causal-span-mdeberta

A multilingual BIO token classifier that tags cause, effect and signal spans, fine-tuned from microsoft/mdeberta-v3-base on the Causal News Corpus (CC0-1.0). Direction is encoded in the label TYPE, so cause -> effect is read straight from the tags -- reversed phrasing like "the crash resulted from brake failure" is handled without a separate orientation lexicon.

Built for reasongraph as the accurate, model-based causal extractor.

Labels

O
B-CAUSE  I-CAUSE     the cause span
B-EFFECT I-EFFECT    the effect span
B-SIGNAL I-SIGNAL    the causal connective

Usage (transformers)

from transformers import AutoTokenizer, AutoModelForTokenClassification
import torch

tok = AutoTokenizer.from_pretrained("Berk/causal-span-mdeberta")
model = AutoModelForTokenClassification.from_pretrained("Berk/causal-span-mdeberta").eval()

enc = tok("Heavy rainfall caused severe flooding.", return_tensors="pt")
ids = model(**enc).logits[0].argmax(-1).tolist()
toks = tok.convert_ids_to_tokens(enc["input_ids"][0])
print([(t, model.config.id2label[i]) for t, i in zip(toks, ids)])

An ONNX build (onnx/model.onnx, inputs input_ids + attention_mask, output per-token logits) is included for fast CPU inference via onnxruntime -- this is what reasongraph loads.

Evaluation

Benchmarked on the Causal News Corpus (CNC) Subtask 2 dev set. This is an honest first-pass model: it is below the shared-task baseline and SOTA. Its value is robust multilingual zero-shot extraction, not leaderboard rank.

Official CNC scorer (evaluation/subtask2: best-combination alignment + FairEval), V2 dev,
with best-span decoding (one highest-confidence span per role):

  Overall  F1 0.550   (precision 0.649, recall 0.477)
  Cause    F1 0.48  |  Effect  F1 0.55  |  Signal  F1 0.63
  Multi-relation sentences: F1 0.315  (this model predicts ONE relation per sentence)

  (naive argmax decoding scores 0.475; best-span is the recommended decode and is
   what reasongraph's consumer applies -- the ONNX model is identical.)

Shared-task context (official scorer):
  1Cademy (2022 winner, test)         0.542   <- this model (0.550 dev) is above it
  Organizer baseline (2023, dev)     ~0.627   <- this model is ~8 F1 below
  BoschAI (2023 winner, test)         0.728

Honest positioning: a strong first-pass BIO tagger, competitive with the 2022 field but
below the 2023 baseline and SOTA. The baseline uses a span-pointer + beam-search +
signal-detector architecture; matching it needs that architecture. This model's real
strength is robust multilingual zero-shot extraction.

Own metrics (clean dev): strict exact-span seqeval micro F1 0.48 ; token-level F1 0.79

Multilingual capability comes from mDeBERTa-v3's zero-shot cross-lingual transfer; the training spans are English only.

Training data

Causal News Corpus V2, subtask 2 (CC0-1.0). ARG0 = cause, ARG1 = effect, SIG* = signal. Tags can nest (a signal inside an argument) and a sentence may carry several relations.

Limitations

• Trained on English news text; other languages rely on zero-shot transfer and are less accurate, especially for distant scripts (e.g. Chinese, Arabic).
• News-domain bias; short, explicit causal statements are handled best.
• Multi-relation sentences receive one predicted causal structure.

License

MIT (weights and code). Training data is CC0-1.0.