Abstract

Television is changing. Increasingly, broadcasts are consumed interactively. This allows broadcasters to provide consumers with additional background information that they may bookmark for later consumption. To support this type of functionality, we consider the task of linking a textual streams derived from live broadcasts to Wikipedia. While link generation has received considerable attention in recent years, our task has unique demands that require an approach that needs to (i) be high-precision oriented, (ii) perform in real-time, (iii) work in a streaming setting, and (iv) typically, with a very limited context. We propose a learning to rerank approach that significantly improves over a strong baseline in terms of effectiveness and whose processing time is very short. We extend this approach, leveraging the streaming nature of the textual sources that we link by modeling context as a graph. We show how our graph-based context model further improves effectiveness. For evaluation purposes we create a dataset of segments of television subtitles that we make available to the research community.