Case-Based Retrieval Interface Adapted to Customer-Initiated Dialogues in Help Desk Operations

Hideo Shimazu, Akihiro Shibata, Katsumi NiheiAAAI 1994

Abstract

Help desk operations often struggle with the "vocabulary problem," where customers and experts use different terminology to describe the same technical issues. This paper presents CARET (Case-based Retrieval Tool), a multi-modal interface that utilizes Case-Based Reasoning (CBR) to bridge this linguistic gap. By employing "schemata"—predefined templates representing procedures, logical diagrams, or physical layouts—the system allows operators to index and retrieve solutions based on "nearmisses," or minor deviations from standard operations. Evaluation on a real-world email service case-base demonstrates that this approach provides efficient retrieval (approx. 2.0 seconds) and high relevance for customer-initiated dialogues.

Introduction

Information retrieval in help desk environments is complicated by the discrepancy between a customer's mental model and the technical documentation of a product. A user might describe a failure chronologically, through a logical flow, or by physical symptoms. Traditional keyword-based search systems often fail to capture the intent or structural context of these varied descriptions. To address this, the authors propose a system that translates non-technical explanations into precise queries by mapping them onto structured domain knowledge.

Methodology: The CARET System

The core innovation of the CARET system is the use of three distinct types of Schemata to represent domain knowledge:

  1. Step-by-step Schema: Represents the chronological sequence of user actions (e.g., the steps to send an email).
  2. Diagram Schema: Represents the internal logical model or data flow of a product.
  3. Physical-appearance Schema: Provides a visual map of the product's interface or hardware components.

The "Nearmiss" Indexing Strategy

Rather than relying on complex keyword indexing, CARET indexes cases as deviations from these standard schemata. Most help desk inquiries arise from users who are performing a task correctly for the most part but encounter a single point of failure. This specific "nearmiss point" serves as a highly effective index for retrieving relevant past cases.

Figure 4: The nearmiss indexing strategy, showing how specific cases are linked to deviations in a standard "Send-Email" schema.

Technical Implementation and RDBMS Integration

CARET is designed to operate on top of standard Relational Database Management Systems (RDBMS), such as Oracle. The system architecture includes a Case Table and a Schema Table, linked via attribute hierarchies to the different schema types.

Figure 6: System architecture showing the integration of the three schema types with the RDBMS-based case-base.

To handle similarity-based retrieval within a standard SQL environment, CARET generates multiple SQL queries representing varying degrees of similarity. It utilizes a weighted nearest-neighbor algorithm to calculate similarity across attribute hierarchies and graphical control elements.

Evaluation and Results

The system was tested using several hundred customer inquiries related to a personal electronic mail service, sourced from NEC internal publications and support engineers.

Key findings include:

  • Retrieval Performance: The system achieved an average response time of 2.0 seconds for a database of 1,500 cases on a SUN Sparc Station 2. This was considered practically acceptable for real-time operations in 1994.
  • Schema Effectiveness: Subjective evaluations indicated that "how-to" (Goal-Plan) questions were most effectively retrieved via Diagram Schemata, whereas "Symptom-Cause-Recovery" questions were better handled by Step-by-step Schemata.

Discussion and Key Insights

The research highlights that indexing the deviation from a correct path is significantly more efficient than indexing the entire problem description. Furthermore, the system's ability to "bridge" different schemata—for instance, retrieving a case stored as a diagram using a procedural query—allows it to accommodate different user perspectives.

However, limitations remain. The schema-based approach is less effective for "context-free" inquiries, such as hardware failures (e.g., an unplugged cable) that have no logical link to the user's software actions. Additionally, the manual effort required by domain experts to define schemata and similarity weights represents a significant knowledge engineering overhead.

Conclusion

CARET demonstrates a practical application of CBR in corporate help desk environments by integrating with existing RDBMS technologies. By focusing on nearmiss points within structured schemata, it successfully mitigates the vocabulary problem inherent in customer-expert dialogues.

Summary updated

BibTeX
@inproceedings{shimazu1994case,
  author = {Hideo Shimazu and Akihiro Shibata and Katsumi Nihei},
  title = {{Case-Based Retrieval Interface Adapted to Customer-Initiated Dialogues in Help Desk Operations}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1994},
}
Figure