470_情感分析方法¶

# ---- 方案1：使用 transformers pipeline（最快验证）----
from transformers import pipeline

# 加载情感分析流水线（内置fine-tuned模型）
sentiment_pipe = pipeline(
    "sentiment-analysis",
    model="lxyuan/distilbert-base-multilingual-cased-sentiments-student",
    # 支持中文的多语言情感模型
)

texts = [
    "这款产品真的太好用了，强烈推荐！",
    "服务态度差，等了半小时没人理",
    "还行吧，一般般"
]

for text in texts:
    result = sentiment_pipe(text)[0]  # {'label': 'positive', 'score': 0.98}
    print(f"{text[:15]}... → {result['label']} ({result['score']:.2f})")

# ---- 方案2：BERT微调（自定义数据集）----
from transformers import BertTokenizer, BertForSequenceClassification
import torch
import torch.nn.functional as F

tokenizer = BertTokenizer.from_pretrained('bert-base-chinese')
model = BertForSequenceClassification.from_pretrained(
    'bert-base-chinese',
    num_labels=3  # 0=负面, 1=中性, 2=正面
)

def predict_sentiment(text, model, tokenizer):
    """预测单条文本情感"""
    model.eval()
    enc = tokenizer(text, return_tensors='pt',
                    max_length=128, truncation=True, padding='max_length')
    with torch.no_grad():
        logits = model(**enc).logits          # (1, 3)
        probs = F.softmax(logits, dim=-1)     # 转成概率
        pred = probs.argmax(dim=-1).item()
    labels = {0: '负面', 1: '中性', 2: '正面'}
    return labels[pred], probs[0][pred].item()

# ---- 方案3：方面级情感分析（ABSA）----
# 把方面词拼接到句子中，让模型判断该方面的情感
def absa_predict(sentence, aspect, model, tokenizer):
    """
    方面级情感：将 sentence 和 aspect 拼接，
    格式：[CLS] 方面词 [SEP] 句子 [SEP]
    """
    # 以方面词作为"问题"，原句作为"段落"，类似QA格式
    enc = tokenizer(
        aspect, sentence,  # (question, context) 格式
        return_tensors='pt',
        max_length=128,
        truncation=True,
        padding='max_length'
    )
    with torch.no_grad():
        logits = model(**enc).logits
        pred = logits.argmax(dim=-1).item()
    return {0: '负面', 1: '中性', 2: '正面'}[pred]

# ---- 方案4：中文词典方法（快速基线）----
# pip install snownlp
from snownlp import SnowNLP

text = "这本书写得非常好，推荐给大家"
s = SnowNLP(text)
print(f"SnowNLP情感分数: {s.sentiments:.3f}")  # 0~1，越大越正面