guidelines/GUIDELINES.md

# Stopword gold annotation guidelines

**Unit:** one surface word token (punctuation and digit tokens are excluded; a multiword token such as
*ağlayacaktı* is labelled once, as written). **Label** `1` = stopword, `0` = not.

**Definition (function, not frequency):** a stopword is a **function word**: it does grammatical work in the
sentence but does not tell what the text is about. Label each token **in context**: the same form can be 1 in
one sentence and 0 in another.

**Test question:** *If I delete this word, can I still tell what the sentence is about?* Yes → 1.

Rule numbers (D2–D27) are referenced in the `note` column and in `stopwords/docs/WALKTHROUGH.md` (Step 2).

## Rules

| # | Case | Example | Label |
|---|---|---|---|
| D2 | Numbers written as words, incl. ordinals and distributives | *iki kişi, bin, yüzde, onuncu, birer, binlerce* | 0 |
| D3 | *bir* as article "a"; determiners and quantifiers | *bir gün; her, bazı, tüm, bütün, birçok, birçokları, hiçbir, diğer, öbür, kimi, her türlü* | 1 |
| D4 | *bir* meaning "one" (a counted quantity, incl. *bir kez*, *bir defada*) | *bir elma iki armut; üç oda bir salon* | 0 |
| D5 | *bir* in time phrases: if translatable as "a/an" → 1 | *bir saat sonra, bir süre* | 1 |
| D6 | Degree and quantity words | *çok, daha, en, az, pek, fazla, gittikçe* | 1 |
| D7 | Focus and emphasis words | *bile, sadece, yalnızca, hatta, tam (= just)* | 1 |
| D8 | Separate *de/da* ("too", "also", concessive "-se de") | *Ali **de** geldi; yapsam **da*** | 1 |
| D9 | Suffix *-de/-da* inside a word | *okul**da*** | not judged |
| D10 | Question particle and its forms | *mi, mu, misin, mıdır, musunuz* | 1 |
| D11 | Pronouns, **all inflected forms**, incl. indefinite and possessive | *o, onu, onun, bunlardan, kendisine, kimse, kimsenin, benimkisi, bizimkiler* | 1 |
| D12 | Question words | *ne, neden (= why), niçin, nasıl, hangi, kim* | 1 |
| D13 | Deictic adverbs | *şimdi, burada, böyle, şöyle, öyle, öylece* | 1 |
| D14 | Existence and negation predicates | *var, vardır, varsa, yok, değil* | 1 |
| D15 | *etmek/eylemek* and similar **light verbs** (meaning is in the preceding noun) | *kabul **etti**, ikram **edildi**, söz **edebilir**, neden **olan**, söz konusu **olduğunda*** | 1 |
| D16 | *olmak* meaning "become" | *doktor **oldu**, ne **olmak** istiyorsunuz, âdet **olmuştur*** | 0 |
| D17 | *yapmak* (always, see D25) | *ev **yaptı**, ziyaret **yapıyor*** | 0 |
| D18 | Postpositions and connective adverbs | *için, ile, gibi, kadar, göre, beri, sonra, önce (incl. "first"), rağmen, karşı, ilgili, ilişkin, dair* | 1 |
| D19 | *sonra/önce* as a noun | *ders **sonrası*** | 0 |
| D20 | Content (manner/time) adverbs | *hızla, düzenli, yıllarca, durmadan, geç, yeni (= newly), şahsen, bizzat* | 0 |
| D21 | Interjections | *ah, valla, lütfen* | 0 |
| D22 | Still unsure after the rules (also semantically light nouns such as *şey*, *kez*; *aynı*, *ilk*) | | 0 + note |
| **D23** | **Relational nouns used as postpositions** (they translate as English prepositions) | ***içinde, arasında, arasındaki, üzerine, üstünde, altında, başına (= per), ardından, yoluyla, aracılığıyla, halinde, yanı sıra, yanında (= besides), başka (= other than)*** | 1 |
| | …but **not** when the noun keeps spatial/content meaning | *sokağın **ortasında**, denizin **kenarında**, kelimelerin **arkasına**, yemek **sonunda**, temaslar **çerçevesinde**, **yanında** pilav* | 0 |
| **D24** | **Discourse and stance adverbs / connectives** | ***elbette, kesinlikle, mutlaka, asla, hiç (= never/at all), artık, yine, işte, yoksa, sizce, falan, bu yüzden, ilk olarak*** | 1 |
| **D25** | ***bulunmak* as an existential or light verb**; *yapmak* is never treated as light (D17) | *fiyat **bulunurken**, uyarıda **bulundu**, bir yerde **bulunuyor*** | 1 |
| **D26** | ***olmak* as copula, existential or auxiliary** (vs "become" D16 and "happen" → 0) | *mutlu **oluyorum**, haberim **olduğunu**, sunacak **olan**, kim **olduğunu**, önlem **olarak**; but ne **oluyor**, ne **olacağı**, **oldu** da → 0* | 1 |
| **D27** | **Idioms:** parts of fixed expressions keep 0, including *bir* in *bir araya* | *bir araya gelmek, ortaya çıkmak, yerine getirmek, karşı karşıya kalmak, haline gelmek, yer almak, farkına varmak* | 0 |

D23–D27 were added during the first full labelling pass (19 Oct 2026) for cases D2–D22 did not cover.

## Known treebank errors
UD Turkish-BOUN occasionally mis-tags tokens (e.g. *Benzin* and *Tasviri* as PRON, *Adjani* as CCONJ, *neden*
"why" as NOUN). Label by the rules, not by the UPOS column. Such rows have `note = LLM: UD-error`.

## Procedure
1. Files are pre-labelled from UD UPOS tags (`stopwords/scripts/prepare_stopword_gold.py`); `check=1` marks uncertain classes.
2. **19 Oct 2026:** a first full pass was done by an LLM (Claude), applying these rules; every changed row has
   `note = LLM: <rule>`. The first line of each file records this (`LLM-LABELLED … HUMAN REVIEW PENDING`).
3. **Human review (required):** the annotator reviews **every** row (Annotate tab of the GUI), corrects `label`,
   and then marks the file as reviewed; the first line becomes `# status: REVIEWED by <name>, <date>`.
4. **Agreement:** a second annotator labels `boun_dev_annotator2.tsv` independently (no pre-labels);
   Cohen's kappa via `stopwords/scripts/stopword_agreement.py`. Disagreements may lead to new rules here.