eval_utils.py

import json
import re
from collections import defaultdict

import evaluate
import nltk
import numpy as np

from nervaluate import Evaluator
# from rouge_score import rouge_scorer
from sacrebleu.metrics import BLEU, CHRF
from sklearn.metrics import f1_score
from tqdm import tqdm
from transformers import AutoTokenizer
import rouge
import bert_score
import string


def load_json(file_path):
    with open(file_path, "r") as f:
        return json.load(f)


def get_micro_at_k(gold, pred, k):
    gold_set = set(gold)
    pred_set = set(pred[:k])
    return len(gold_set & pred_set), len(gold_set), len(pred_set)


def evaluate_bail(gold_data, pred_data):
    gold_labels = []
    pred_labels = []
    for id, label in gold_data.items():
        gold_labels.append(label)
        pred_labels.append(pred_data.get(id, 0))

    f1 = f1_score(gold_labels, pred_labels, average="macro")
    print("Macro-F1 on HLDC-all-districts test set:", f1)
    return {"mF1": f1}

def get_BLEU_score(ref_text_all, machine_text_all):
    sc_all = []
    for i in range(len(ref_text_all)):
        ref_text = ref_text_all[i]
        machine_text = machine_text_all[i]
        tok_ref_text = nltk.word_tokenize(ref_text)
        tok_machine_text = nltk.word_tokenize(machine_text)
        sc = nltk.translate.bleu_score.sentence_bleu([tok_ref_text], tok_machine_text, weights = (0.5,0.5))
        sc_all.append(sc)
    return sum(sc_all)/len(sc_all)

def evaluate_cjpe(gold_data, pred_data):
    # Evaluate prediction
    gold_labels = []
    pred_labels = []
    for id, label in gold_data["prediction"].items():
        gold_labels.append(label)
        pred_labels.append(pred_data["prediction"].get(id, 0))

    f1 = f1_score(gold_labels, pred_labels, average="macro")
    prediction_result = {"cjpe-eval": f1}
    print("Macro-F1 on ILDC test:", prediction_result)
    
    R = []
    B = []
    rl_evaluator = rouge.Rouge(metrics=['rouge-l'], max_n=2, limit_length=False, apply_avg=True)
    for x in range(1, 6):
        gold_explanations = []
        pred_explanations = []
        for k,v in gold_data['explanation'].items():
            gold_explanations.append(v[f'expert_{x}'])
            pred_explanations.append(pred_data['explanation'][k])
        print("Metrics for expert", x, "...", end=' ')
        rougex = rl_evaluator.get_scores(pred_explanations, gold_explanations)['rouge-l']['f']
        bleux = get_BLEU_score(gold_explanations, pred_explanations)
        R.append(rougex)
        B.append(bleux)
        print("Done.")
        
    
    rouge_score = sum(R)/len(R)
    bleu_score = sum(B)/len(B)

    explanation_result = {
        "cjpe-exp-eval": {
            "rouge": rouge_score,
            "bleu": bleu_score,
        }
    }
    print("Explanability for ILDC Expert:", explanation_result)
    #return {**prediction_result, **explanation_result}
    return {"mF1": f1, "ROUGE-L": rouge_score, "BLEU": bleu_score}

def span2bio(txt, roles):
    roles = sorted(roles, key = lambda x:x['start'])        
    roles_left = [r['start'] for r in roles]    
    
    ttxt = re.findall(r'[{}]|\w+'.format(string.punctuation), txt)
                            
    c = 0
    cr = -1
    prev = 'O'
    troles = []
    for tok in ttxt:
        if c >= len(txt):
            break
        
        while txt[c] == ' ':
            c += 1
        
        else:
            if c in roles_left: # Start of a new role
                ind = roles_left.index(c)
                cr = roles[ind]['end']
                prev = 'I-' + roles[ind]['label']
                troles.append('B-' + roles[ind]['label'])
            else:
                if c < cr: # Assign previous role
                    troles.append(prev)
                else: # Assign 'O'
                    troles.append('O')
            
            c += len(tok)
        
    if len(ttxt) != len(troles):
        troles += ['O'] * (len(ttxt) - len(troles))
        
    assert len(ttxt) == len(troles)
    return ttxt, troles

def evaluate_lner(gold_data, pred_data, text_data):
    with open("ner_labels.txt") as f:
        labels = f.read().strip().split("\n")

    results_per_fold = {}
    for fold in range(1, len(gold_data) + 1):
        gold = gold_data[f"fold_{fold}"]
        pred = pred_data[f"fold_{fold}"]
        text = text_data[f"fold_{fold}"]

        texts, gold_labels, pred_labels = [], [], []

        for id, gold_label in tqdm(gold.items()):
            txt = text[id]
            pred_label = pred.get(id, [])

