Utilizing a Semantic Sliding-Window Recovery algorithm to recover orphaned annotations in revised texts

Sammanfattning

As digital document workflows steadily rely on collaborative editing and heavy revisions, annotations frequently lose their anchor point. This ''broken anchor'' problem is exacerbated by structural and semantic rewrites, where traditional methods such as lexical matching heuristics fail to re-attach orphaned annotations. To address this, this thesis introduces a Semantic Sliding Window Recovery (SSWR) algorithm. By utilizing sentence-level transformer models, SSWR translates text into spatial embeddings, allowing the system to systematically scan revised documents and re-attach annotations based on semantic equivalence rather than character-level syntax. SSWR was evaluated against a Levenshtein-based lexical baseline across organic Wikipedia revisions, a live thesis workflow, and synthetic edge-cases. Results demonstrate that SSWR displays a higher recovery rate, proving robust against structural splits and text expansion. However, the algorithm's sentence-level chunking introduces a ceiling on boundary precision, highlighting a trade-off between semantic neighborhood recovery and exact token span accuracy.