意图识别:管理指令解析

INTENT_LABELS = ["mute", "kick", "warn", "unmute", "query_log", "query_rule", "chat", "abuse"]
label2id = {label: i for i, label in enumerate(INTENT_LABELS)}
id2label = {i: label for i, label in enumerate(INTENT_LABELS)}

model = AutoModelForSequenceClassification.from_pretrained(
    "hfl/chinese-roberta-wwm-ext",
    num_labels=len(INTENT_LABELS),
    label2id=label2id,
    id2label=id2label,
)

def preprocess(examples):
    tokens = tokenizer(examples["text"], padding="max_length", truncation=True, max_length=64)
    tokens["labels"] = [label2id[intent] for intent in examples["intent"]]
    return tokens

# 其余训练代码和第 12 篇完全一致,只需修改:
# - num_labels = 8(意图数)
# - max_length = 64(管理指令更短)
# - num_train_epochs = 5(意图类别多,多训几轮)

import re

class IntentRouter:
    def __init__(self, bert_classifier):
        self.bert = bert_classifier
        self.command_pattern = re.compile(
            r"^/(mute|kick|warn|unmute)\s+@(\S+)(?:\s+(\d+[mhd]))?"
        )

    def parse(self, text, sender_role="member"):
        if sender_role != "admin":
            return {"intent": "not_admin", "confidence": 1.0}

        cmd_match = self.command_pattern.match(text)
        if cmd_match:
            return {
                "intent": cmd_match.group(1),
                "target": cmd_match.group(2),
                "duration": cmd_match.group(3),
                "confidence": 1.0,
                "source": "rule",
            }

        result = self.bert.predict(text)

        if result["confidence"] < 0.7:
            return {"intent": "uncertain", "confidence": result["confidence"]}

        return {**result, "source": "model"}

import re

def extract_slots(text, intent):
    """从文本中提取槽位(规则方式)"""
    slots = {}

    at_match = re.search(r"@(\S+)", text)
    if at_match:
        slots["target"] = at_match.group(1)

    duration_match = re.search(r"(\d+)\s*(分钟|小时|天|min|h|d|m)", text)
    if duration_match:
        num = int(duration_match.group(1))
        unit = duration_match.group(2)
        unit_map = {"分钟": "m", "min": "m", "m": "m",
                     "小时": "h", "h": "h",
                     "天": "d", "d": "d"}
        slots["duration"] = f"{num}{unit_map.get(unit, 'm')}"

    return slots

# 示例
extract_slots("把@张三禁言30分钟", "mute")
# → {"target": "张三", "duration": "30m"}

from sklearn.metrics import confusion_matrix, ConfusionMatrixDisplay
import matplotlib.pyplot as plt

cm = confusion_matrix(true_labels, pred_labels)
disp = ConfusionMatrixDisplay(cm, display_labels=INTENT_LABELS)
disp.plot(xtick_rotation=45)
plt.tight_layout()
plt.savefig("intent_confusion_matrix.png")