eminesener/sentence_splitter

Turkish Sentence Boundary Detection Data

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.

Data for our two Turkish sentence splitters: a rule-based cascade and our own Naive Bayes classifier. Both treat sentence splitting as a binary decision over end-of-sentence candidates (runs of . ? ! … with closing quotes/brackets, and line ends). Companion repositories: normalization, stopwords.

Files

File	Format	Contents
abbreviations/abbreviations.tsv	form ⇥ class ⇥ source ⇥ wiki_dot ⇥ wiki_nodot	abbreviation lexicon, 394 dotted forms from the TDK Kısaltmalar Dizini (2019) and the SplitExp trie, filtered by Turkish Wikipedia counts (written with vs. without a period); classes title / address / general / other
abbreviations/abbreviations.txt	one form per line	the same forms
abbreviations/abbreviations_additions.tsv	form ⇥ class ⇥ reason	hand-written additions with reasons
boun/boun_<split>.docs.jsonl	one JSON object per split: id, text, sentences (gold [start, end] character spans), sent_ids	UD Turkish-BOUN sentences joined into one running text (space after terminal punctuation, newline otherwise)
boun/boun_<split>.candidates.tsv	offset ⇥ marker ⇥ label ⇥ prev_word ⇥ next_word ⇥ left_context ⇥ right_context ⇥ sent_id	every boundary candidate with its gold label (1 = sentence end)
boun/boun_stats.json	JSON	candidate counts by marker type; every gold sentence end is a candidate
results/	JSON, Markdown	dev scores of all systems, per-genre and per-marker breakdowns, 5-fold CV of the Naive Bayes settings
figures/	PNG	flow diagrams (labels in Turkish)

Candidate counts: train 7,803 sentences / 8,415 candidates; dev 979 / 1,049; test 979 / 1,047.

Results (dev; test splits will be scored once for the final report)

System	BOUN F1	SplitExp F1
always split (baseline)	0.965	0.801
rule-based cascade	0.979	0.974
Naive Bayes, lexical features (B)	0.981	0.968

Naive Bayes is trained on BOUN train only; its settings were chosen by 5-fold cross-validation on BOUN train.

Sources

Source	Licence	Use
UD Turkish-BOUN (UniversalDependencies/UD_Turkish-BOUN); Türk et al. 2022, LRE	CC BY-SA 4.0	train / dev / test documents and labels in boun/
SplitExp (ideateknoloji/SplitExp, Dataset/sbd.json, commit cd6e457); Özbey & Dinçsoy 2019, SIU	MIT	second test set (685 news documents; dev/test split by alternating sorted ids); abbreviation trie. Raw file not included, download from the repository
TDK Kısaltmalar Dizini (Türk Dil Kurumu, 2019 PDF, via the Internet Archive)	not stated	dotted abbreviations in the lexicon, with attribution
Turkish Wikipedia, dump 20231101	CC BY-SA 3.0 / GFDL	ambiguity filter counts for the abbreviation lexicon