Ericu950 / CharDiff-grc-tagger-parser

CharDiff-grc tagger + parser

A joint morphosyntactic tagging and dependency parsing model for Ancient Greek: factored XPOS position tagging (9 heads decoded jointly), edit-script lemmatization, UPOS tagging, and biaffine dependency parsing, all from a single shared backbone via an ELMo-style scalar mix over all 33 encoder layers. Fine-tuned from Ericu950/CharDiff-grc-doc_clean on Celano's AGDT treebank

Results (test split)

See the paper for the full cross-encoder ablation (this backbone vs. six external subword encoders under the identical fine-tuning recipe).

Usage

import torch
from transformers import AutoModel
from huggingface_hub import snapshot_download

repo = "Ericu950/CharDiff-grc-tagger-parser"

local_dir = snapshot_download(
    repo,
    allow_patterns=["*.json", "*.txt", "*.py", "*.model"],
)

import sys
sys.path.insert(0, local_dir)
from processing_char_bert_joint import CharBertJointProcessor

model = AutoModel.from_pretrained(repo, trust_remote_code=True)
processor = CharBertJointProcessor.from_pretrained(local_dir)
model.eval()

words = "μῆνιν ἄειδε θεὰ Πηληϊάδεω Ἀχιλῆος".split()
batch = processor([words])
with torch.no_grad():
    out = model(**batch)

rows = processor.decode(out, batch)


def print_parse(decoded):
    sentences = [decoded] if decoded and isinstance(decoded[0], dict) else decoded

    cols = ["id", "form", "lemma", "upos", "xpos", "head", "deprel"]
    for s_idx, sent in enumerate(sentences, start=1):
        table = [
            {
                "id": str(i),
                "form": r["form"],
                "lemma": r["lemma"],
                "upos": r["upos"],
                "xpos": r["xpos"],
                "head": str(r["head"]),
                "deprel": r["deprel"],
            }
            for i, r in enumerate(sent, start=1)
        ]
        widths = {c: max(len(c), *(len(row[c]) for row in table)) for c in cols}
        header = "  ".join(c.upper().ljust(widths[c]) for c in cols)
        if len(sentences) > 1:
            print(f"\n# sentence {s_idx}")
        print(header)
        print("-" * len(header))
        for row in table:
            print("  ".join(row[c].ljust(widths[c]) for c in cols))


print_parse(processor.decode(out, batch))

ID  FORM       LEMMA      UPOS  XPOS       HEAD  DEPREL
-------------------------------------------------------
1   μῆνιν      μῆνις      n     n-s---fa-  2     OBJ   
2   ἄειδε      ἀείδω      v     v2spma---  0     PRED  
3   θεὰ        θεά        n     n-s---fv-  2     ExD   
4   Πηληϊάδεω  Πηληιάδης  n     n-s---mg-  5     ATR   
5   Ἀχιλῆος    Ἀχιλλεύς   n     n-s---mg-  1     ATR   

Unlike the base backbone, this model's input includes a capitalization channel (fine-tune-only, zero-initialized at pretraining time) and operates on whole, word-tokenized sentences rather than raw unsegmented text — the processor expects a list of word strings (or a list of sentences, each a list of words) and handles the character-level packing and word-level pooling internally.

Citation

See the CharDiff-grc repository and the associated paper (citation to be added on publication).