TAXI: A Taxonomic Assistant

Thomas Y. GallowayAAAI 1987

Introduction

TAXI (Taxonomic Assistant) is a "Knowledge Manipulation System" (KMS) designed to streamline the construction and exploration of complex taxonomies. Developed to address the bottlenecks in knowledge base construction common in the late 1980s, TAXI utilizes a discrimination tree structure to automate the logistical "bookkeeping" of data organization. By handling the repetitive tasks of sorting and dependency tracking, the system allows human researchers to focus on the high-level conceptual work of defining domain relationships.

The Challenge of Knowledge Engineering

Constructing large-scale knowledge bases, particularly for diagnostic systems, was historically a daunting task. The author identifies a fundamental tension in the field: while taxonomies are essential for organizing declarative knowledge, building them manually is both error-prone and tedious. Conversely, fully automating the process is often impossible because it requires the very domain knowledge the system is intended to capture.

Furthermore, the paper argues that total automation is counterproductive to the human expert's growth. The process of manually organizing data often leads to "aha!" moments and a deeper understanding of the domain. TAXI was designed to resolve this tension by acting as an assistant that handles the "grunge work"—such as re-sorting data when definitions change—while keeping the human expert in control to facilitate active learning.

System Methodology and Features

TAXI is built upon a discrimination tree representation. The author chose this structure over more complex directed acyclic graphs (such as KL-ONE) because a tree ensures each node has only one parent, making the hierarchy significantly easier for humans to visualize and comprehend.

Key features of the system include:

  • Attribute Aid: Based on "Personal Construct Theory," this tool presents the user with three random objects and asks them to define an attribute that distinguishes one from the other two. This method helps users discover latent attributes they might have otherwise overlooked.
  • Automated Reclassification: One of TAXI's most powerful features is its ability to handle structural changes. If a user modifies an attribute or a "type" definition at a high level, the system automatically ripples those changes down the tree, re-sorting all affected objects without manual intervention.
  • Heuristic Suggestions: To assist in tree construction, the system can suggest the "best" attribute to use for the next level of classification. Users can request an even distribution to create a balanced, efficient tree, or a skewed distribution to quickly isolate specific terminal nodes for diagnostic purposes.
  • Meta-Types: The system supports both numeric and non-numeric data, allowing users to define specific ranges (e.g., 1 to 100) to maintain data integrity across the taxonomy.

The Philosophy of the "Assistant"

A central thesis of the TAXI project is that the construction of a taxonomy is a primary way for a human to understand a domain. By participating in decisions regarding attribute order and value assignment, the user gains a deeper intuitive grasp of the data. TAXI is not intended to be an autonomous agent; it is a tool meant to augment human intelligence by removing the cognitive load of manual data management.

This approach also helps identify "confusability" within a dataset. In diagnostic contexts, TAXI allows engineers to see which objects (such as different diseases) consistently end up in the same or adjacent classes. This identifies areas where the data is ambiguous, prompting the engineer to develop specialized procedures to distinguish between similar cases.

Results and Future Directions

The initial implementation of TAXI consists of approximately 40 integrated tools for construction, editing, and exploration. Qualitatively, the system has proven successful in automating the trial-and-error nature of taxonomy construction. By allowing users to quickly swap the order of discriminators, they can view the same domain from multiple perspectives—for example, re-sorting a biological database by "habitat" rather than "morphology" in seconds.

While the current version of TAXI provides significant efficiency gains, the author notes several areas for future development:

  • Multi-Value Support: Currently, objects are limited to a single value per attribute.
  • Advanced Representations: Future iterations may move beyond simple discrimination trees to support more sophisticated schemes like lattices.
  • Comparative Tools: There is a recognized need for tools that can compare two different taxonomies of the same domain side-by-side to evaluate their relative logic and efficiency.

In conclusion, TAXI represents a shift toward collaborative knowledge engineering, where the machine handles the structural integrity of the data while the human provides the conceptual insight.

Summary updated

BibTeX
@inproceedings{galloway1987taxi,
  author = {Thomas Y. Galloway},
  title = {{TAXI: A Taxonomic Assistant}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1987},
}
Figure