Introduction
Standard search engines are optimized for "lookup" tasks where users have a clear, well-defined information need. However, they often fail to support exploratory search—scenarios where users are unfamiliar with a domain or have "vague" requirements. This paper presents a novel search interface designed to bridge this gap by introducing a "vague querying" mechanism and a compact, visual representation of search results that provides a "visual fingerprint" of the target pages.
System Architecture
The prototype system is built on a three-tier architecture: a custom crawler for data acquisition, an Apache Solr-based backend for indexing and retrieval, and a highly interactive SVG/JavaScript frontend. This structure allows for real-time adjustments to query weights and dynamic rendering of visual result components based on user interaction.
Interaction Design
The interface departs from traditional search paradigms in two significant ways: the input of query terms and the visual representation of results.
The Confidence Slider
Instead of a static text box, the UI employs a horizontal "confidence" slider. As users enter terms, they appear as interactive tags. These tags can be moved horizontally across a central axis to indicate preference:
- Positive Weighting (Left): Moving a tag to the left indicates higher importance or "confidence" in the term.
- Negative Weighting (Right): Moving a tag to the right indicates a desire to exclude the term (effectively a visual NOT operator).
Mathematically, a term's position $x$ is mapped to the interval $[-1.0, 1.0]$. The sign determines the inclusion/exclusion status, while the absolute value $|x|$ serves as a boost factor within the search index.
Result Representation: The Visual Fingerprint
To facilitate rapid scanning and "visual scent," each search result is condensed into a horizontal "result bar" rather than a standard text snippet. This allows for a high density of information in a compact space.
Each result bar consists of three primary components designed to aid pre-attentive processing:
- Color Bar: A histogram representing the dominant colors of the page, generated by clustering the top 600 pixels using the KMeans algorithm. This helps users recognize the "vibe" or brand of a site.
- Salient Extract: A thumbnail of the most visually "interesting" region of the page, identified via the Achanta saliency algorithm which targets high-contrast regions.
- Word Cloud: A summary of unique content. Terms are scored using a TF-IDF-like formula against the Brown Corpus to highlight page-specific keywords rather than common language.
Evaluation and Results
The system was evaluated through a two-step usability study with 17 participants. In a formative task involving finding hiking destinations while avoiding specific terrains (e.g., the Alps), users intuitively utilized the slider to weight terms. Qualitative feedback was overwhelmingly positive; on a 5-point scale (where 1 is best), the "representation of results" and "weighting ability" both scored highly (mostly 1s and 2s).
Furthermore, a "recognition" test demonstrated that participants could successfully use visual cues (colors and thumbnails) to identify and deprioritize irrelevant results even when textual descriptions were ambiguous.
Conclusion and Future Work
The proposed interface successfully treats user uncertainty as a functional parameter rather than a query failure. By providing visual fingerprints, the system supports faster relevance judging during exploratory tasks. While the prototype demonstrated effectiveness on small-scale indices, future work will focus on scalability for larger datasets and optimization for mobile devices where horizontal space is at a premium.
Summary updated