Abstract
Building large-scale hypermedia systems, specifically "ASK systems" designed to simulate expert conversations, faces a significant bottleneck known as indexer saturation. As databases grow, human editors struggle to maintain a coherent web of links across thousands of potential connections. This paper proposes a semi-automated indexing framework that combines human-provided "partial representations" of story content with machine-driven inference rules and concept hierarchies to suggest relevant links, ensuring system scalability and conversational coherence.
Introduction: The Challenge of Indexer Saturation
In the early 1990s, researchers at the Institute for the Learning Sciences pioneered ASK systems—multimedia databases that allow users to "converse" with experts by navigating through related video and text segments. The effectiveness of these systems depends on a dense network of links that follow conversational logic (e.g., following a story about a "failed experiment" with one about "alternative approaches").
However, as these systems exceed 100 stories, they encounter indexer saturation. A human knowledge engineer can no longer recall every story well enough to identify all relevant connections for new content. Manual searching becomes prohibitively expensive, while fully automated natural language processing of the era remained too brittle for complex narrative understanding. The authors propose a "middle way": a machine-assisted indexing partnership where humans extract key features and machines propose potential links.
Methodology: Frames, Hierarchies, and Inference
The proposed framework relies on three core pillars to bridge the gap between raw text and structured links.
1. The Partial Representation (Frames)
Instead of full-text analysis, indexers populate a domain-independent "frame" for each story scene. These frames capture "naive intentionality"—the underlying motivations and outcomes of the story's agents.
Key slots include:
- AgentRole: The role of the primary actor (e.g., Student, Athlete).
- IntentionType: The specific goal or action (e.g., "get good grades").
- OutcomeTypes: The result of the action (e.g., success or failure).
- SituationType: The broader context (e.g., "conflict with others").
2. Concept Hierarchies
Fillers for these slots are organized into taxonomies rather than flat keyword lists. This hierarchical structure allows the system to recognize relationships between stories that use different but related terminology.
For instance, "getting good grades" is categorized under "graduating." By traversing these hierarchies, the system can identify that two stories are conceptually related even if their specific actions differ.
3. Inference Procedures
The system utilizes rules based on the ASK model of hypermedia, which organizes links into eight "poles" across four dimensions: Refocusing (Context/Specifics), Causality (Causes/Results), Comparison (Analogies/Alternatives), and Advice (Opportunities/Warnings). By analyzing the frames and their positions in the concept hierarchies, the system can propose specific link types. For example, it might suggest a "Comparison/Alternative" link between two stories where agents share the same goal but employ different strategies.
Evaluation and Insights
The primary testbed for this approach was School Stories, a hypermedia environment documenting K-12 educational experiences.
Key Results
- Superset Performance: In an evaluation of 16 randomly selected stories, the automated system identified a superset of the links previously found by human indexers, suggesting that the machine-assisted approach can effectively combat link sparsity.
- Precision in "Near-Misses": The system successfully distinguished between similar actions based on their underlying intentionality. It could differentiate between a student leaving class due to boredom versus leaving due to a broken promise by the teacher, assigning different link types accordingly.
The Human-in-the-Loop Philosophy
The authors emphasize that the goal is not total automation. Humans remain superior at feature extraction (filling the frames) and final relevance judgment. The machine's role is to manage the massive search space and maintain consistency, significantly reducing the cognitive load on the human indexer.
Limitations and Future Directions
While promising, the system's performance is heavily dependent on the quality of the hand-crafted concept hierarchies and the consistency of the human indexers. Future work aims to move toward dynamic link generation, where the system creates links in real-time based on a user's specific browsing path, rather than relying on a static pre-indexed network.
Summary updated