MHIPEX

Multilingual Historical Person–Place Relation Extraction
A Calibrated Transformer Ensemble with Relation-Specific Language-Adaptive Routing

---

Overview

MHIPEX is a research framework for extracting person–place relations from multilingual historical newspaper archives. Given a (person, location) entity pair and its surrounding newspaper article, MHIPEX classifies two distinct relations:

The system operates across three languages (English, French, German) and is evaluated using macro-recall (MR) across both relations — a metric that penalizes models which ignore rare classes.

Authors: Aman Jaiswal & Dr. Sarika Jain
Affiliation: National Institute of Technology Kurukshetra, India
Target Journal: Knowledge-Based Systems (Elsevier)

---

Key Results

Official HIPE-2026 Test Set

Development Set (2,081 pairs)

• +43.9% relative improvement over mBERT baseline
• +6.0% relative improvement over best single model (mDeBERTa-v3)
• Statistical significance confirmed: McNemar's χ² = 40.01, p = 2.53 × 10⁻¹⁰

---

Architecture

MHIPEX employs a dual-encoder architecture with post-hoc ensemble routing:

 Newspaper Article + Entity Pair
              │
    ┌─────────┴─────────┐
    ▼                   ▼
 Input Enrichment    Input Enrichment
 (<P>, <L>, <DATE>,  (<P>, <L>, <DATE>,
  <LANG> markers)     <LANG> markers)
    │                   │
    ▼                   ▼
 mDeBERTa-v3         XLM-RoBERTa
  (278M)              Large (560M)
    │                   │
 CLS + Mean Pooling  CLS + Mean Pooling
 + Multi-Sample      + Multi-Sample
   Dropout (K=3)       Dropout (K=3)
    │                   │
 Dual Heads          Dual Heads
 (at: 3-way,         (at: 3-way,
  isAt: 2-way)        isAt: 2-way)
    │                   │
    └─────────┬─────────┘
              ▼
   RLAE: Relation-Specific
   Language-Adaptive Ensemble
   (per-relation, per-language
    β weights + τ thresholds)
              │
              ▼
       Final Predictions

What Makes RLAE Novel

RLAE (Relation-Specific Language-Adaptive Ensemble) is a conditional Mixture-of-Experts gating mechanism that learns independent mixing weights for each (relation type × language) combination. Unlike global ensembling:

• Spatial reasoning (at) benefits from mDeBERTa-v3's disentangled attention
• Temporal reasoning (isAt) varies by language — XLM-R Large dominates for French, while mDeBERTa-v3 is stronger for German
• The weight matrix β is fully interpretable and requires no model retraining

---

Experiments

All experiments are designed to run on Kaggle with 2 × NVIDIA T4 GPUs (16 GB VRAM each). Each script is fully self-contained — paste into a single notebook cell and run.

Quick Start (Kaggle)

1. Open Kaggle → New Notebook
2. Set Accelerator → GPU T4 × 2, Internet → ON
3. Paste the contents of any .py script into a single cell
4. Run — data downloads automatically from this repository

Quick Start (Local)

git clone https://github.com/amanraj74/MHIPEX.git
cd MHIPEX
pip install torch transformers datasets scikit-learn pandas tqdm
python mhipex_master_pipeline.py

---

Project Structure

MHIPEX/
│
├── paper/
│   ├── main.tex                              # Full paper source (journal format)
│   ├── compile_pdf.py                        # LaTeX → PDF with embedded figures
│   └── figures/
│       ├── architecture1.png                 # System architecture diagram
│       ├── fig1_label_distribution.png       # Class distribution visualization
│       ├── fig2_confusion_matrices.png       # Confusion matrices (at + isAt)
│       └── fig6_error_analysis.png           # Error category breakdown
│
├── data/
│   ├── de-train.jsonl / de-dev.jsonl         # German data
│   ├── en-train.jsonl / en-dev.jsonl         # English data
│   └── fr-train.jsonl / fr-dev.jsonl         # French data
│
├── mhipex_master_pipeline.py                 # End-to-end: train → evaluate → report
├── kaggle_mhipex_v31_a1.py                   # Training cell 1: setup + mDeBERTa-v3
├── kaggle_mhipex_v31_a2.py                   # Training cell 2: XLM-R Large
├── kaggle_mhipex_rlae.py                     # RLAE weight optimization
├── kaggle_mhipex_ablations.py                # Ablation study (A0–A6)
├── kaggle_mhipex_crossval.py                 # Cross-dataset / zero-shot transfer
├── kaggle_mhipex_entity_marker_baseline.py   # Soares et al. entity-marker baseline
├── kaggle_mhipex_kg.py                       # Single KG augmentation experiment
├── kaggle_mhipex_multikg.py                  # Multi-KG comparison experiment
├── kaggle_mhipex_ocr_robustness.py           # OCR noise robustness experiment
├── run_mcnemar.py                            # Statistical significance test
├── gen_architecture.py                       # Architecture figure generator
│
├── *_results.csv                             # Experiment result files
├── results/                                  # Official test set metrics
└── README.md

---

Requirements

---

Paper

Title: MHIPEX: A Calibrated Transformer Ensemble for Multilingual Person–Place Relation Extraction in Historical Newspapers

Authors: Aman Jaiswal, Sarika Jain (NIT Kurukshetra)

Paper Structure

Key Tables

---

Citation

@article{jaiswal2026mhipex,
  title     = {MHIPEX: A Calibrated Transformer Ensemble for Multilingual
               Person--Place Relation Extraction in Historical Newspapers},
  author    = {Jaiswal, Aman and Jain, Sarika},
  journal   = {Knowledge-Based Systems},
  publisher = {Elsevier},
  year      = {2026},
  note      = {Under review}
}

---

License

This project is released under the MIT License. If you use any part of this work in your research, please cite our paper.

---

Built with ❤️ at NIT Kurukshetra