evaluate_models.py

"""
Azerbaijani NER Benchmark Evaluation Script

Evaluates all four Azerbaijani NER models on the benchmark test set
and prints a comparison table with precision, recall, and F1-score.

Usage:
    pip install transformers datasets seqeval torch
    python evaluate_models.py
"""

from datasets import load_dataset
from transformers import pipeline, AutoTokenizer, AutoModelForTokenClassification
from seqeval.metrics import precision_score, recall_score, f1_score, classification_report
import torch

# ─── Models to evaluate ───────────────────────────────────────────────────────
MODELS = [
    {
        "name": "mBERT Azerbaijani NER",
        "repo": "IsmatS/mbert-az-ner",
        "params": "180M",
    },
    {
        "name": "XLM-RoBERTa Base",
        "repo": "IsmatS/xlm-roberta-az-ner",
        "params": "125M",
    },
    {
        "name": "XLM-RoBERTa Large",
        "repo": "IsmatS/xlm_roberta_large_az_ner",
        "params": "355M",
    },
    {
        "name": "Azeri-Turkish BERT",
        "repo": "IsmatS/azeri-turkish-bert-ner",
        "params": "110M",
    },
]

# IOB2 label mapping (index → tag string used in the benchmark dataset)
ID2LABEL_BENCHMARK = {
    0: "O",
    1: "B-PERSON",
    2: "I-PERSON",
    3: "B-LOCATION",
    4: "I-LOCATION",
    5: "B-ORGANISATION",
    6: "I-ORGANISATION",
    7: "B-DATE",
    8: "I-DATE",
}


def load_benchmark():
    """Load the benchmark test split."""
    print("Loading Azerbaijani NER Benchmark …")
    dataset = load_dataset("IsmatS/azerbaijani-ner-benchmark", split="test")
    print(f"  Loaded {len(dataset)} sentences.\n")
    return dataset


def align_predictions(predictions, label_ids, id2label):
    """
    Convert token-level model outputs to word-level IOB2 tag sequences,
    aligning with the gold labels from the benchmark.

    Args:
        predictions: list of dicts returned by the HF pipeline
                     (word-level, after aggregation_strategy='first')
        label_ids:   list[int] of gold NER tag ids for one sentence
    Returns:
        pred_tags:   list[str] of predicted tag strings
        gold_tags:   list[str] of gold tag strings
    """
    gold_tags = [id2label.get(lid, "O") for lid in label_ids]
    # The pipeline with aggregation_strategy='first' returns one entry per word
    pred_tags = [p["entity"] for p in predictions]

    # Truncate or pad to match gold length
    min_len = min(len(pred_tags), len(gold_tags))
    pred_tags = pred_tags[:min_len]
    gold_tags = gold_tags[:min_len]

    return pred_tags, gold_tags


def evaluate_model(model_info, dataset):
    """
    Run a single model over the full benchmark and return seqeval metrics.
    """
    repo = model_info["repo"]
    print(f"  Loading model: {repo}")

    device = 0 if torch.cuda.is_available() else -1

    ner = pipeline(
        "token-classification",
        model=repo,
        tokenizer=repo,
        aggregation_strategy="first",
        device=device,
    )

    all_pred_tags = []
    all_gold_tags = []

    for i, example in enumerate(dataset):
        tokens = example["tokens"]
        label_ids = example["ner_tags"]
        text = " ".join(tokens)

        try:
            predictions = ner(text)
        except Exception:
            # On error fall back to all-O predictions
            predictions = [{"entity": "O", "word": t} for t in tokens]

        pred_tags, gold_tags = align_predictions(
            predictions, label_ids, ID2LABEL_BENCHMARK
        )
        all_pred_tags.append(pred_tags)
        all_gold_tags.append(gold_tags)

        if (i + 1) % 500 == 0:
            print(f"    Processed {i + 1}/{len(dataset)} examples …")

    precision = precision_score(all_gold_tags, all_pred_tags)
    recall = recall_score(all_gold_tags, all_pred_tags)
    f1 = f1_score(all_gold_tags, all_pred_tags)

    return {
        "precision": precision,
        "recall": recall,
        "f1": f1,
        "detailed_report": classification_report(all_gold_tags, all_pred_tags),
    }


def print_results_table(results):
    """Print a formatted comparison table."""
    header = f"{'Model':<35} {'Params':>8}  {'Precision':>10}  {'Recall':>8}  {'F1-Score':>9}"
    separator = "-" * len(header)

    print("\n" + separator)
    print("AZERBAIJANI NER BENCHMARK — MODEL COMPARISON")
    print(separator)
    print(header)
    print(separator)

    for model_info, metrics in results:
        name = model_info["name"]
        params = model_info["params"]
        p = metrics["precision"] * 100
        r = metrics["recall"] * 100
        f = metrics["f1"] * 100
        print(f"{name:<35} {params:>8}  {p:>9.2f}%  {r:>7.2f}%  {f:>8.2f}%")

    print(separator)

    # Highlight the best model
    best = max(results, key=lambda x: x[1]["f1"])
    print(f"\nBest model: {best[0]['name']}  (F1 = {best[1]['f1']*100:.2f}%)\n")


def main():
    dataset = load_benchmark()
    results = []

    for model_info in MODELS:
        print(f"\nEvaluating: {model_info['name']}")
        metrics = evaluate_model(model_info, dataset)
        results.append((model_info, metrics))

        # Print per-entity-type report
        print(f"\n  Detailed report for {model_info['name']}:")
        for line in metrics["detailed_report"].splitlines():
            print("    " + line)

    print_results_table(results)


if __name__ == "__main__":
    main()