between the new document and its nearest neighbor is considered in order to decide if it is a first story or not. In this methodology, the selection of an appropriate similarity metric and the selection of an informative document representation are essential. As a result many different metrics have been studied in the literature, such as the cosine similarity (Petrovi´c et al., 2012), the KL-Divergence (Karkali et al., 2013), and the named entities overlap (Kumaran and Allan, 2004). So far, terms and entities occurring in a document have been the main representation units, boiling down the task to detecting documents with novel terms and entities in a …

Content truncated to 500 words