Back to search

International Hellenic University

Machine learning and neural methods for federated patent search

Abstract

dc:description

This Ph.D. dissertation investigates the specialized and highly demanding domain of patent information retrieval, aiming to bridge the gap between traditional search methodologies and modern Artificial Intelligence (AI) capabilities. Professional patent search is characterized as a mission-critical, recall-oriented task where the failure to retrieve a single relevant document can result in severe consequences, including patent invalidation, infringement litigation, and substantial economic loss. Despite advancements in Natural Language Processing, standard neural models often falter in this domain due to the unique characteristics of patent documents, which employ highly technical and often intentionally vague legal terminology to define the scope of protection. To address these inherent complexities—specifically extreme document length, intricate structure, and the vocabulary mismatch problem—this research develops and validates novel frameworks for two critical stages of the Federated Patent Search pipeline: results merging and document re-ranking. A central challenge addressed in this work is the effective retrieval of long, structured documents in the Transformer era. Standard neural architectures struggle with patent length due to quadratic computational complexity. Common solutions to this length limitation include summarization, document segmentation, or specialized long-context models. This thesis adopts a segmentation-based approach, motivated and theoretically grounded in the "scope hypothesis". This hypothesis posits that patent relevance is often localized within specific structural sections—such as the Abstract, Description, or Claims—rather than being uniformly distributed. By leveraging these distinct components, the proposed methodology preserves fine-grained technical details essential for determining novelty, which are typically lost in holistic document processing. In the first part of the thesis, we address the challenge of merging results from distributed, heterogeneous resources where relevance scores are often incomparable or withheld. We introduce the Machine Learning Models for Results Merging (MLRM) framework, which repurposes a Centralized Sample Index (CSI) as a dynamic training ground to normalize local scores into a global relevance space. Through extensive experimentation, we demonstrate that ensemble learning methods, specifically Random Forests, significantly outperform established heuristic baselines such as CORI and SSL. This establishes a new benchmark for results merging, proving particularly robust in uncooperative environments where search engines operate as "black boxes". The second major contribution is the Query-Aware Patent Re-ranking (QAPR) framework, designed to optimize the final ranking by treating patents as sequences of semantically independent structural components. QAPR utilizes a hybrid neural architecture to fuse three distinct categories of relevance signals: lexical signals capturing exact term matches via BM25, semantic signals utilizing Sentence-BERT (SBERT) embeddings to capture deep conceptual context, and structural signals modeling the explicit relationships and importance weights between different document sections. A key innovation is the introduction of a query-adaptive weighting mechanism that dynamically adjusts the influence of lexical versus semantic signals based on the discriminative power (IDF) of query terms. Experimental evaluation on the CLEF-IP and WPI-PR datasets confirms that QAPR yields statistically significant improvements, establishing a new standard for structure-aware, long-document retrieval in the patent domain.

Degree

thesis:*
Grantor dc:publisher
International Hellenic University
Year dc:date
2026

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Stamatis, Vasileios
  • Σταμάτης, Βασίλειος

Subjects

dc:subject × 12

Rights

Language dc:language
eng

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/61634

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Stamatis, Vasileios; Σταμάτης, Βασίλειος. Machine learning and neural methods for federated patent search. International Hellenic University, 2026. http://hdl.handle.net/10442/hedi/61634