Turkish stylometry walkthrough¶

A complete runnable example: attribute a small Turkish short-story corpus using MFW + Ateşman readability + Burrows Delta. The stories are public-domain Ömer Seyfettin texts from Turkish Wikisource.

Setup¶

uv pip install 'bitig[turkish]'
python -c "import stanza; stanza.download('tr')"
bitig init seyfettin --language tr
cd seyfettin

This scaffolds a project directory with study.yaml pre-configured for Turkish. Confirm with:

bitig info

The language row shows tr.

Corpus¶

Place 3-5 Turkish short stories in corpus/ as UTF-8 .txt files. A good public-domain source is Ömer Seyfettin on Turkish Wikisource — dozens of early-20th- century short stories are already transcribed there.

Add a corpus/metadata.tsv:

filename    author  year
bomba.txt   Omer_Seyfettin  1910
kesik_biyik.txt Omer_Seyfettin  1911
forsa.txt   Omer_Seyfettin  1913
pembe_incili_kaftan.txt Omer_Seyfettin  1917

A real study would include several authors. For a single-author demo, pair Seyfettin with a few stories by Refik Halit Karay (also public domain) to give Delta something to discriminate.

Running the study¶

bitig ingest corpus/ --language tr --metadata corpus/metadata.tsv
bitig run study.yaml --name first-run

bitig ingest runs Stanza through spacy-stanza. The first run parses every document and caches the DocBins; subsequent runs hit the cache and finish in seconds.

What you get¶

A default Turkish study computes:

MFW (top 1000 tokens, z-scored relative frequencies)
Turkish function words — loaded from resources/languages/tr/function_words.txt, derived from UD Turkish BOUN closed-class tokens
Ateşman and Bezirci-Yılmaz readability indices
Burrows Delta + PCA/MDS reduction plots

The output folder results/first-run/ contains:

result.json with Delta scores and provenance
table_*.parquet feature matrices
PNG / PDF figures (distance heatmap, PCA scatter)
a provenance.json that records the corpus hash, seed, and full resolved config

Worked example: 28 short stories by Ömer Seyfettin¶

The examples/turkish_seyfettin/ directory in the repository ships a complete, end-to-end run against 28 short stories by Ömer Seyfettin (1884-1920; in the Turkish public domain since 1991), scraped from tr.wikisource.org by fetch_corpus.py and committed alongside the study so the numbers below are reproducible to the exact random draw.

python examples/turkish_seyfettin/fetch_corpus.py --n 30   # ~30s, polite to Wikisource
python -m bitig run examples/turkish_seyfettin/study.yaml --name seyfettin

Corpus. 28 stories survive the >200-token cutoff; story lengths range from 326 to 4 455 tokens (median ≈ 1 700). Wikisource transcriptions are CC BY-SA 4.0; attribution and source URLs live in examples/turkish_seyfettin/manifest.json.

Study. Five hundred most frequent words (z-scored, min_df = 2) + Turkish function-word frequencies → Burrows Delta self-attribution + PCA + Ward hierarchical clustering. Single-author setups don't admit between-author verification (Imposters / classify), so this is exploratory within-author stylometry, not authorship attribution.

Burrows Delta self-attribution¶

Across the 28-story leave-one-out grid, every story's nearest training neighbour is itself (accuracy = 1.0). For a single-author corpus that's the trivial result — every document is closer to its own MFW profile than to any other — but it also confirms that no story has been mistakenly tagged with a different author's metadata. The interesting signal lives one rank deeper, in the rank-2 nearest neighbour, which surfaces stylistically adjacent stories. The pairwise distance matrix used for ranking is preserved in results/seyfettin/burrows/result.json.

PCA on the MFW-500 lexical space¶

PCA scatter

PC1 explains 7.7 % of the variance, PC2 explains 7.2 %. That no single component dominates is itself diagnostic: lexical variance inside one author is spread across many small axes rather than concentrated in one or two. Compare this to a between-author PCA (e.g. the Federalist tutorial) where PC1 alone routinely captures 30 %+.

