Deprecated word Replacement in Technical Documents : A Comparitive Study

Sammanfattning

Background Industrial technical documents evolve: new components and naming rules appear while legacy wording remains. This creates inconsistent terminology and slows maintenance. This thesis studies how to detect deprecated terms (non-preferred under an internal standard) and suggest context-preserving sentence replacements in about 100 pages of anonymized English technical documents. Objectives (i) Build an editor-in-the-loop pipeline that flags candidate deprecated terms and proposes meaning-preserving replacements; (ii) compare a non-retrieval NLP baseline with a RAG approach for context-aware grouping; (iii) evaluate both using clustering and topic-coherence metrics plus expert judgment; and (iv) produce artefacts that can seed a company deprecated→preferred terminology standard. Methods Approach A (NLP baseline): LLM-assisted term extraction is validated with spaCy and regex patterns. A representative sentence is selected per term, and term-sentence pairs are embedded using Sentence-BERT (all-MiniLM-L6-v2). Clustering of K-Means with k selected using combined Silhouette and Calinski–Harabasz criteria produces grouping lists facing the editor. Approach B (RAG): Passages are organized and indexed in FAISS, then the top-k passages are retrieved for each term using a retrieval-oriented embedding model. An LLM uses this retrieved context to group terms and rewrite sentences, replacing deprecated terms with preferred ones while preserving meaning. Both approaches are evaluated with cosine-based intra/inter-cluster scores and a separation score, topic coherence (cv, UMass), and a domain-expert Likert questionnaire with sentence-level acceptability checks. Results RAG produced tighter clusters with stronger between-cluster separation, while inter-cluster similarity stayed comparable across methods. Topic coherence was mixed: cv was similar (with more stable medians for RAG), whereas UMass slightly favored the NLP baseline. Experts preferred RAG for contextual consistency and interpretability, and most suggested sentence replacements were accepted after review. Both approaches yielded large time savings versus manual work (around 90%). Conclusions Conditioning on retrieved evidence gives more interpretable and context-grounded terminology grouping and editing than a pure embedding-and-cluster baseline, at higher computational cost. A pragmatic strategy is to use RAG for high-precision grouping and edits, and the NLP baseline for faster triage. The resulting cluster lists and sentence diffs can be used to grow a living deprecated→preferred standard in editor-in-the-loop practice.