Letizia: An Agent That Assists Web Browsing

IJCAI 1995

TL;DR

This paper introduces Letizia, a user interface agent that assists web browsing by autonomously exploring links in the background and proactively suggesting relevant content to the user. It aims to alleviate information overload by inferring user interests from their browsing behavior, offering a more intuitive and less interruptive approach than traditional search engines.

What Problem Does This Paper Address?

The explosive growth of the World Wide Web has created a significant challenge for users trying to find relevant information. Traditional methods often require users to explicitly formulate queries and wait for results, interrupting their browsing flow. Existing solutions like automated search programs or web crawlers can be cumbersome, demanding user intervention and potentially leading to idle waiting periods. The paper addresses the need for an intelligent assistant that can seamlessly integrate with the user's browsing experience, anticipating their needs and reducing the cognitive load of navigating vast amounts of information.

Methods

Letizia operates as a "behavior-based interface agent," meaning it learns and makes inferences from observing the user's actions rather than relying on a pre-programmed knowledge representation. It interleaves the user's manual browsing with its own concurrent, autonomous exploration of links. Letizia employs a best-first search strategy augmented with heuristics to infer user interest. These heuristics are derived from observing user behavior, such as following links, saving references, and returning to documents. The agent models user interest as a preference ordering of links within a document, aiming to recommend a certain percentage of the most promising links. It also incorporates a "persistence of interest" model, recognizing that user interests often extend beyond a single search.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Inferring Interest from Behavior: The core idea is that a user's actions on the web (clicking links, saving pages) are strong indicators of their interests, even without explicit queries.
  • Proactive Assistance: Letizia doesn't wait for a user to ask for help; it actively explores in the background, making its suggestions available when the user might need them most (e.g., when unsure what to do next).
  • Interleaving Browsing and Search: The agent seamlessly integrates its search activities with the user's browsing, avoiding interruptions and leveraging the "idle time" when a user is reading a page.
  • "Persistence of Interest": User interests are not fleeting. Letizia remembers past expressed interests and continues to look for relevant information, saving users from repeatedly restating their preferences.
  • Focus on Link Preference: Instead of abstractly judging document interest, Letizia focuses on recommending the most relevant links within a given context, directly addressing the user's immediate decision-making process.

Key Results

The paper describes the conceptual framework and heuristics for Letizia's operation. While specific quantitative results are not presented, the qualitative findings highlight:

  • Letizia can infer user interest in topics like "Agents" by observing browsing patterns across various websites.
  • It can suggest relevant articles from a personal publication list based on previously inferred interests.
  • The agent's suggestions are dynamically recomputed based on changes in user behavior or at the user's request.
  • Letizia never takes control of the user interface, always providing suggestions that the user can choose to accept or ignore.

Limitations and Future Work

  • Limited Natural Language Understanding: Letizia currently relies on a keyword-based content model and lacks full natural language understanding capabilities. The authors suggest that partial natural language processing could significantly improve accuracy.
  • Heuristic Weakness: While individual heuristics are weak, their combination is intended to provide a useful approximation of user interests. The accuracy of these heuristics could be further refined.
  • Netscape Programmability: The implementation was limited by the extent to which Netscape could be programmed via AppleEvents.
  • Future Work: The authors mention the potential for integrating more sophisticated natural language processing. They also note that the extensible object-oriented architecture facilitates the incorporation of new heuristics for determining document interest. The paper also touches upon the need for "robot ethics" conventions for autonomous web agents.

Summary updated

BibTeX
@inproceedings{1995letizia,
  title = {{Letizia: An Agent That Assists Web Browsing}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1995},
}
Figure