Capturing User Intent for Information Retrieval

Hien NguyenAAAI 2004

Abstract

Traditional Information Retrieval (IR) systems often fail to account for the cognitive state of the user, treating search as a static transaction rather than a dynamic process of discovery. This paper proposes the IPC model, a cognitive framework that decomposes user intent into three distinct components: Interests, Preferences, and Context. By leveraging Bayesian Networks and Decision Theory, the system dynamically captures these elements from search behavior to proactively refine queries and improve retrieval performance.

Introduction: The Cognitive Gap in IR

Standard IR frameworks typically operate on a simple input-output loop: a user provides a query, and the system returns documents based on keyword matching. However, this ignores the complex cognitive reality of the user. The author identifies four major hurdles that traditional systems fail to address:

  1. Non-observability: The system cannot directly observe the user's internal knowledge state.
  2. Uncertainty: User interactions and feedback are often ambiguous.
  3. Vagueness: Users frequently struggle to articulate their exact information needs in a formal query.
  4. Dynamics: A user’s knowledge and needs evolve as they interact with and learn from retrieved information.

The IPC model aims to bridge this gap by providing a structured way for the system to involve itself in the user's cognitive process, moving from literal keyword matching to intent-based retrieval.

The IPC User Model

The core of this research is the decomposition of intent into three functional modules:

1. Context (C)

The Context component answers why the user is searching. It is represented as a Directed Acyclic Graph (DAG) of concepts and relations. Rather than relying on a static profile, the Context is built dynamically by finding intersections among documents that the user marks as relevant during the session.

2. Interests (I)

Interests represent what the user cares about. This module consists of a set of concepts and their associated "interest levels." These levels are derived by applying a spreading activation algorithm over the Context network, allowing the system to identify related topics that the user might not have explicitly mentioned but are likely relevant to their goal.

3. Preferences (P)

Preferences answer how the search should be conducted. Modeled using a Bayesian Network (BN), this component captures the user’s preference for specific query-modification tools. For example, it tracks whether a user prefers the system to expand their query with synonyms or constrain it with more specific technical terms.

By combining these three elements, the system can proactively modify the user's original query graph, adding relevant concepts and relations to disambiguate the search and align it more closely with the user's actual intent.

Experimental Evaluation

The IPC model was evaluated in two distinct phases. The initial study utilized a specialized medical database containing 521 records and 35 queries. In this controlled environment, the IPC model achieved precision and recall scores that were at least as high as, and in several cases superior to, the system operating without the user model.

However, pilot usability testing highlighted a significant challenge: system transparency. Users often found the system's proactive query modifications "opaque," as they did not always understand why the system was changing their search terms. This lack of interpretability occasionally hindered task efficiency despite the improved retrieval metrics.

Future Work and Scalability

To address the limitations identified in the initial study, the author outlines several paths for future development:

  • Decision-Theoretic Framework: The paper proposes a "best move" logic. By treating the IR process as a series of decisions, the system could use a decision tree to predict whether it should expand a query, filter results, or ask the user for more feedback based on the current IPC state.
  • Interpretability: Future iterations will focus on making the system's actions more transparent to the user, ensuring they understand the rationale behind automated query refinements.
  • Large-Scale Benchmarking: To prove the model's robustness, the author plans to test the IPC framework against classic IR "gold standard" collections, including CRANFIELD, NPL, MEDLINE, CACM, and the TREC disks.

Conclusion

The IPC model represents a significant step toward bridging the gap between Artificial Intelligence and Information Retrieval. By using formal knowledge representation techniques like Bayesian Networks and spreading activation, the system moves beyond static keyword matching toward a more nuanced, cognitive understanding of user intent.

Summary updated

BibTeX
@inproceedings{nguyen2004capturing,
  author = {Hien Nguyen},
  title = {{Capturing User Intent for Information Retrieval}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2004},
}
Figure