Beyond Instructions: A Taxonomy of Information Types in How-to Videos

Title of the Paper

Beyond Instructions: A Taxonomy of Information Types in How-to Videos

Paper Information

• Subject Areas: Human-Computer Interaction, Information Taxonomy, Video Content Analysis and Navigation
• Keywords: How-to Videos, Information Types, Video Content Analysis, Taxonomy, Video Navigation, Dataset, Human-Computer Interaction

Research Background and Problem

• Identified Problem/Challenge: How-to videos contain not only instructions for completing tasks but also various other forms of related information. However, these information types are scattered throughout the video, making it difficult for users to locate specific information they need. Existing video navigation techniques, often based on chapters or scripts, are insufficient for quickly accessing specific information.
• Significance: How-to videos are an important resource for users to learn new skills due to their detailed explanations. However, the complexity of the information and the unpredictable structure can lead to poor user experience. Enhancing the efficiency of accessing and locating relevant information is crucial.
• Motivation and Related Work: While some studies have explored the classification of lecture videos or software tutorial videos, research on the classification of information in how-to videos remains unexplored. The researchers aim to create a taxonomy for how-to videos to provide theoretical and practical support for improving tasks such as navigation, presentation, and creation of video content.

Solution

• Method/Solution: The authors developed a taxonomy to identify and organize the types of information present in how-to videos.
  • Data was sourced from 120 short videos covering various topics, selected from the HowTo100M dataset.
  • Through iterative open coding of 4,000 sentences from 48 videos, 21 information types were identified and categorized into 8 groups. The remaining 72 videos were used to validate the taxonomy.
• Innovations:
  • Proposed a comprehensive taxonomy of information types, systematically defining the diverse types of information present in how-to videos for the first time.
  • Provided an annotated dataset, HTM-Type, containing 120 videos with a total of 9.9k labeled sentences corresponding to the taxonomy.
• Implementation Steps and Techniques:
  • Selected appropriate samples, transcribed audio, and segmented sentences.
  • Developed the taxonomy iteratively using open coding, leveraging video context and textual content.
  • Analyzed the distribution of classified information in videos using statistical techniques such as the Kruskal-Wallis test, exploring the impact of task types and narrative styles on information type distribution.
  • Designed a video browsing interface based on the taxonomy as a research probe for user navigation functionality.

Research Outcomes

• Specific Outcomes:
  • Identified 21 information types grouped into 8 categories: Greeting, Overview, Method, Supplementary, Explanation, Description, Conclusion, and Miscellaneous.
  • Created the HTM-Type dataset, providing foundational data for automated classification based on the taxonomy.
  • Experimental results confirmed the strong applicability of the taxonomy, with significant effects of task type, video style, and goals on information type needs.
• Advantages Over Existing Solutions:
  • The taxonomy supports more granular video analysis and user navigation needs, overcoming the limitations of chapter-based navigation.
  • Users can quickly locate specific parts of a video based on particular information types of interest.
• Experiments/Evaluation Results:
  • "Method" information accounted for 47.5% of video duration, making it the dominant type, but the importance of other types, such as explanations and supplementary information, was also significantly highlighted.
  • User studies demonstrated that the taxonomy-based video navigation interface significantly improved task efficiency.
  • Attributes such as tool descriptions and warning information were shown to be indispensable for task-following.
• Limitations and Future Directions:
  • The current taxonomy is primarily text-based and does not fully integrate complex visual information. Future research could enhance the taxonomy through multimodal analysis.
  • The taxonomy is mainly applicable to medium-length (5 to 15 minutes) YouTube videos and may not fully adapt to long videos (e.g., live streams) or short video platforms (e.g., TikTok).
  • Further optimization of automated information classification pipelines is needed, along with exploration of broader applications of the taxonomy in video recommendation, analysis, and creation.