            txt_seg, gold_bio = span2bio(txt, gold_label)
            _, pred_bio = span2bio(txt, pred_label)

            texts.append(txt_seg)
            gold_labels.append(gold_bio)
            pred_labels.append(pred_bio)

        evaluator = Evaluator(gold_labels, pred_labels, tags=labels, loader="list")

        results, results_per_tag, _, _ = evaluator.evaluate()

        f1_scores = [results_per_tag[l]["strict"]["f1"] for l in results_per_tag]
        avg_f1 = sum(f1_scores) / len(f1_scores)
        print(f"Strict Macro-F1 on Fold {fold}:", avg_f1)
        results_per_fold[f"fold_{fold}"] = avg_f1

    print("Strict macro-F1 on L-NER Dataset:", results_per_fold)
    return {"strict mF1": sum(results_per_fold.values())/len(results_per_fold)}


def evaluate_rr(gold_data, pred_data):
    all_gold_labels = []
    all_pred_labels = []
    with open("rr_label_vocab.json") as f:
        label_vocab = json.load(f)


    for id, gold_labels in gold_data.items():
        pred_labels = pred_data.get(id, ["None"] * len(gold_labels))
        for i in range(len(gold_labels)):
            g = gold_labels[i]
            p = pred_labels[i]
            if g not in label_vocab: continue
            for pp in p.split():
                if pp in label_vocab:
                    p = pp
                    break
            if p not in label_vocab: continue
            all_gold_labels.append([label_vocab[g]])
            all_pred_labels.append([label_vocab[p]])

    f1 = f1_score(all_gold_labels, all_pred_labels, average="macro")
    print(f"Macro-F1 on combined test set:", f1)
    return {"mF1": f1}


def evaluate_lsi(gold_data, pred_data):
    with open("lsi_label_vocab.json") as f:
        label_vocab = json.load(f)

    gold_matrix = np.zeros((len(gold_data), len(label_vocab)))
    pred_matrix = np.zeros((len(gold_data), len(label_vocab)))

    for i, (id, gold_labels) in enumerate(gold_data.items()):
        pred_labels = pred_data.get(id, [])

        for label in gold_labels:
            if label in label_vocab:
                gold_matrix[i, label_vocab[label]] = 1

        for label in pred_labels:
            if label in label_vocab:
                pred_matrix[i, label_vocab[label]] = 1

    f1 = f1_score(gold_matrix, pred_matrix, average="macro")
    print("Macro-F1 on ILSI test set:", f1)
    return {"mF1": f1}


def evaluate_pcr(gold_data, pred_data):
    f1_scores = []
    for k in range(1, 21):
        correct, gold_total, pred_total = 0, 0, 0
        for id, gold_candidates in tqdm(gold_data.items(), desc="pcr"):
            pred_candidates = pred_data.get(id, [])
            gold_candidates = [c for c in gold_candidates if c != id]
            pred_candidates = [c for c in pred_candidates if c != id]

            c, g, p = get_micro_at_k(gold_candidates, pred_candidates, k)
            correct += c
            gold_total += g
            pred_total += p

        precision = correct / pred_total if pred_total > 0 else 0
        recall = correct / gold_total if gold_total > 0 else 0
        f1 = (
            2 * precision * recall / (precision + recall)
            if precision + recall > 0
            else 0
        )
        f1_scores.append(f1)

        print(f"Micro-F1@{k} on IL-PCR test set:", f1)

    max_f1 = max(f1_scores)
    index_max = f1_scores.index(max_f1) + 1
    return {"muF1@K": f"{max_f1:.2f}@{index_max}"}


def evaluate_summ(gold_data, pred_data):
    gold_summaries = []
    pred_summaries = []

    for id, gold_summary in gold_data.items():
        if id in pred_data:
            gold_summary = re.sub(r"\s+", " ", gold_summary.replace("\n", " ")).strip()
            pred_summary = re.sub(r"\s+", " ", pred_data[id].replace("\n", " ")).strip()

            gold_summaries.append(gold_summary)
            pred_summaries.append(pred_summary)

    
    rl_evaluator = rouge.Rouge(metrics=['rouge-n','rouge-l'], max_n=2, limit_length=False, apply_avg=True)
    rl_scores = rl_evaluator.get_scores(pred_summaries, gold_summaries)
    print("Rouge:", {k:v['f'] for k,v in rl_scores.items()}, flush=True)
    
    _, _, bs = bert_score.score(pred_summaries, gold_summaries, lang="en", verbose=True, device='cuda')
    print("BERTSCORE:", bs.mean().item())
    return {'ROUGE-L': rl_scores['rouge-l']['f'], 'BERTSCORE': bs.mean().item()}


def evaluate_lmt(gold_data, pred_data):
    tokenizer = AutoTokenizer.from_pretrained("ai4bharat/indic-bert")
    bleu = BLEU()
    chrfp = CHRF(word_order=2)
    gleu = evaluate.load("google_bleu")

