A Conversational Approach to Process-oriented Case-based Reasoning

Christian Zeyen, Gilbert Müller, Ralph BergmannIJCAI 2018 · DOI

TL;DR

This paper introduces C-POCBR, a novel conversational approach to process-oriented case-based reasoning (POCBR) that automatically generates questions to help users find existing workflows. By reducing the need for users to manually craft complex queries, C-POCBR aims to significantly lower the communication effort and expertise required for workflow retrieval.

What Problem Does This Paper Address?

Process-oriented case-based reasoning (POCBR) helps users find and adapt existing workflows (like recipes or business processes) that have been successful in the past. However, a major hurdle is that current POCBR systems typically require users to define very detailed queries, often in specialized languages. This is challenging because users might not have a clear idea of what they're looking for or lack the domain knowledge to articulate it precisely. This paper addresses this gap by proposing a conversational system that interactively guides users through the retrieval process, making it more accessible and less demanding.

Methods

The core of the C-POCBR approach lies in its ability to automatically generate questions. This is achieved through an "offline phase" where the system analyzes the existing workflow cases. Key steps include:

  • Feature Extraction: Workflows are represented as semantically labeled directed graphs. The system extracts "features" from these graphs, which can be as simple as individual nodes (tasks or data) or more complex partial workflows. These features are then generalized to allow for broader queries.
  • Feature Ranking: Features are ranked based on their ability to distinguish between different workflows. The "similarity variance" (simVar) measure is used, prioritizing features that, when known, most effectively narrow down the set of potential matching workflows.
  • Conversational Dialog (Online Phase): When a user initiates a search, the system engages in a dialog. It presents questions based on the ranked features. Users can respond by:
  • Ignoring a question: This helps prune irrelevant features.
  • Answering a question: The user indicates whether a feature is desired or undesired, which updates the internal query and refines the set of candidate workflows.
  • Excluding a suggested workflow: This temporarily removes a workflow from consideration and triggers a re-evaluation.
  • Selecting a solution: The user identifies the desired workflow, terminating the search.

The system employs different question types (feature, specialization, enlargement) to progressively elicit more specific information from the user.

Datasets

The paper uses a case base of 61 cooking workflows derived from the existing Cooking-CAKE system. For evaluation, 60 search scenarios were created, each describing a target workflow in plain text.

Key Insights and Ideas

  • Conversational interaction significantly lowers the barrier to entry for complex retrieval tasks. By asking targeted questions, the system compensates for users' potential lack of detailed knowledge or inability to formulate precise queries.
  • Automatic question generation is crucial for making conversational systems practical. Manually creating questions for every case would be prohibitively time-consuming. The paper demonstrates that features extracted from workflow structures can serve as effective subjects for these questions.
  • Feature ranking based on distinguishing power is key to efficient dialogs. Prioritizing features that most effectively reduce uncertainty leads to quicker convergence on the desired workflow.
  • The hierarchical nature of features (specific vs. generalized) and their relationships are vital for intelligent dialog management. This allows the system to infer irrelevant features and guide the conversation effectively.

Key Results

  • Hypothesis H1 (Desired workflow retrieval): Fully confirmed. The simulated user, answering questions correctly, always retrieved the target workflow.
  • Hypothesis H2 (User retrieval success): Confirmed. Human users successfully retrieved desired workflows in most cases (14 out of 16 conversations).
  • Hypothesis H3 (Reduced communication effort): Partially confirmed. While total conversation time was comparable to traditional POCBR, the C-POCBR approach required significantly fewer questions (average 9 vs. 30). The authors note that with better workflow presentation and explanation, C-POCBR could be even faster.

Limitations and Future Work

  • Workflow Presentation: The authors acknowledge that the current presentation of retrieved workflows in C-POCBR could be improved to help users identify them more quickly. Additional explanation functions are suggested.
  • Domain Complexity: The evaluation was conducted on cooking workflows. The authors suggest evaluating the approach in domains with more complex workflows, where conversational retrieval is expected to outperform traditional methods even more significantly.
  • Adaptation Integration: Future work could explore integrating interactive adaptation of workflows with the interactive retrieval process to create a more powerful problem-solving tool.
  • Broader Applicability: The questioning strategy could be extended to other CCBR applications with complex case representations, provided feature hierarchies and co-occurrences are identified.

Summary updated

BibTeX
@inproceedings{zeyen2018conversational,
  author = {Christian Zeyen and Gilbert Müller and Ralph Bergmann},
  title = {{A Conversational Approach to Process-oriented Case-based Reasoning}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2018},
  doi = {10.24963/ijcai.2018/762},
}
Figure