This past March we demonstrated how Google's AI Workshop "Semantic Similarity For Natural Language" experiment could be used to segment an evening television news show into discrete stories. To explore this approach further, we've applied the exact same approach to another show, ABC World News Tonight With David Muir ABC on February 1, 2020 4:00pm-4:30pm PST. You can see how the Vanderbilt Television News Archive's human indexers have segmented the show.
Once again we split the show into lines of captioning and compared every two lines of captioning with the previous two lines of captioning. The final results spreadsheet is a list of airtime offsets with the probability that each pair of lines ending at that time offset were entailed by the previous two lines. We went in by hand and added in the story transition point for each of the major stories in the program, which can be compared with the probability scores. With a bit of filtering it is clear that the probability scores offer a powerful indicator of segmentation points!