传统方法快速出 baseline

from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import train_test_split
from sklearn.metrics import classification_report

# 1. 加载数据
texts, labels = load_dataset("data/dataset_v1.csv")

# 2. 划分训练/测试集
X_train, X_test, y_train, y_test = train_test_split(
    texts, labels, test_size=0.2, random_state=42, stratify=labels
)

# 3. TF-IDF 向量化
vectorizer = TfidfVectorizer(max_features=5000)
X_train_vec = vectorizer.fit_transform(X_train)  # 在训练集上学词表
X_test_vec = vectorizer.transform(X_test)         # 用同一个词表转换测试集

# 4. 训练分类器
clf = LogisticRegression(max_iter=1000)
clf.fit(X_train_vec, y_train)

# 5. 评估
y_pred = clf.predict(X_test_vec)
print(classification_report(y_test, y_pred, target_names=["正常", "对骂"]))

# 推荐配置
vectorizer = TfidfVectorizer(
    max_features=5000,
    ngram_range=(1, 2),     # 单词 + 二元组("脑子有","有病")
    min_df=2,               # 至少出现在 2 个文档中
    sublinear_tf=True,      # TF 取对数
)

import fasttext

# 训练(一行代码)
model = fasttext.train_supervised(
    input="data/train.txt",      # FastText 格式的训练文件
    epoch=25,                     # 训练轮数
    lr=0.5,                       # 学习率
    wordNgrams=2,                 # 考虑二元组
    dim=50,                       # 词向量维度
    loss="softmax",               # 分类损失函数
)

# 预测
result = model.predict("你脑子有病吧")
# 返回: (('__label__1',), array([0.95]))  → 对骂,置信度 95%

# 评估
result = model.test("data/test.txt")
# 返回: (样本数, precision, recall)
print(f"样本数: {result[0]}, P: {result[1]:.3f}, R: {result[2]:.3f}")

import numpy as np

def extract_features(text: str, abuse_keywords: list) -> np.ndarray:
    """提取手工特征(拼接到 TF-IDF 向量后面)"""
    features = []

    # 长度特征
    features.append(len(text))
    features.append(len(text.split()))

    # 标点特征
    features.append(text.count("!") + text.count("!"))
    features.append(text.count("?") + text.count("?"))

    # 脏话词表特征
    hit_count = sum(1 for kw in abuse_keywords if kw in text.lower())
    features.append(hit_count)
    features.append(hit_count / max(len(text), 1))  # 脏话密度

    # 大写比例(检测 SB,NM 等)
    alpha_chars = [c for c in text if c.isalpha()]
    upper_ratio = sum(1 for c in alpha_chars if c.isupper()) / max(len(alpha_chars), 1)
    features.append(upper_ratio)

    return np.array(features)

# 实验设计
# ─────────────────────────────────────
# 数据集: dataset_v1.csv(200+ 条,1:1 比例)
# 划分: 80% 训练,20% 测试,stratify 分层
# 指标: F1-score(对骂类)
# 方法:
#   A. TF-IDF(unigram) + LR
#   B. TF-IDF(bigram) + LR + 手工特征
#   C. FastText

# 预期结果(量级参考):
#   A: F1 ≈ 0.75-0.85
#   B: F1 ≈ 0.80-0.90(特征工程提升 5-10%)
#   C: F1 ≈ 0.80-0.88(子词优势)

# 错误分析代码框架
errors = []
for text, true_label, pred_label in zip(X_test, y_test, y_pred):
    if true_label != pred_label:
        error_type = "漏报" if true_label == 1 else "误报"
        errors.append({"text": text, "type": error_type, "true": true_label, "pred": pred_label})

print(f"总错误: {len(errors)} 条")
print(f"  漏报: {sum(1 for e in errors if e['type'] == '漏报')} 条")
print(f"  误报: {sum(1 for e in errors if e['type'] == '误报')} 条")

# 打印错误样本
for e in errors[:10]:
    print(f"  [{e['type']}] {e['text']}")