Banano kirtis v2 🍌 — Lithuanian stress annotator (kirčiuoklis)

Predicts, for every word in a sentence, which syllable carries the stress and which accent type it has (trumpinė/grave, tvirtapradė/acute, tvirtagalė/tilde) — from sentence context, so homographs come out right: kasojè … šuõ kãsa duõbę, màno brólis mãno, dù sū́nus … vyrèsnis sūnùs, šiõs kalbõs … vìsos kal̃bos.

v2 replaces kristijonas/banano-kirtis (v1). Same label scheme, different base (large instead of base) and — the part that matters — different training data: written text stressed through a morphology-aware pipeline instead of stress recovered from speech transcriptions.

Vegetable-garden lineage of open Lithuanian speech models: svogūnas 🧅 → paprika 🌶️ → bananas 🍌.

What it is

XLMRobertaForTokenClassification, 37 classes, one prediction per word (first sub-token):

0                      no mark on this word
1 + syllable*3 + acc   acc: 0 = short/grave, 1 = acute, 2 = tilde
                       syllable = 0-based index of the stressed syllable nucleus

Fine-tuned from FacebookAI/xlm-roberta-large on meldynamics/lt-stressed-corpus: 108,324 sentences / 1.14 M labelled words from the MATAS and ALKSNIS corpora. Those corpora carry morphology, not stress: the stress was added by the dataset's author, who queried the VDU Kirtis service for candidate stressed forms and matched their morphological labels against the corpus's own (manually checked) annotation. So the marks are automatic but morphology-aware, not hand-placed. Tokens whose stress came from either fallback path (phonology_engine, lki.words, together ~0.4% of word tokens) were excluded from the loss, as were digits and multiple-best ties. 6 epochs, lr 2e-5, batch 32, max 256 tokens, bf16, one A100, ~30 min.

Evaluation — what the numbers mean

3,000 held-out corpus sentences (28,883 words), never seen in training:

Remaining errors in the recommended setup: 0.41 % wrong syllable, 0.09 % right syllable / wrong accent type. Most of the model-alone gap (1.6 %) is single letters and abbreviations that the corpus leaves bare — the lexicon layer handles them.

These are agreement with the corpus's stressing, not verified accuracy. There is still no expert-adjudicated open benchmark for Lithuanian stress. The corpus is large and consistent, which is why the number is useful; it is not human-verified truth. The v1 gap is partly convention: v1 learnt from speech, where clitics are unstressed; v2 marks function words the normative way (ir̃, sù, priẽ, õ). If you have, or would help build, an expert-annotated test set, please get in touch.

Known limitations

• Foreign names (Rossi, Wildberries): the model tends to leave them unmarked, which is the safe behaviour; check confidence.
• Rare homograph readings lose to the frequent one when context is thin: the verb gė́lė ("it ached") comes out as the noun gėlė̃ (10:1 in the corpus).
• Acronyms, digits, single letters should be passed through untouched before calling the model (see below).
• About 3× slower than v1 (560 M parameters): ~55 ms per sentence on a laptop CPU, ~0.5 ms batched on an A100.

Usage

import torch
from transformers import AutoModelForTokenClassification, AutoTokenizer

MD = "kristijonas/banano-kirtis-v2"
tok = AutoTokenizer.from_pretrained(MD)
model = AutoModelForTokenClassification.from_pretrained(MD).eval()

words = "Mano brolis mano , kad rytoj lis .".split()
enc = tok([words], is_split_into_words=True, return_tensors="pt")
with torch.no_grad():
    pred = model(**enc).logits.argmax(-1)[0]

first = {}
for pos, wid in enumerate(enc.word_ids(0)):
    if wid is not None and wid not in first:
        first[wid] = pos
for i, w in enumerate(words):
    if not w[0].isalpha():
        continue                          # punctuation is context only; its label is noise
    p = int(pred[first[i]])
    if p == 0:
        print(w, "→ no mark")
    else:
        syl, acc = (p - 1) // 3, ["short", "acute", "tilde"][(p - 1) % 3]
        print(f"{w} → syllable {syl + 1}, {acc}")

Feed words cased, punctuation as its own tokens (it is context). Rendering the combining mark onto the right letter is a separate step — see below; it is where most of the difficulty is.

Recommended: a lexicon in front

The model is best used behind a dictionary built from the same corpus: forms with a single attested stress (130 k) come from the lexicon, forms with several attested stresses (1.8 k homographs) let the model choose among the attested variants, everything else (out-of-vocabulary words, names) goes to the model. That is the 99.4 % row above. The reference implementation (kirtis_v2.py, rules.py, build_lexicon.py) is in the training repo and is short; the lexicon is a plain JSON form → (syllable, accent).

