如何优化FuzzyWuzzy匹配逻辑，精准识别用户支持请求的关联应用？

问题背景 #

公司用户每日在工作频道提交多份支持请求，涉及Okta、Slack、Cloudflare等应用，希望通过Bot自动推送对应归档的解决方案。当前采用Python的FuzzyWuzzy库，用fuzz.partial_token_set_ratio计算用户消息与预设请求类型（如OKTA_ACCESS、OKTA_UNLOCK、CLOUDFLARE）的匹配分数，设定分数>90时判定匹配。

但现有逻辑存在误匹配问题：当用户发送如下消息时，Cloudflare和Okta的匹配分数均>90，但实际用户问题仅与Cloudflare相关：

Hello team, it looks like I am having an issue with Cloudflare. I have been added to the following Okta group.
Can someone look into this for me, please?

现有代码如下：

from enum import Enum
from fuzzywuzzy import fuzz

class RequestType(Enum):

    OKTA_ACCESS = "okta access"
    OKTA_UNLOCK = "okta unlock"
    CLOUDFLARE = "cannot connect to cloudflare"

class RequestsMatcher:

    def __init__(self, message: str) -> None:
        
        scores = [
            (
                type,
                fuzz.partial_token_set_ratio(message, type.value)
            )
            for type in RequestType
        ]

        mvs = [score for score in scores if score[1] > 75]

        hits = [score for score in scores if score[1] > 90]

        print(scores)
        self.scores = scores
        self.highest = mvs
        self.hit = next(iter(hits), None)

优化方案 #

1. 聚焦问题分句匹配 #

用户核心诉求通常和"issue"、"problem"、"cannot"这类问题描述词绑定，先提取包含这类词汇的分句，仅针对分句做匹配，避免无关内容干扰。

修改后示例代码：

import re
from enum import Enum
from fuzzywuzzy import fuzz

class RequestType(Enum):
    OKTA_ACCESS = "okta access"
    OKTA_UNLOCK = "okta unlock"
    CLOUDFLARE = "cloudflare"  # 简化匹配文本，聚焦应用核心标识

class RequestsMatcher:
    def __init__(self, message: str) -> None:
        # 提取包含问题关键词的分句
        issue_keywords = r"(?i).*(issue|problem|cannot|trouble|unable|issue with).*"
        issue_sentences = [sent.strip() for sent in message.split('.') if re.match(issue_keywords, sent)]
        
        # 优先用问题分句匹配，无匹配时 fallback 到全消息
        target_text = ' '.join(issue_sentences).lower() if issue_sentences else message.lower()
        
        scores = [
            (req_type, fuzz.partial_token_set_ratio(target_text, req_type.value.lower()))
            for req_type in RequestType
        ]
        
        # 按分数排序，取最高分判断是否达标
        sorted_scores = sorted(scores, key=lambda x: x[1], reverse=True)
        top_score = sorted_scores[0][1]
        self.hit = sorted_scores[0] if top_score > 90 else None
        
        print(scores)
        self.scores = scores

2. 简化匹配模板+分层阈值 #

把RequestType的value简化为"应用名+核心动作"的极简形式，比如CLOUDFLARE = "cloudflare"、OKTA_UNLOCK = "okta unlock"，减少冗余文本带来的误匹配。同时给不同类型设置分层阈值：

• 仅应用名的匹配（如Cloudflare）：阈值设为90
• 带动作的匹配（如okta unlock）：阈值设为95，避免单纯提到应用名就触发匹配

3. 关键词位置加权 #

计算分数时，给靠近问题词汇的应用名额外加权。比如检测到应用名出现在"issue"、"problem"前后3个词范围内，分数加15分；若出现在无关陈述中（如示例里的"I have been added to the following Okta group"），分数减10分。

4. 切换至意图分类方案 #

如果模糊匹配精度始终达不到要求，可基于历史支持请求数据，用spaCy做实体识别+意图分类，或用scikit-learn训练朴素贝叶斯/逻辑回归分类模型，通过标注数据让模型学习用户诉求与应用的关联，能大幅提升匹配准确率。

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