Topic Segmentation Sample Clauses

Topic Segmentation. For tuning and evaluating our approach, we selected ASR transcripts of any two segments, with the goal of retrieving the seam shot of these segments by automatic means. We restricted these sets by only allowing segments taken from the same day, in order to ensure that the topics of these segments are distinct (especially important for topics with a high longevity in this period, such as the delay in Berlin’s airport construction). Keyword extraction has been re-run on each of these double-segments. With an average of seven segments per day, all possible combinations of distinct segments gave 6830 testing instances. Of these, we used 5800 combinations as development set, and reserved 1030 combinations as test set.‌ 4.2.3.1 Approach Our approach for topic segmentation combines visual cues from Shot Segmen- tation with information from ASR. In video production cuts, dissolves and wipes are used to visualize shifts in time or space including topic shifts in news productions. Furthermore, ASR provides the basis to determine lexical cohesion. First, we consider repetitions of extracted keywords in different shots. Second, we indirectly include associated words (e.g. generalizations, statistical associations, etc.) in a similar way by analyzing Wikipedia links and extract associated words for each of the extracted named entities. After combining both sources of information, shot boundaries in areas of low cohesive strength are considered to indicate topic transitions. 4.2.3.2 Evaluation A qualitative analysis of the shots which were preferred as split points rather than the segment seam produced interesting results, including: (a) split points whenever an interview started, because the interviewee employed different wording, (b) split points when a news point was commented on a higher level, and (c) many split points in between brief news summary sessions that were merged as one segments. One recurring peculiarity were weather reports, which normally were split well if the weather remained constant but received a split point in-between if the forecast predicted a change in the upcoming weather (e.g. “tomorrow the sun finally shines again, bringing warm temperatures to the north. ”).‌ For this work package, this means that we should not rely on editorial split points as offered by content providers for our algorithms. Especially for news summaries, a user might be interested only in a few sub-topics that are covered there, but in order to adress these preferences,...