从即将停用的LUIS迁移至Azure CLU：实现gather_number意图100%识别

在LUIS中可通过指定pattern实现意图100%置信度识别，现从即将停用的LUIS迁移至Azure CLU，但CLU不支持pattern，相关逻辑在迁移中丢失。我们需要在CLU中构建模型，实现对各类含数字语句的gather_number意图100%置信度识别，但当前测试中部分场景未达标——实体识别准确率为100%，仅意图置信度不足。

测试结果 #

当前CLU项目配置 #

{
    "projectFileVersion": "2022-10-01-preview",
    "stringIndexType": "Utf16CodeUnit",
    "metadata": {
        "projectKind": "Conversation",
        "settings": {
            "confidenceThreshold": 0
        },
        "projectName": "TestNumbers",
        "multilingual": false,
        "description": "Testing any number",
        "language": "en-us"
    },
    "assets": {
        "projectKind": "Conversation",
        "intents": [
            {
                "category": "None"
            },
            {
                "category": "gather_number"
            }
        ],
        "entities": [
            {
                "category": "mynumber",
                "compositionSetting": "combineComponents",
                "prebuilts": [
                    {
                        "category": "Quantity.Number"
                    }
                ]
            }
        ],
        "utterances": [
            {
                "text": "oh k 100",
                "language": "en-us",
                "intent": "gather_number",
                "entities": [],
                "dataset": "Train"
            },
            {
                "text": "take 100",
                "language": "en-us",
                "intent": "gather_number",
                "entities": [],
                "dataset": "Train"
            },
            {
                "text": "100 now 200 later",
                "language": "en-us",
                "intent": "gather_number",
                "entities": [],
                "dataset": "Train"
            },
            {
                "text": "only 100",
                "language": "en-us",
                "intent": "gather_number",
                "entities": [],
                "dataset": "Train"
            },
            {
                "text": "100 only",
                "language": "en-us",
                "intent": "gather_number",
                "entities": [],
                "dataset": "Train"
            },
            {
                "text": "100",
                "language": "en-us",
                "intent": "gather_number",
                "entities": [
                    {
                        "category": "mynumber",
                        "offset": 0,
                        "length": 3
                    }
                ],
                "dataset": "Train"
            },
            {
                "text": "this time 100",
                "language": "en-us",
                "intent": "gather_number",
                "entities": [],
                "dataset": "Train"
            },
            {
                "text": "no 100",
                "language": "en-us",
                "intent": "gather_number",
                "entities": [],
                "dataset": "Train"
            },
            {
                "text": "yes 100",
                "language": "en-us",
                "intent": "gather_number",
                "entities": [],
                "dataset": "Train"
            },
            {
                "text": "for now 100",
                "language": "en-us",
                "intent": "gather_number",
                "entities": [],
                "dataset": "Train"
            },
            {
                "text": "great 100",
                "language": "en-us",
                "intent": "gather_number",
                "entities": [],
                "dataset": "Train"
            },
            {
                "text": "its 100",
                "language": "en-us",
                "intent": "gather_number",
                "entities": [],
                "dataset": "Train"
            },
            {
                "text": "it is 100",
                "language": "en-us",
                "intent": "gather_number",
                "entities": [],
                "dataset": "Train"
            },
            {
                "text": "ok 100",
                "language": "en-us",
                "intent": "gather_number",
                "entities": [],
                "dataset": "Train"
            },
            {
                "text": "how about 100",
                "language": "en-us",
                "intent": "gather_number",
                "entities": [],
                "dataset": "Train"
            },
            {
                "text": "got to pay 100",
                "language": "en-us",
                "intent": "gather_number",
                "entities": [],
                "dataset": "Train"
            },
            {
                "text": "my number is 100",
                "language": "en-us",
                "intent": "gather_number",
                "entities": [],
                "dataset": "Train"
            }
        ]
    }
}

优化方案 #

1. 补充多样化训练样本 #

• 新增单数字（如"1"、"5"）、带小数的数字（如"3212.44"、"383.22"）的gather_number意图样本，覆盖测试中置信度不足的输入场景
• 补充"200 may be"这类数字出现在句中但属于None意图的样本，明确区分“含数字但不属于收集数字意图”的边界情况
• 给None意图添加更多无数字的多样化文本，比如"random words"、"no number here"，强化模型对非目标意图的判断

2. 完善实体标注 #

• 给所有gather_number意图的训练 utterance 标注mynumber实体，让模型建立“意图-实体”的强关联，而不是仅单个样本标注
• 对于包含多个数字的样本（如"100 now 200 later"），也要标注所有数字实体，提升模型对多数字场景的意图判断准确性

3. 利用CLU规则引擎替代LUIS Pattern功能 #

在CLU中通过意图规则实现类似Pattern的强匹配逻辑：

• 创建规则：当文本中识别到mynumber实体时，强制将gather_number意图的置信度设为100%
• 创建反向规则：当文本中未识别到mynumber实体时，强制将None意图的置信度设为100%

4. 后处理逻辑兜底 #

如果规则引擎配置受限，可在调用CLU模型后添加简单的后处理逻辑：

• 若返回结果中存在mynumber实体，直接将gather_number意图的置信度改为100%
• 若不存在mynumber实体，直接将None意图的置信度改为100%

内容的提问来源于stack exchange，提问作者Athadu