Abstract
Professional searchers, including legal and scientific experts, require high-recall results and transparent retrieval mechanisms. Traditional personalization methods, which often rely on black-box popularity metrics, are ill-suited for the complex, exploratory nature of professional workflows. This paper advocates for the use of Professional Knowledge Graphs (PKGs) to bridge the gap between user expertise and search relevance. By structuring user knowledge as an interpretable graph, we enable a two-stage retrieval process that improves ranking accuracy while providing users with the transparency necessary to trust and control their search experience.
Introduction
Standard search engines are optimized for general web queries where popularity—measured by aggregate click data—serves as a reliable proxy for relevance. However, professional search environments present unique challenges that render these models ineffective. Professionals often require high recall, meaning they must find all relevant documents rather than just a single "best" answer. Furthermore, their search process is inherently exploratory and longitudinal. Finally, professional needs are highly specific; the "sparse data" problem means there is rarely enough aggregate click data from other users to inform a personalized ranking for a niche legal case or scientific inquiry.
The Professional Knowledge Graph (PKG) Approach
To address these challenges, we propose shifting from query-centric retrieval to a model based on the Professional Knowledge Graph (PKG). A PKG is an automatically generated structure where nodes represent terms and documents relevant to a specific user's domain of expertise.
In this model, edges represent relationships such as document similarity, term co-occurrence, or the "representativeness" of a term for a document (calculated using weights like tf-idf). Retrieval is executed in two stages:
- Initial Retrieval: A standard algorithm identifies the top 1,000 candidate documents.
- Re-ranking: These candidates are re-ranked based on their "centrality" within the user’s PKG, ensuring that results align with the user's established knowledge base.
Transparency and Explainability
For professionals, explainability is a functional requirement. If a ranking changes without a clear reason, the user loses confidence in their workflow. Because the PKG is a visual, human-understandable structure, it serves as the foundation for transparent interfaces.
By exposing the graph to the user, the system can provide explicit justifications for its recommendations—for instance, highlighting that a document is suggested because it connects to specific topics the user has previously researched. This "transparency by design" allows users to explore and even manually adjust the influences on their search results.
Evaluation and Insights
Pilot experiments using the two-stage PKG re-ranking method demonstrated a small but statistically significant improvement in ranking accuracy over non-personalized baselines. This suggests that even sparse user data, when structured as a graph, can enhance relevance.
The research also draws on Dervin’s "Knowledge Gap" model, viewing search as a process of identifying where new information should "attach" to a user's existing knowledge map. However, this approach introduces the risk of "filter bubbles." A critical challenge remains in balancing the exploitation of a known user profile with the exploration of novel, potentially relevant directions.
Future Directions
The transition to explainable, personalized professional search requires addressing several key hurdles:
- Data Scarcity: The community needs richer datasets, like the iSearch physics dataset, but supplemented with longitudinal search logs and interaction data.
- Dynamic Modeling: PKGs must evolve as users gain expertise or shift project focuses.
- User-Centric Metrics: Evaluation must move beyond simple click-through rates to measure transparency, user trust, and task-completion efficiency through qualitative studies.
By prioritizing transparency and user control, we can move toward search engines that act as true partners in professional discovery.
Summary updated