Top loadings — PC1: baktı, değildi, açtı, durdu, hafif, iki, şeyler, gelince. PC1 separates stories that lean on simple-past 3rd-person narration verbs (baktı, açtı, durdu) from those without.

Top loadings — PC2: idi, ediyordu, onu, o, olduğu, nihayet, etti, durdu. PC2 picks up the past-imperfective auxiliary (idi, ediyordu) and the 3rd-person pronoun cluster (onu, o, olduğu) — i.e., the choice between extended-state description and event-driven narration.

The biplot overlays the top-12 loading vectors onto the same 2-D projection:

PCA biplot

Open the interactive plotly biplot ↗ — hover to read story IDs and arrowhead labels.

Ward hierarchical clustering (k = 4)¶

Ward dendrogram

Cutting at four flat clusters yields:

Cluster	n	Member story IDs
0	23	the bulk of the corpus (`aleko`, `bomba`, `kasag`, `forsa`, …)
1	3	`bir_refikin_defter_i_ihtisasat_ndan`, `elma`, `hediye`
2	1	`bir_kay_s_n_tesiri`
3	1	`keramet`

The three-story cluster turns out to be the three shortest pieces in the corpus (329 / 517 / 457 tokens). The two singletons are also short (554 and 511 tokens). What the dendrogram is mostly surfacing, in other words, is a length-driven signal: z-scored MFW counts get noisy below ~600 tokens, so short stories drift away from the main cloud regardless of topic. This is not a bug in the method — it is the natural consequence of MFW estimation variance at small N, and it is the single most useful thing this analysis tells you:

Open the interactive plotly dendrogram ↗

If you are running stylometry on Turkish short prose, raise the per-document token floor to at least 1 000 — or move to character n-grams, which are far more length-tolerant — before drawing topic / period conclusions from the sub-cluster structure.

Caveats and what is not shown here¶

A genuine authorship-attribution demonstration needs at least one additional PD-old Turkish prose author with comparable Wikisource coverage, which Wikisource:tr does not currently have for the early-republican period (Refik Halit Karay's transcriptions are present but the underlying texts will not enter the Turkish public domain until 2036). For attribution work we recommend pairing Seyfettin with a different-genre control corpus (e.g. parliamentary speeches, Türkçe Wikipedia featured articles by topic, or your own institutional corpus) rather than a contemporaneous literary author.

Customising¶

Edit study.yaml to swap features or methods. For example, to use contextual embeddings instead of MFW:

features:
  - id: bert_tr
    type: contextual_embedding
    # model auto-resolves to `dbmdz/bert-base-turkish-cased` via the language registry
    pool: mean

For a heavier Turkish encoder, point model: at any HuggingFace checkpoint:

features:
  - id: bert5urk
    type: contextual_embedding
    model: stefan-it/bert5urk
    pool: mean

Notes on Turkish specifics¶

Morphology. Turkish is agglutinative; a token like evlerinizden packs ev+ler+iniz+den into one form. Stanza's Turkish BOUN model lemmatises and tags these correctly, which matters for POS-ngram and dependency-based features.
Syllable counting. Both Ateşman and Bezirci-Yılmaz count syllables using a vowel-counter specialised for Turkish orthography (including ı, ğ, ş, ç, ü, ö).
Function words. The bundled list leans on Turkish's closed-class postpositions, conjunctions, and discourse particles (e.g. ile, ancak, fakat, çünkü, ki, ise).

Troubleshooting¶

ModuleNotFoundError: No module named 'spacy_stanza' — run uv pip install 'bitig[turkish]'.
FileNotFoundError: ... stanza_resources/tr/default.zip — run python -c "import stanza; stanza.download('tr')". The model is about 600 MB.
Very slow first ingest on MPS. Stanza's Turkish model does not yet support Apple Silicon MPS. Expect CPU parse rates on first run; subsequent runs are cache hits.