WaterlooClarke: TREC 2015 Temporal Summarization Track

The Temporal Summarization Track looks at providing meaningful summaries of major events and sub-events as they occur. Difficulties arise due to the unique nature of the temporal summarization task in which the corpora is constantly changing along with the known information about the event [1]. This year, the temporal summarization track consists of three tasks, two filtering and summarization tasks and an ontopic summarization task. In the filtering and summarization tasks, relevant content must be extracted from a continuous stream of documents and gradually summarized while maintaining low redundancy, latency and verbosity. In the third task, only the summarization must take place on a corpora filled with many documents related to the event [2]. The overall goal of the track is to provide a sequence of short meaningful sentence length updates over the duration of the events taking into account sentence redundancy and latency. The summaries are scored using an expected latency gain metric (ELG) in which sentences are rewarded for containing handpicked key updates (nuggets) and are penalized for redundancy, latency and verbosity [3]. This track can help users deal with all of the vast information that comes as major events progress through the use of general impersonal updates. Our approach for this track consists of a two step process. The first stage is a data preprocessing stage which extracts the information from the supplied web documents and only retains the information considered necessary for the second stage. The second stage is the filtering and summarization stage used to find and push the relevant sentence length updates. We have applied this methodology to both the second and third task. A total of six different runs have been submitted with all of them being judged. The run scores are comparable to those in 2014, however, with a larger latency gain and lower latency comprehensiveness. The submitted task 2 runs appear to have a slightly better performance than those of task 3. However, there were no apparent differences between the task 2 or the task 3 runs.