KazakhStopEval: Stop-Word Inventories, Validation Benchmarks, and NLP Evaluation Artifacts for Kazakh Library OCR Collections

### Dataset Description This dataset provides the stop-word inventories, manual validation sets, baseline stemmer, and downstream evaluation benchmarks described in the paper *"Building an NLP Pipeline for Kazakh-Language Library Collections: Stop-Word Extraction and Genre Classification"* (AIST 2026 / CCIS). The resources are derived from a 122.56-million-token Kazakh-centred corpus combining 2,028 OCR-scanned books from the National Library of the Republic of Kazakhstan and 231,729 Kazakh Wikipedia articles. ### Contents of the Deposit: 1. **Stop-Word Inventories (`stopwords/`):** - `grammar_dict.json` (335 normalised grammar-dictionary terms) - `stopwords_single_tfidf.txt` (2,991 statistical single-token candidates at 5% threshold) - `stopwords_collocation_phrases.txt` (3,462 collocation pairs) - `stopwords_combined.txt` (3,414 combined surface forms) - `stopwords_multilingual.txt` (3,930 multilingual forms including Kazakh, Russian, English) 2. **Validation Data (`validation/`):** - `manual_stopword_review_sample_200_labeled.csv` (stratified 200-word sample with manual expert annotations and Wilson confidence intervals) - `manual_stopword_review_template.csv` (unlabeled annotation template for replication) 3. **Evaluation Sets (`proxy_evaluation/`):** - `queries.csv` & `qrels.csv` (50 queries: 39 topical + 11 known-item, with associated relevance judgments) - `classification_metrics.csv` & `topic_metrics.csv` (ablation results across configurations A–E and stemmed variant F) 4. **Tools & Baselines (`tools/`):** - `kazakh_stemmer.py` (lightweight rule-based suffix-stripping stemmer for Kazakh) ### Metadata for Mendeley: - **Category:** Computer Science > Artificial Intelligence / Computational Linguistics / Information Retrieval - **Keywords:** `Kazakh NLP`, `Stop Words`, `Digital Libraries`, `Low-Resource Languages`, `Information Retrieval`, `Agglutinative Morphology`, `LibRAG` - **License:** CC BY 4.0