Placing the mark

syllable indexes the word's syllable nuclei, so you must split the word the same way the labels were made. The rules (each learnt by shipping the wrong output first):

1. Only true diphthongs merge into one nucleus: ai au ei ui ie uo (plus oi ou eu in loans). Every other vowel pair is hiatus — two syllables (te-a-tras).
2. A sonorant (l m n r) joins the nucleus only after a SHORT vowel (a e i u) and only when no vowel follows: kal̃nas, ir̃; after a long vowel it is a plain coda: dė̃l.
3. i after a consonant and before a ą o u ų ū is a palatalization glide, not a nucleus member: dirbtiniù, vėliaũ, siū́lo. Before i e ė it is a real syllable (pri-i-ma, pasi-ė-mė); ie is a real diphthong (priẽš, píeva).
4. Within a two-element nucleus: acute on the FIRST element (káulas, píeva), grave on the SECOND (Ukraìna), tilde on the SECOND vowel (kaĩp, priẽš) or on the sonorant of a mixed diphthong (ir̃, kar̃tu, klieñtas) — except native ie/uo + sonorant, where it stays on the vowel (kasdiẽn, dviẽm).

This implementation round-trips 99.95 % of the 1.16 M corpus tokens (label → mark → label). Copy freely.

VOW = set("aeiouyąęėįųū"); SON = set("jlmnr"); SHORT_VOW = set("aeiu")
DIPH = {"ai", "au", "ei", "ui", "ie", "uo", "oi", "ou", "eu"}
GLIDE_VOW = set("aąoųuū")
COMB = {"short": "̀", "acute": "́", "tilde": "̃"}

def nucleus_groups(word):
    w, groups, cur = word.lower(), [], []
    for i, c in enumerate(w):
        nxt = w[i + 1] if i + 1 < len(w) else ""
        if c in VOW:
            if not cur:
                cur = [i]; continue
            prev = w[cur[-1]]
            glide = (len(cur) == 1 and prev == "i" and cur[0] > 0
                     and w[cur[0] - 1] not in VOW and c in GLIDE_VOW)
            if prev + c in DIPH or glide:
                cur.append(i)
            else:
                groups.append(cur); cur = [i]
        elif cur and c in SON and nxt not in VOW and w[cur[-1]] in SHORT_VOW:
            cur.append(i); groups.append(cur); cur = []
        else:
            if cur: groups.append(cur)
            cur = []
    if cur: groups.append(cur)
    return groups

def mark_word(word, syl, acc):
    groups = nucleus_groups(word)
    if not groups or syl >= len(groups):
        return word                      # cannot place: leave the word bare
    g = groups[syl]; w = word.lower()
    vowels = [i for i in g if w[i] in VOW]
    if (len(vowels) >= 2 and w[vowels[0]] == "i" and w[vowels[1]] in GLIDE_VOW
            and (vowels[0] == 0 or w[vowels[0] - 1] not in VOW)):
        vowels = vowels[1:]              # glide i: the accent belongs to the next vowel
    if acc == "tilde" and len(g) > 1:
        if w[g[-1]] in SON:
            after = w[g[-1] + 1] if g[-1] + 1 < len(w) else ""
            if (len(vowels) >= 2 and w[vowels[0]] + w[vowels[1]] in ("ie", "uo")
                    and not (w[vowels[0]] == "i" and after == "t")):
                pos = vowels[1]
            else:
                pos = g[-1]
        else:
            pos = vowels[1] if len(vowels) >= 2 else g[-1]
    elif acc == "short" and len(vowels) >= 2:
        pos = vowels[1]
    else:
        pos = vowels[0] if vowels else g[0]
    return word[:pos + 1] + COMB[acc] + word[pos + 1:]

Output is NFC-normalised by most toolchains (à becomes one code point). For TTS training data that keeps combining marks separate, normalise the way your corpus does.

Things you should also do

• Do not accent what is not Lithuanian. Acronyms (JAV), digits, single letters (D.) and foreign names: pass them through untouched.
• Do not trust confidence as knowledge. The model returns a confident label for any input; leaving an unknown word unmarked is more useful than a fluent guess.

Attribution

• Base: FacebookAI/xlm-roberta-large (MIT)
• Training data: meldynamics/lt-stressed-corpus, built by Aleksas Pielikis from the MATAS and ALKSNIS corpora (VDU / CLARIN-LT, CC BY 4.0), with the stress added via the VDU Kirtis service and his vdu-nlp-services — ačiū už atvirus duomenis.
• Built by the Kalamo team.

Feedback, and especially hard Lithuanian test cases, are very welcome.