Assorted Textual Features and Dynamic Push Strategies for Real-time Tweet Notification

In this paper, we describe our systems and corresponding results submitted to the RealTime Summarization (RTS) track at the 2016 Text Retrieval Conference (TREC). The task involved identifying relevant tweets based on a user’s interest profile. In Scenario A of the task, tweets relevant to an interest profile were pushed to a live user in real-time. In Scenario B, a daily digest of relevant tweets was sent to a user. We submitted three automatic runs for each scenario. Our overall method for identifying relevant tweets was based on 1) automatically identifying key textual features from a set of interest profiles provided by the Track organizers, 2) expanding the textual phrases with their paraphrases, and 3) exploiting the features for message filtering and relevance measurement after novelty recognition. We experimented with different push strategies to decide when to deliver a message to a user. The evaluation results (by mobile and NIST assessors) show that our system ranked 3rd for Scenario A and 6th for Scenario B.