PRESY System: Contextual Query Reformulation

• PRESY System is a context-based query reformulation tool that integrates static user profiles and dynamic search histories to improve web search precision.
• It employs a sequential workflow—from user identification to query reformulation—using .Net implementation to enhance both relevance and diversity in search results.
• Experimental evaluations on Google, Yahoo, and Bing indicate improved top-result relevance and reduced redundancies, demonstrating its practical impact in information retrieval.

The PRESY System is a context-based query reformulation tool designed to enhance web information retrieval by leveraging both user profiles and evolving search contexts. Its architectural and methodological framework integrates static user data and dynamically-accumulated search history to modify user queries before submitting them to search engines, with the goal of increasing the precision and selectivity of returned results. PRESY was implemented in a .Net environment and evaluated with popular search engines using quantitative information retrieval metrics (Bouramoul et al., 2011).

1. Architectural Overview

PRESY operates as an external system that mediates between the end user and standard web search engines. Its design is centered on two principal forms of context:

• Static Context: Derived from user-identifying information, including attributes such as age, sex, native language, domain interests, and expertise.
• Dynamic Context: Constructed from accumulated historical data, specifically from previous search sessions and the analysis of documents (e.g. web page titles) returned by search engines.

The sequential workflow is:

1. User Identification: Extraction or construction of the user’s static profile. At first login, the user supplies personal attributes and interests.
2. Static Context Acquisition: Retrieval of stored static profile data for contextualization.
3. Dynamic Context Update: As search sessions proceed, the dynamic context is incrementally updated by extracting relevant terms from the result sets, subject to user approval.
4. Query Reformulation: The user’s query is programmatically expanded or modified via context-derived terms.
5. External Search: The reformulated query is dispatched to the selected search engine (Google, Bing, Yahoo).
6. Result Processing and Context Enrichment: Returned search results are parsed for additional candidate expansion terms; these are validated by the user for future use, further refining the dynamic context.

The system’s core cycle is depicted conceptually as:

User → (Identification & Static Context) → Query Input             ↓     Query Reformulation (Static + Dynamic Contexts)             ↓     Call to Search Engine             ↓     Result Display + Dynamic Context Enrichment

2. Construction of the Contextual Base

PRESY’s context base is the union of static and dynamic user-specific information, formally stored as attribute-value pairs:

• Static Context Construction: On first use, the system requests and stores immutable/tacit user information and declared interests, e.g., {"Language": "French", "Interest": "Archaeology", "Expertise Level": "University"}.
• Dynamic Context Construction: Post-query, the system parses the titles of retrieved web pages, applies stop-word (anti-dictionary) filtering, and then proposes the remainder to the user for inclusion in the dynamic context.

The operational sequence is:

Input Page Title → Tokenization → Stop-word Elimination → User Validation → Dynamic Context Update

Dynamic context thus attains increasing resolution with continued use, adapting to both the evolving search intent and the topical granularity emanating from search engine feedback.

3. Query Reformulation Process

Query reformulation is algorithmically the fusion of user input (initial query $Q_0$) and context ($C$). The refined query $Q_1$ is expressed as:

$Q_1 = Q_0 \cup f(C)$

Here, $f(C)$ denotes a function that selects and weights relevant context terms (from both static and dynamic sources) based on concordance with the query. For example, if $Q_0$ = "guelma" and $C$ contains "algeria", "university", "archaeology", the output might be "guelma algeria university archaeology".

The reformulated query is directly issued to the targeted search engine, and system logic may present multiple candidate reformulations, auto-selecting the optimal form.

The reformulation mechanism is both deterministic (in terms of attribute-value matching) and adaptive; repeated cycles reinforce the dynamic context, enhancing subsequent reformulations with greater topical specificity and query term selectivity.

4. Experimental Design and Results

Experimental validation of PRESY was performed using 15 queries, partitioned into 10 simple thematic cases (e.g., travel, news, culture) and 5 complex or specialist cases. Key parameters and results:

• Search Engines: Google, Yahoo, Bing
• Evaluation Criteria:
  • C1: Relevance of first three results.
  • C2: Relevance of last seven results out of ten.
  • C3: Redundancy (repetition of results from the same domain).

Each metric was scored on a normalized 10-point scale over all queries. Core improvements are summarized:

Search Engine	C1 (First 3 Results)	C2 (Last 7 Results)	C3 (Redundancy)
Google	6.62 → 7.69	5.60 → 6.77	(Improvement ~0.79)
Yahoo	5.78 → 6.11	4.92 → 4.18	(Improvement ~0.99)
Bing	3.38 → 4.23	3.94 → 4.87	(Improvement ~0.69)

The most significant gain ($\Delta$C1 = 1.07, $\Delta$C2 = 1.17) was observed in Google’s top results, directly evidencing the increase in content relevance and the decrease in redundant results. In some configurations (such as Yahoo with a high number of keywords), minimal regression in certain metrics was observed (C2 decreased by -0.24), suggesting sensitivity to engine-specific query parsing.

5. Comparative and Analytical Perspective

PRESY’s contextual reformulation was consistently superior to non-contextual (plain) queries for selectivity and precision, as measured per Information Retrieval convention:

• On the leading search engines, mean relevance increased for both the highest-rank (top-3) and subsequent (4–10) results.
• The reduction in redundant listings implies that personalized contextualization constrains the result set to more informationally diverse origins.
• Contextual dependency (especially static interest/expertise data) appears pivotal to improvements in themed or complex-query regimes.
• Engine-specific query handling may limit the effectiveness of context expansion, demonstrated by variable performance on Yahoo.

This suggests further adaptation to query syntactic constraints of heterogeneous search backends could optimize efficacy.

6. Future Directions and Theoretical Implications

Identified avenues for enhancement and potential research trajectories include:

• Automated Dynamic Context Acquisition: Minimizing or obviating user validation via unsupervised or weakly supervised NLP mechanisms.
• Advanced Profile Construction: Integrating longitudinal behavioral analytics or richer session histories to inform static and dynamic contexts with greater granularity.
• Broader Domain Application: Validation against domain-specific search engines, vertical markets, or other modalities (e.g., enterprise data lakes).
• Algorithmic Enhancement: Application of advanced NLP or ML models for automated term selection, weighting, and synonym/hypernym expansion within $f(C)$ to further optimize reformulation strategies.

The theoretical model underscores the impact of user-centric context on retrieval selectivity—a plausible implication is that such dual-base (static and dynamic) personalization frameworks could generalize to other information discovery platforms.

7. Significance and Applications

PRESY’s principal contribution is the explicit and modular exploitation of user-centric static and dynamic context for the purpose of automatic query transformation in web information retrieval. The demonstrated improvements over baseline querying, including enhanced relevance and diversity of returned content, point to direct applicability in search environments characterized by user inexperience, evolving topical focus, or information overload. Further, the methodology offers a framework for integrating emerging computational linguistics techniques and for adapting to multi-engine environments. The separation of static and dynamic context, operationalized through user validation and incremental expansion, provides a scalable template for subsequent research or deployment in both general-purpose and specialized search scenarios.