    G = defaultdict(lambda: defaultdict(list))
    P = defaultdict(lambda: defaultdict(list))

    for dataset in gold_data:
        for id, gold_text in gold_data[dataset].items():
            lang = id.split("/")[1].strip()
            gold_tokens = " ".join(tokenizer.tokenize(gold_text))
            pred_tokens = " ".join(tokenizer.tokenize(pred_data[dataset][id]))
            G[dataset][lang].append(gold_tokens)
            P[dataset][lang].append(pred_tokens)

    bleu_scores, chrfpp_scores, gleu_scores = [], [], []

    for dataset in G:
        print("Dataset", dataset)
        dataset_bleu, dataset_chrfpp, dataset_gleu = [], [], []

        for lang in G[dataset]:
            gold = G[dataset][lang]
            pred = P[dataset][lang]

            bleu_score = bleu.corpus_score(pred, [gold]).score
            chrfpp_score = chrfp.corpus_score(pred, [gold]).score
            gleu_score = gleu.compute(predictions=pred, references=gold)["google_bleu"]

            dataset_bleu.append(bleu_score)
            dataset_chrfpp.append(chrfpp_score)
            dataset_gleu.append(gleu_score)

        bleu_scores.append(sum(dataset_bleu) / len(dataset_bleu))
        chrfpp_scores.append(sum(dataset_chrfpp) / len(dataset_chrfpp))
        gleu_scores.append(sum(dataset_gleu) / len(dataset_gleu))

    return {
        "BLEU": sum(bleu_scores) / len(bleu_scores),
        "GLEU": sum(gleu_scores) / len(gleu_scores),
        "chrF++": sum(chrfpp_scores) / len(chrfpp_scores),
    }


def create_output_json(evaluation_results):
    output = {
        "Method": "GPT-5 (2-shot)",
        "Submitted By": "IL-TUR",
        "Github Link": "dummy submission",
        "L-NER": {"strict mF1": evaluation_results["lner"]["strict mF1"]},
        "RR": {"mF1": evaluation_results["rr"]["mF1"]},
        "CJPE": {
            "mF1": evaluation_results["cjpe"]["mF1"],
            "ROUGE-L": evaluation_results["cjpe"]["ROUGE-L"],
            "BLEU": evaluation_results["cjpe"]["BLEU"],
        },
        "BAIL": {"mF1": evaluation_results["bail"]["mF1"]},
        "LSI": {"mF1": evaluation_results["lsi"]["mF1"]},
        "PCR": {"muF1@K": evaluation_results["pcr"]["muF1@K"]},
        "SUMM": {
            "ROUGE-L": evaluation_results["summ"]["ROUGE-L"],
            "BERTSCORE": evaluation_results["summ"]["BERTSCORE"] #"-",  # Placeholder BERTSCORE
        },
        "L-MT": {
            "BLEU": evaluation_results["lmt"]["BLEU"],
            "GLEU": evaluation_results["lmt"]["GLEU"],
            "chrF++": evaluation_results["lmt"]["chrF++"],
        },
    }
    return [output]  # Wrap in a list to match the desired format


def main():
    # gold_data = load_json("IL_TUR_eval_gold.json")
    # pred_data = load_json("IL_TUR_eval_submission2.json")
    gold_data = load_json("submissions/baseline/IL_TUR_eval_gold.json")
    pred_data = load_json("submissions/baseline/IL_TUR_eval_submission_dummy.json")
    pred_data = gold_data
    evaluation_results = {}

    for task in pred_data.keys():
        print(f"Task: {task}")

        if task == "bail":
            evaluation_results[task] = evaluate_bail(gold_data[task], pred_data[task])
        elif task == "cjpe":
            nltk.download('punkt')
            evaluation_results.update(evaluate_cjpe(gold_data[task], pred_data[task]))
        elif task == "lner":
            text_data = load_json("lner-text.json")
            evaluation_results[task] = evaluate_lner(
                gold_data[task], pred_data[task], text_data
            )
        elif task == "rr":
            evaluation_results[task] = evaluate_rr(gold_data[task], pred_data[task])
        elif task == "lsi":
            evaluation_results[task] = evaluate_lsi(gold_data[task], pred_data[task])
        elif task == "pcr":
            evaluation_results[task] = evaluate_pcr(gold_data[task], pred_data[task])
        elif task == "summ":
            nltk.download('punkt')
            evaluation_results[task] = evaluate_summ(gold_data[task], pred_data[task])
        elif task == "lmt":
            evaluation_results[task] = evaluate_lmt(gold_data[task], pred_data[task])

