Jgmorenof/MILRIT_AT_CLEF2026

MILRIT @ CLEF HIPE-2026 — Multi-view Person–Place Relation Classifier

This is the submitted model of team MILRIT for the CLEF HIPE-2026 shared task on Person–Place Relation Extraction from multilingual (DE/EN/FR) historical newspaper text.

The task is pairwise classification of (person, location) mentions extracted from a document into:

• at ∈ {TRUE, PROBABLE, FALSE} — is the person textually linked to the place?
• isAt ∈ {TRUE, FALSE} — is the person currently there, within ~1 month of the publication date?

Model

• Base encoder: microsoft/mdeberta-v3-base, weight-shared (Siamese) across views
• Heads: at_head (Linear 768→3), isAt_head (Linear 768→2), on the [CLS] vector
• Training: multi-view self-distillation — each training pair is presented in three parallel text forms (raw OCR text, LLM chain-of-thought analysis, LLM self-consistency analysis), sharing one encoder, optimized jointly with cross-entropy on all views plus a JSD consistency loss and an InfoNCE contrastive loss across views.
• Inference: uses the raw OCR text view only (encode_single) — the auxiliary views and the projection head are training-time-only scaffolding to push reasoning-derived signal into the encoder without needing an LLM at inference time.
• Checkpoint: epoch 7, dev-set global macro Recall = 0.6618

Files

• config.json, model.safetensors — HF-format weights for MilritMultiViewForRelationClassification, the inference-only port of the training architecture (encoder + at_head + isAt_head; the contrastive proj head is kept in the weights for completeness but unused at inference).

• configuration_milrit.py, modeling_milrit.py, __init__.py — custom PretrainedConfig/PreTrainedModel classes wired into config.json via auto_map, so the model loads with trust_remote_code=True:

  from transformers import AutoModel, AutoTokenizer
  
  tok = AutoTokenizer.from_pretrained("jgmorenof/milrit-hipe2026-multiview", trust_remote_code=True)
  model = AutoModel.from_pretrained("jgmorenof/milrit-hipe2026-multiview", trust_remote_code=True)
  

• spm.model, tokenizer.json, tokenizer_config.json, special_tokens_map.json, added_tokens.json — mDeBERTa-v3 tokenizer, extended with task-specific special tokens

• train_results.json — best epoch and dev-set global macro Recall

Citation

If you use this model, please cite:

@inproceedings{pham2026milrit,
  author    = {Ha Ngan Pham and Jose G. Moreno and Antoine Doucet},
  title     = {{MILRIT} at {HIPE} 2026: From Generalist Relation Extraction to Multi-view Learning for Historical Person--Place Relations},
  booktitle = {CLEF 2026 Working Notes},
  series    = {CEUR Workshop Proceedings},
  address   = {Jena, Germany},
  month     = sep,
  year      = {2026}
}