Personalcasting: Tailored Broadcast News

  title={Personalcasting: Tailored Broadcast News},
  author={Mark T. Maybury and Warren R. Greiff and Stanley Boykin and Jay M. Ponte and Chad McHenry and Lisa Ferro},
  journal={User Modeling and User-Adapted Interaction},
Broadcast news sources and newspapers provide society with the vast majority of real-time information. Unfortunately, cost efficiencies and real-time pressures demand that producers, editors, and writers select and organize content for stereotypical audiences. In this article we illustrate how content understanding, user modeling, and tailored presentation generation promise personalcasts on demand. Specifically, we report on the design and implementation of a personalized version of a… CONTINUE READING

17 Figures & Tables

Extracted Numerical Results

  • Video news segmentation performance ranges from 50 to 80% balanced precision and recall. In particular, segmentation algorithms using multimodal cues and trained on a range of broadcast sources such as CNN, MS-NBC or ABC perform with 53% precision and 78% recall (Boykin and Merlino 1999). Broadcast specific models (e.g., ones using visual anchor booth recognition cues specific to a particular program such as ITN) raise the performance to 96% precision and recall.
  • For example, using 20 users performing relevance assessments and information extraction tasks, we demonstrated that users exhibit over 90% precision and recall using displays such as those in Figure 11b in less than half the time required to search digital video sequentially.
  • Thus, the precision on the general term “Iraq” is 20/20 or 100%.
  • Since the user provides only a single document for relevance feedback and the words “Tariq” and “Aziz” appear near the end of a twenty term expansion list, the precision performance of this feedback is only 5% (third column of Table 1).
  • As shown in the fourth column of Table 1, when the user selects either the terms “Tariq” or “Aziz” from the term expansion list, the system returns exactly five documents that pertain to the user’s original information need, thus achieving a precision of 5/5 or 100%.
  • When the user selects specific concrete terms such “gaza” or “hamas”, precision rises to 95%.
  • When the user selects those documents and requests similar ones, the precision rises to 100%.
  • If they select the term “shooting”, the precision of the returned document set is 18/20 or 90% (two irrelevant documents are returned about a shooting of a marine and a U.N. protester shooting).
  • A baseline version of the system over a range of broadcast sources (e.g., CNN, MS-NBC, and ABC) performed segmentation on average with 38% precision and 42% recall across all multimodal cues (i.e., textual, audio, and visual cues). In contrast, performance for the best combination of multimodal cues rose to 53% precision and 78% recall.
  • , the performance rises to 96% precision and recall.