In the mid-1990s, artificial intelligence transitioned from theoretical laboratory exercises to robust, industrial-scale applications. This era was defined by the "VIDEOS" track at major conferences like IJCAI 1995, which showcased systems capable of handling multi-billion dollar operations and massive text corpora. This paper examines two such milestones: the CLASSIC knowledge representation system for product configuration and the innovative Scatter/Gather and TileBars interfaces for information retrieval.
The Configuration Challenge: Moving Beyond Flat Rules
Configuring complex systems—such as telecommunications hardware or high-end home theaters—presents a significant challenge for traditional rule-based AI. These systems are inherently hierarchical and interconnected; a single choice (e.g., selecting a high-definition display) triggers a cascade of requirements for cables, power supplies, and compatible receivers.
The CLASSIC system, based on Description Logic (DL), addressed these issues by moving beyond simple "if-then" logic. It utilized:
- Taxonomic Reasoning: Organizing components into a formal hierarchy (e.g., a "TV" is a type of "Electronic Component").
- Active Inference: Automatically deducing necessary components based on user-defined constraints, such as requiring a specific screen size for a "high-quality" designation.
- Explanation Facility: Providing transparency by allowing users to query why a specific component was added or why a configuration error occurred.
Visualizing Information: Scatter/Gather and TileBars
As digital text collections expanded to the gigabyte scale (notably the TIPSTER corpus), users faced a "vocabulary problem." Standard search engines returned lists of titles that offered little insight into the document's internal structure or the context of search terms.
To solve this, researchers introduced two complementary paradigms:
- Scatter/Gather: This tool clusters a collection into topical groups ("scatters"). Users select interesting clusters ("gather") to form a new sub-collection, iteratively refining their search through browsing rather than just querying.
- TileBars: This visualization uses "TextTiling" to segment documents. A grid (TileBar) represents each document, where the darkness of a square indicates the frequency of query terms in that segment. This allows users to see if terms appear together or are scattered, providing essential spatial context.
Impact and Real-World Results
The deployment of CLASSIC at AT&T demonstrated the immense commercial potential of Description Logic. The system was used to process over $2.5 billion in orders, significantly reducing processing intervals and personnel requirements while eliminating data duplication.
In the realm of Information Retrieval, the Scatter/Gather and TileBars methods transformed search from a "black box" into a transparent, iterative process. TileBars, in particular, allowed users to filter out documents with only passing references to a topic, focusing instead on those where search terms overlapped in meaningful ways.
Conclusion
These systems proved that logic-based AI and advanced visualization could solve high-stakes industrial problems. While challenges remained—specifically the "knowledge engineering bottleneck" required to build initial ontologies—the success of CLASSIC and the IR interfaces laid the groundwork for modern automated configuration and visual search technologies.
Summary updated