{
    "prompt": "從以下政府文本中，提取指定的實體類別，并按照JSON格式輸出結果。  xxx",      
    "predict": "\n```json\n{\n  \"行政區劃和地理實體\": [],\n  \"政府機構和組織\": [],\n ... \"特定術語和關鍵詞\": [\"改革創新\", \"體制機制障礙\", \"行政管理體制\", \"城鄉一體化發展機制\"]\n}\n```", 
    "label": "```json\n{\n  \"行政區劃和地理實體\": [],\n  \"政府機構和組織\": [], ... \"特定術語和關鍵詞\": [\"改革創新發展\", \"高質量發展\", \"高品質生活\", \"體制機制障礙\"]\n}\n```"
    ...
}

import os
import re
import json
from dataclasses import dataclass

@dataclass
class EvalNode:
    # 默認值
    predict_right_num: int = 0
    predict_num: int = 0
    label_num: int = 0

    @property
    def accuracy(self):
        return self.predict_right_num / (self.predict_num + 1e-6)

    @property
    def recall(self):
        return self.predict_right_num / (self.label_num + 1e-6)

    @property
    def f1(self):
        return 2 * self.accuracy * self.recall / (self.accuracy + self.recall + 1e-6)

@dataclass
class NEREvaluate:
    ent_class: list[str]
    input_file: str
    avg_accuracy: int = 0
    avg_recall: int = 0
    avg_f1: int = 0
    predict_error: int = 0

    def __post_init__(self):
        self._evaluate_by_jsonl()

    def _evaluate_by_jsonl(self):

        with open(self.input_file, "r", encoding="utf-8") as f:
            self.total_ent = {ent: EvalNode() for ent in self.ent_class}

            for line in f:
                data = json.loads(line)

                # 大模型采取的是序列到序列到文本生成，不能轉換為字典的數據跳過即可
                try:
                    matches = re.search("\{.*?\}", data["predict"], re.DOTALL)
                    if matches:
                        predict = eval(matches.group(0))
                except:
                    self.predict_error += 1
                    predict = {ent: [] for ent in self.ent_class}

                try:
                    # 防止label出錯
                    matches = re.search("\{.*?\}", data["label"], re.DOTALL)
                    if matches:
                        labels = eval(matches.group(0))
                except:
                    continue

                # 每個不同的實體類別單獨計數
                for ent_name in self.ent_class:
                    label_set = set(labels.get(ent_name, []))
                    predict_set = set(predict.get(ent_name, []))
                    self.total_ent[ent_name].predict_right_num += len(
                        label_set & predict_set
                    )
                    self.total_ent[ent_name].predict_num += len(predict_set)
                    self.total_ent[ent_name].label_num += len(label_set)

        for ent in self.ent_class:
            self.avg_accuracy += self.total_ent[ent].accuracy
            self.avg_recall += self.total_ent[ent].recall
            self.avg_f1 += self.total_ent[ent].f1

        self.avg_accuracy = self.avg_accuracy / len(self.ent_class)
        self.avg_recall = self.avg_recall / len(self.ent_class)
        self.avg_f1 = self.avg_f1 / len(self.ent_class)

ENT_CLASS = [
    "行政區劃和地理實體",
    "政府機構和組織",
    "法律法規與政策",
    "人物與職務",
    "產業",
    "產業園區",
    "高新技術",
    "項目與工程",
    "社會服務與基礎設施",
    "特定術語和關鍵詞",
]

ner_eval = NEREvaluate(
        ENT_CLASS, "xxx/generated_predictions.jsonl")
    )

for key, item in ner_eval.total_ent.items():
    print(key, item.f1)

行政區劃和地理實體 0.6292412411907403
人物與職務 0.45714235316380664
...
特定術語和關鍵詞 0.29256472295174