Turkish Stopword Lists
Part of a Turkish text preprocessing toolkit built for CMPE 561 (Natural Language Processing), Application Project 1, Boğaziçi University, 2026. All code is our own; no external NLP library is used in the systems. Status: progress report, 20 October 2026.
General-domain Turkish stopword lists built three ways, as the assignment requires: a static list curated from an existing resource, a dynamic list extracted from corpus statistics, and their hybrid union. Companion repositories: normalization, sentence_splitter.
At run time a stopword is removed when its Turkish-lowercased form is in the chosen list. Hybrid = static ∪ dynamic (797 words).
Files
| File | Format | Contents |
|---|---|---|
lists/stopwords_static_expanded.txt |
one word per line | static list, 686 words: stopwords-iso repaired (37 mis-encoded entries), 31 entries removed (8 content words, 23 number words), + 223 inflected forms and the question-particle paradigm, + 23 closed-class additions |
lists/stopwords_static.txt |
one word per line | intermediate cleaned list, 440 words |
lists/stopwords_dynamic.txt |
one word per line | dynamic list, 200 words: Wikipedia + news, worst-rank combination, Poisson measure, k chosen on dev |
curation/stopwords_static_exclusions.tsv |
word ⇥ reason |
every removed entry with its reason |
curation/stopword_inflections.tsv |
lemma ⇥ forms |
inflected forms of 24 function words (hand-written from Turkish grammar) |
curation/stopwords_static_additions.tsv |
word ⇥ rule ⇥ note |
23 added closed-class words (rules D23/D24) |
guidelines/GUIDELINES.md |
Markdown | our stopword definition and annotation rules D2–D27 |
rankings/wikipedia/ranked_<measure>.tsv |
term ⇥ score |
full rankings of 581,412 word types from a 20,000-article Wikipedia sample by term frequency, document frequency, entropy and a Poisson score |
results/ |
JSON, Markdown | all dev/test scores, static-list ablation, per-genre scores, error analysis, Jaccard distances, corpus statistics |
figures/ |
PNG | test comparison, static list step by step, dynamic vs. static, threshold curves, Zipf plot, flow diagrams (titles in Turkish) |
Lines starting with # in the list files are provenance comments. The news rankings are not included because the
news corpus has no stated licence.
Results (test, provisional)
| System | Size | P | R | F1 | Acc |
|---|---|---|---|---|---|
| NLTK | 53 | 97.6 | 43.7 | 60.3 | 84.4 |
| Can et al. 2008 | 147 | 95.8 | 71.8 | 82.1 | 91.5 |
| Lucene | 209 | 91.8 | 74.6 | 82.3 | 91.3 |
| stopwords-iso (raw) | 504 | 88.8 | 81.3 | 84.9 | 92.2 |
| Static (ours) | 686 | 93.0 | 89.4 | 91.2 | 95.3 |
| Dynamic (ours) | 200 | 75.9 | 68.7 | 72.1 | 85.6 |
| Hybrid (ours) | 797 | 78.4 | 90.5 | 84.0 | 90.6 |
Evaluation is token-level on 100 + 100 sentences of UD Turkish-BOUN dev/test (20 per genre; 952 / 1,045 word tokens),
positive class = stopword. The reference labels were revised by an LLM (Claude, Anthropic) following
GUIDELINES.md and are awaiting human review; all scores are provisional, and the reference files will be added
after the review.
Sources
| Source | Licence | Use |
|---|---|---|
| stopwords-iso Turkish (stopwords-iso/stopwords-tr) | MIT | base of the static list |
| NLTK (nltk/nltk_data), Lucene (apache/lucene), spaCy (explosion/spaCy), trstop (uzay00/trstop), Can et al. 2008 (JASIST 59(3)) | Apache 2.0 / Apache 2.0 / MIT / MIT / article | comparison only, not included |
Turkish Wikipedia, dump 20231101 (wikimedia/wikipedia, 20231101.tr) |
CC BY-SA 3.0 / GFDL | dynamic extraction (20,000-article sample, seed 13) |
| TTC-4900 news (savasy/ttc4900); Yıldırım & Yıldız 2018 | not stated; not redistributed | second corpus for the cross-corpus filter |
| UD Turkish-BOUN (UniversalDependencies/UD_Turkish-BOUN) | CC BY-SA 4.0 | reference sentences |
Main literature
Nothman, Qin & Yurchak 2018 (NLP-OSS); Can et al. 2008 (JASIST); Saif et al. 2014 (LREC); Sarica & Luo 2021 (PLOS ONE); Schofield, Magnusson & Mimno 2017 (EACL); Fan, Arora & Treude 2023 (NLBSE).
- Downloads last month
- -