文本分类基线：把文本变成可评估的预测结果

log P(类别) + Σ 特征次数 × log P(特征 | 类别)

import math
from collections import Counter, defaultdict


TRAIN = [
    ("接口连续超时", "incident"),
    ("任务执行失败", "incident"),
    ("服务无法访问", "incident"),
    ("节点连接中断", "incident"),
    ("磁盘空间不足", "capacity"),
    ("内存使用率过高", "capacity"),
    ("消息队列积压", "capacity"),
    ("缓存容量接近上限", "capacity"),
    ("服务已经恢复", "recovery"),
    ("节点重启成功", "recovery"),
    ("批处理任务完成", "recovery"),
    ("网络状态恢复正常", "recovery"),
]

TEST = [
    ("网络请求超时", "incident"),
    ("磁盘使用率达到上限", "capacity"),
    ("定时任务恢复运行", "recovery"),
    ("接口调用失败", "incident"),
]


def features(text: str) -> list[str]:
    clean = "".join(text.lower().split())
    chars = list(clean)
    pairs = [clean[i : i + 2] for i in range(len(clean) - 1)]
    return chars + pairs


class MultinomialNB:
    def fit(self, rows: list[tuple[str, str]]) -> None:
        self.label_counts = Counter(label for _, label in rows)
        self.term_counts = defaultdict(Counter)
        self.term_totals = Counter()
        self.vocab = set()

        for text, label in rows:
            counts = Counter(features(text))
            self.term_counts[label].update(counts)
            self.term_totals[label] += sum(counts.values())
            self.vocab.update(counts)
        self.total_rows = len(rows)

    def predict_one(self, text: str) -> tuple[str, dict[str, float]]:
        counts = Counter(features(text))
        scores = {}
        vocab_size = len(self.vocab)

        for label, label_count in self.label_counts.items():
            score = math.log(label_count / self.total_rows)
            denominator = self.term_totals[label] + vocab_size
            for term, count in counts.items():
                numerator = self.term_counts[label][term] + 1
                score += count * math.log(numerator / denominator)
            scores[label] = score

        prediction = max(scores, key=scores.get)
        return prediction, scores


model = MultinomialNB()
model.fit(TRAIN)

correct = 0
for text, expected in TEST:
    predicted, scores = model.predict_one(text)
    correct += predicted == expected
    mark = "✓" if predicted == expected else "✗"
    print(f"{mark} {text:<12} expected={expected:<8} predicted={predicted}")

print(f"accuracy={correct / len(TEST):.2f}")