基于 PyTorch + BERT 意图识别与模型微调

意图识别（Intent Detection）是自然语言处理（Natural Language Processing，NLP）的一个子模块。意图识别说大白话就是：能够理解提炼用户输入的内容最终目的、意图是什么，能够从用户的自然语言中理解他到底想要做什么！所以通常我们会先进行意图分类定义，就那绩效业务来说，一般会定义：

发起考核任务 ------> create_task
创建考核方案 ------> create_plan
查看考核评分 ------> query_score
进行任务归档 ------> task_archive

当用户输入：发起张三2026年全年的绩效考核任务，就能把该内容理解为【发起考核任务】------> create_task；将2026年1月份任务就行归档，就理解为【进行任务归档】------> task_archive。

了解了意图识别概念后，接下里先了解一个概念叫做【槽位slot】，槽位是意图识别中的：参数、信息点，是实现具体操作所需的关键数据。还是用上面的案例举例，发起张三2026年全年的绩效考核任务，那么意图和槽位分别是：

意图：发起考核任务 ------> create_task
槽位：员工名 ------> 张三；时间 ------> 2026年全年；

理解了意图识别和槽位，接下来使用 Pytorch + Bert 实现一个意图识别的代码案例，首先引入pip依赖：

python 复制代码

pip install torch transformers datasets scikit-learn numpy

意图识别、模型微调代码如下：

python 复制代码

import torch
from transformers import BertTokenizer, BertForSequenceClassification, Trainer, TrainingArguments
from datasets import Dataset
import numpy as np
from sklearn.metrics import accuracy_score, precision_recall_fscore_support

# --------------------------
# 1. 数据准备
# --------------------------
texts = [
    "我想发起一个考核任务", "启动新的绩效任务", "我要创建任务", "发起考核任务", "新建任务", "我要开始任务", "创建新的考核任务", "帮我发起任务",
    "帮我创建考核方案", "新增考核方案", "我要制定考核计划", "创建考核模板", "我要新增方案", "帮我设计考核方案", "制定新的考核方案", "建立考核方案",
    "我想查看我的考核评分", "查看考核结果", "查询我的绩效", "能不能查一下我的评分", "帮我看考核成绩", "获取考核分数", "查一下我的考核评分", "查询成绩",
    "请将这个任务归档", "归档已完成的任务", "任务完成后归档", "把任务存档", "完成任务归档", "整理任务归档", "归档考核任务", "任务处理完毕归档"
]

labels = [
    "create_task","create_task","create_task","create_task","create_task","create_task","create_task","create_task",
    "create_plan","create_plan","create_plan","create_plan","create_plan","create_plan","create_plan","create_plan",
    "query_score","query_score","query_score","query_score","query_score","query_score","query_score","query_score",
    "task_archive","task_archive","task_archive","task_archive","task_archive","task_archive","task_archive","task_archive"
]

label2id = {label: idx for idx, label in enumerate(sorted(set(labels)))}
id2label = {idx: label for label, idx in label2id.items()}
numeric_labels = [label2id[label] for label in labels]

dataset = Dataset.from_dict({
    "text": texts,
    "label": numeric_labels
})

# 划分训练/验证集
train_test = dataset.train_test_split(test_size=0.2, shuffle=True, seed=42)
train_dataset = train_test["train"]
eval_dataset = train_test["test"]

# --------------------------
# 2. 分词
# --------------------------
tokenizer = BertTokenizer.from_pretrained("bert-base-chinese")

def tokenize(examples):
    return tokenizer(examples["text"], padding="max_length", truncation=True, max_length=32)

train_dataset = train_dataset.map(tokenize, batched=True)
eval_dataset = eval_dataset.map(tokenize, batched=True)

train_dataset.set_format(type="torch", columns=["input_ids","attention_mask","label"])
eval_dataset.set_format(type="torch", columns=["input_ids","attention_mask","label"])

# --------------------------
# 3. 加载 Bert 模型
# --------------------------
model = BertForSequenceClassification.from_pretrained(
    "bert-base-chinese",
    num_labels=len(label2id),
    id2label=id2label,
    label2id=label2id
)

# --------------------------
# 4. 训练参数
# --------------------------
training_args = TrainingArguments(
    output_dir="./bert_intent_model",
    evaluation_strategy="epoch",
    save_strategy="epoch",
    learning_rate=2e-5,
    per_device_train_batch_size=8,
    per_device_eval_batch_size=8,
    num_train_epochs=15,
    weight_decay=0.01,
    logging_dir="./logs",
    logging_steps=10,
    load_best_model_at_end=True,
    metric_for_best_model="accuracy",
    no_cuda=True
)

# --------------------------
# 5. 指标函数
# --------------------------
def compute_metrics(eval_pred):
    logits, labels = eval_pred
    preds = np.argmax(logits, axis=-1)
    precision, recall, f1, _ = precision_recall_fscore_support(labels, preds, average="weighted")
    acc = accuracy_score(labels, preds)
    return {"accuracy": acc, "precision": precision, "recall": recall, "f1": f1}

# --------------------------
# 6. Trainer
# --------------------------
trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=train_dataset,
    eval_dataset=eval_dataset,
    compute_metrics=compute_metrics
)

# --------------------------
# 7. 训练
# --------------------------
trainer.train()
trainer.save_model("./bert_intent_model")

# --------------------------
# 8. 推理函数
# --------------------------
def predict_intent(text):
    inputs = tokenizer(text, return_tensors="pt", padding=True, truncation=True, max_length=32)
    with torch.no_grad():
        outputs = model(**inputs)
        pred = torch.argmax(outputs.logits, dim=1).item()
    return id2label[pred]

# --------------------------
# 9. 测试 demo
# --------------------------
test_texts = [
    "我要发起一个新的考核任务",
    "帮我创建一个考核方案",
    "我想查看我的考核评分",
    "请把这个任务归档",
    "启动新的绩效任务",
    "新增考核方案",
    "查看考核结果",
    "归档已完成的任务",
    "我需要创建任务",
    "帮我整理考核方案",
    "能不能查一下我的评分",
    "任务完成后归档"
]

for text in test_texts:
    intent = predict_intent(text)
    print(f"输入: {text}  --->  预测意图: {intent}")

最终测试结果：