    # convert the evaluation results to the required format
    for task, result in evaluation_results.items():
        if isinstance(result, dict):
            for subtask, subresult in result.items():
                if isinstance(subresult, dict):
                    for subsubtask, subsubresult in subresult.items():
                        evaluation_results[task][subtask][
                            subsubtask
                        ] = f"{subsubresult:.2f}"
                else:
                    if isinstance(subresult, str):
                        evaluation_results[task][subtask] = subresult
                    else:
                        evaluation_results[task][subtask] = f"{subresult:.2f}"
        else:
            if isinstance(result, str):
                evaluation_results[task] = result
            else:
                evaluation_results[task] = f"{result:.2f}"

    blank_scores = {
        "lner": {"strict mF1": "-"},
        "rr": {"mF1": "-"},
        "cjpe": {"mF1": "-", "ROUGE-L": "-", "BLEU": "-"},
        "bail": {"mF1": "-"},
        "lsi": {"mF1": "-"},
        "pcr": {"muF1@K": "-"},
        "summ": {"ROUGE-L": "-", "BERTSCORE": "-"},
        "lmt": {"BLEU": "-", "GLEU": "-", "chrF++": "-"},
    }

    print("--------------------------Evaluation Summary--------------------------")
    for task, result in evaluation_results.items():
        print(f"{task}: {result}")
    print("---------------------------------------------------------------------")

    # for tasks that were not present in the submission, add blank scores
    for task in gold_data.keys():
        if task not in pred_data:
            evaluation_results[task] = blank_scores[task]

    # Generate the output JSON
    output_json = create_output_json(evaluation_results)
    with open("evaluation_results.json", "w") as f:
        json.dump(output_json, f, indent=2)
    print("Evaluation results saved to evaluation_results.json")


def get_evaluation_scores(gold_data, submission_data):
    evaluation_results = {}

    for task in submission_data.keys():
        print(f"Task: {task}")

        if task == "bail":
            evaluation_results[task] = evaluate_bail(
                gold_data[task], submission_data[task]
            )
        elif task == "cjpe":
            nltk.download('punkt')
            evaluation_results.update(
                evaluate_cjpe(gold_data[task], submission_data[task])
            )
        elif task == "lner":
            text_data = load_json("lner-text.json")
            evaluation_results[task] = evaluate_lner(
                gold_data[task], submission_data[task], text_data
            )
        elif task == "rr":
            evaluation_results[task] = evaluate_rr(
                gold_data[task], submission_data[task]
            )
        elif task == "lsi":
            evaluation_results[task] = evaluate_lsi(
                gold_data[task], submission_data[task]
            )
        elif task == "pcr":
            evaluation_results[task] = evaluate_pcr(
                gold_data[task], submission_data[task]
            )
        elif task == "summ":
            nltk.download('punkt')
            evaluation_results[task] = evaluate_summ(
                gold_data[task], submission_data[task]
            )
        elif task == "lmt":
            evaluation_results[task] = evaluate_lmt(
                gold_data[task], submission_data[task]
            )

    # convert the evaluation results to the required format
    for task, result in evaluation_results.items():
        if isinstance(result, dict):
            for subtask, subresult in result.items():
                if isinstance(subresult, dict):
                    for subsubtask, subsubresult in subresult.items():
                        evaluation_results[task][subtask][
                            subsubtask
                        ] = f"{subsubresult:.2f}"
                else:
                    if isinstance(subresult, str):
                        evaluation_results[task][subtask] = subresult
                    else:
                        evaluation_results[task][subtask] = f"{subresult:.2f}"
        else:
            if isinstance(result, str):
                evaluation_results[task] = result
            else:
                evaluation_results[task] = f"{result:.2f}"

    blank_scores = {
        "lner": {"strict mF1": "-"},
        "rr": {"mF1": "-"},
        "cjpe": {"mF1": "-", "ROUGE-L": "-", "BLEU": "-"},
        "bail": {"mF1": "-"},
        "lsi": {"mF1": "-"},
        "pcr": {"muF1@K": "-"},
        "summ": {"ROUGE-L": "-", "BERTSCORE": "-"},
        "lmt": {"BLEU": "-", "GLEU": "-", "chrF++": "-"},
    }

    # for tasks that were not present in the submission, add blank scores
    for task in gold_data.keys():
        if task not in submission_data:
            evaluation_results[task] = blank_scores[task]

    print("--------------------------Evaluation Summary--------------------------")
    for task, result in evaluation_results.items():
        print(f"{task}: {result}")
    print("---------------------------------------------------------------------")
    output_json = create_output_json(evaluation_results)

    return output_json


if __name__ == "__main__":
    main()