Sequential Clustering and Contextual Importance Measures for Incremental Update Summarization

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 1071–1082, Osaka, Japan, December 11-17 2016. Sequential Clustering and Contextual Importance Measures for Incremental Update Summarization Markus Zopf, Eneldo Loza Menc´ıa, and Johannes F¨urnkranz Research Training Group AIPHES / Knowledge Engineering Group Department of Computer Science, Technische Universit¨at Darmstadt Hochschulstraße 10, 64289 Darmstadt, Germany {zopf@aiphes,eneldo@ke,juffi@ke}.tu-darmstadt.de Abstract Unexpected events such as accidents, natural disasters and terrorist attacks represent an informa- tion situation where it is crucial to give users access to important and non-redundant information as early as possible. Previous work uses either a fast but inaccurate pipeline approach or a pre- cise but slow clustering approach. Instead, we propose to use sequential clustering for grouping information so that we are able to publish sentences at each time step. By doing so, we combine the best of both clustering and pipeline approaches and create a fast and precise real-time sys- tem. Experiments on the TREC Temporal Summarization 2015 shared task dataset show that our system achieves better results compared to the state-of-the-art. 1 Introduction Events such as accidents, natural disasters and terrorist attacks provide an important and challenging information problem. Shortly after such an event has occurred, the information situation is usually unclear. Initially, only vague information about the event may become available, for example that an earthquake has occurred, but details such as magnitude, epicenter, and whether a tsunami has to be expected are not known at this early point in time. Such information becomes only available as the event develops over time. In such situations, it is crucial for responders, crisis management organizations, and victims to get information as soon as possible. However, it is impossible for humans to monitor vast amount of textual information contained in sources such as news articles, tweets, social media posts, forum discussions, and live blogs. In incremental update summarization (IUS) (McCreadie et al., 2014b) the detection of important information in a timely manner is a major challenge. Since the information sources are also highly redundant, detecting and filtering redundancy is a second important challenge in IUS. In the past, two types of systems were used for IUC: pipeline systems (McCreadie et al., 2014a; Mc- Creadie et al., 2015) and ordinary clustering systems (Kedzie et al., 2014; Zhao et al., 2014; Yingzhe Yao and Fan, 2015; McCreadie et al., 2015). Pipeline systems process document or sentence and decide about importance of information and novelty immediately and achieve therefore a good timeliness. However, they publish a large amount of unimportant information (low precision) or miss important information (low recall) since they do not make use of the redundancy of information as signal for importance, which is a usually a very good feature in newswire documents (Nenkova et al., 2006). Traditional clustering systems on the other hand are able to exploit redundancy and therefore produce better summaries accord- ing to precision and recall. Since they collect documents (for example all documents within one hour) before they decide whether or not to publish …

Content truncated to 500 words