基于LLM问答模型的语句情感分类需求及代码实现尝试

现有实现的核心问题 #

1. 模型选型错误：你使用的是基于SQuAD数据集微调的BERT问答模型，这类模型的核心能力是从给定上下文中提取问题的答案，而非进行情感分类。它不会输出"积极/消极/中性"这类标签，反而会从输入句子里抓取相关词汇（比如"love""terrible"），完全不符合需求。
2. 输入逻辑错误：问答任务需要明确的"问题+上下文"结构，但你的prompt将问题和待分类句子混在一起，上下文仅传入原句，模型无法理解要输出指定分类标签的要求。
3. 冗余代码：导入了大量未使用的库（如Roberta相关组件、重复的AutoTokenizer），增加了不必要的资源占用。

优化方案 #

方案一：使用专门的情感分类预训练模型（推荐） #

直接选用Hugging Face生态中针对情感分类任务微调的模型，这类模型能直接输出标准化的情感标签和置信度，准确率和效率远高于用问答模型适配。

示例代码：

import pandas as pd
from transformers import pipeline

# 初始化情感分类管道，采用轻量且高效的预训练模型
sentiment_analyzer = pipeline(
    "sentiment-analysis",
    model="distilbert-base-uncased-finetuned-sst-2-english"
)

# 构建数据集（保留你的原代码）
positive_sentences = [
    "I love this product!",
    "The weather is beautiful today.",
    "The team did an excellent job.",
    "She is a very talented musician."
]

negative_sentences = [
    "I am not satisfied with the service.",
    "The food was terrible at that restaurant.",
    "The movie was a complete disappointment.",
    "He made a lot of mistakes in the project."
]

sentences = positive_sentences + negative_sentences
df = pd.DataFrame({'snippet': sentences})

# 批量处理情感分类，效率远高于逐行迭代
results = sentiment_analyzer(df['snippet'].tolist())

# 将结果合并到DataFrame中
df['sentiment'] = [res['label'].lower() for res in results]
df['confidence'] = [res['score'] for res in results]

print(df)

方案二：适配问答模型实现情感分类（不推荐） #

如果必须使用现有问答模型，需要重构输入逻辑，明确告知模型可选的情感标签范围：

示例代码：

import pandas as pd
from transformers import pipeline, BertTokenizer, BertForQuestionAnswering

# 加载问答模型
tokenizer = BertTokenizer.from_pretrained('bert-large-uncased-whole-word-masking-finetuned-squad')
model = BertForQuestionAnswering.from_pretrained('bert-large-uncased-whole-word-masking-finetuned-squad')
question_answerer = pipeline("question-answering", model=model, tokenizer=tokenizer)

# 构建数据集
positive_sentences = [
    "I love this product!",
    "The weather is beautiful today.",
    "The team did an excellent job.",
    "She is a very talented musician."
]

negative_sentences = [
    "I am not satisfied with the service.",
    "The food was terrible at that restaurant.",
    "The movie was a complete disappointment.",
    "He made a lot of mistakes in the project."
]

sentences = positive_sentences + negative_sentences
df = pd.DataFrame({'snippet': sentences})

# 重构prompt与上下文，明确给出可选标签
for index, row in df.iterrows():
    sentence = row["snippet"]
    context = f"Sentence: {sentence}. The possible sentiment labels are positive, negative, neutral."
    question = "What is the sentiment of the sentence?"
    
    result = question_answerer(question=question, context=context)
    
    if "answer" in result:
        print(f"Article {index+1} sentiment: {result['answer']}")
    else:
        print(f"Article {index+1} sentiment not found.")

额外优化建议 #

• 清理冗余代码：删除未使用的库导入，减少内存占用。
• 支持中性分类：若需要区分中性情感，可选用支持三分类的模型（如cardiffnlp/twitter-roberta-base-sentiment），该模型输出LABEL_0（消极）、LABEL_1（中性）、LABEL_2（积极），可自行映射为中文标签。
• 轻量化部署：在资源有限的环境中，优先选用DistilBERT、MobileBERT等轻量化模型，兼顾速度与准确率。
• 批量处理：避免使用iterrows()逐行处理，改用批量输入提升效率。

内容的提问来源于stack exchange，提问作者Economist_Ayahuasca