Müşteri şikayetleri üzerinden doğal dil işleme teknikleri ile beyaz eşyalarda arıza teşhisi

The thesis investigates the application of Natural Language Processing (NLP) for fault classification in the white goods sector, using real-world service data. two-stage hybrid pipeline is developed: first, unstructured customer complaint texts are transformed into structured representations through both rule-based preprocessing and Large Language Model (LLM)-based enrichment. Then, multiple machine learning classifiers are trained to predict repair codes, which include the Random Forest, Naive Bayes, LightGBM, BiLSTM CNN-BiLSTM and DistilBERT. The system is evaluated on a large-scale dataset of 103,000 technical service records. Experimental results reveal that semantically enriched representations generated by GPT-4o-mini consistently outperform both unprocessed and rule-based inputs. Across all classifiers, the LLM-Rich preprocessing strategy yields the highest performance, achieving complexity-weighted average F1 scores of up to 63,4\%, compared to 54,1\% for the best unprocessed input and 52,2\% for the best rule-based alternative. Furthermore, the study introduces a complexity score for each repair code, allowing analysis of model performance relative to task difficulty. Results indicate that LLM-enhanced inputs not only improve accuracy but also increase robustness across high-complexity classes. Overall, the results indicate that semantically enriched NLP pipelines offer a scalable and powerful basis for semi-automated fault diagnosis, particularly in low-resource languages like Turkish, and greatly strengthen the robustness of the corresponding classification models.