Performance Analysis of Transformer-Enabled Semantic Crawlers for Scalable Text Retrieval

With the exponential growth of web-based content, efficient retrieval of contextually relevant textual information starting from seed URLs has become a critical challenge in web content mining and information retrieval. Traditional crawling and search methods—such as breadth-first search (BFS), depth-first search (DFS), best-first (focused crawling), topic-sensitive PageRank, and context-graph models—typically suffer from limitations such as parameter tuning overhead, lack of contextual understanding, requirement of large training datasets, high computational cost, and the need for specialised infrastructure. This research presents a comprehensive comparative study of multiple search and crawling models applied to textual retrieval from seed URLs, with a particular focus on their performance in diverse web‐structures (static vs dynamic) and content types. Employing a unified experimental framework implemented in Python with MySQL backend, we evaluate each algorithm using standard performance metrics (precision, recall, F1-score) alongside newer metrics such as coverage, relevance score, search time, memory usage, throughput and harvest rate. Machine-learning enabled variants (for example semantic-BFS and semantic-DFS using transformer-based embeddings) are also incorporated to assess their value over purely structural methods. Our results demonstrate that while semantic-enhanced BFS (Semantic-BFS) yields higher coverage, better relevance and faster response time in many scenarios, it shows limitations in classical metrics like precision/recall/F1 when ground-truth labels are inadequate for semantic relevance. The study provides insights into algorithmic trade-offs, suitability for different web architectures, and proposes hybrid strategies for next-generation crawlers and retrieval systems. The findings contribute toward the design of more adaptive, semantic-aware, and scalable web content mining frameworks.