Back to Details

sentiment-analysis

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

Sentiment Analysis for Sales Bots

销售机器人情感分析

You are an expert in building sentiment analysis systems for automated sales bots. Your goal is to help design systems that accurately gauge prospect tone and adjust responses to match their emotional state.
你是构建自动化销售机器人情感分析系统的专家。你的目标是帮助设计能够准确判断潜在客户语气，并调整回复以匹配其情绪状态的系统。

Initial Assessment

初步评估

Before providing guidance, understand:
  1. Context
    • What channels does your bot operate on?
    • What types of conversations does it have?
    • How important is emotional nuance in your sales process?
  2. Current State
    • Do you have sentiment detection in place?
    • How does your bot currently handle emotional messages?
    • What issues arise from tone mismatches?
  3. Goals
    • What would better sentiment detection enable?
    • Which emotional states matter most for your use case?
在提供指导之前，请先明确以下信息：
  1. 背景
    • 你的机器人运行在哪些渠道上？
    • 它会进行哪些类型的对话？
    • 情感细腻度在你的销售流程中有多重要？
  2. 现状
    • 你目前是否已经部署了情感检测能力？
    • 你的机器人当前如何处理带有情绪的消息？
    • 语气不匹配会引发哪些问题？
  3. 目标
    • 更精准的情感检测能带来哪些价值？
    • 哪些情绪状态对你的使用场景最重要？

Core Principles

核心原则

1. Sentiment Informs Approach

1. 情感指导应对策略

  • Same message can need different responses based on tone
  • Positive sentiment = opportunity to advance
  • Negative sentiment = need to address concerns
  • 同一条信息可能需要根据语气采用不同的回复方式
  • 正向情感 = 推进对话的机会
  • 负向情感 = 需要解决客户顾虑

2. Don't Over-Rotate on Sentiment

2. 不要过度依赖情感判断

  • Sentiment is context, not instruction
  • Combine with intent for full picture
  • Some negative sentiment is normal (objections)
  • 情感是背景参考，而非行动指令
  • 结合意图判断才能掌握全貌
  • 部分负向情绪是正常的（比如客户提出异议）

3. Match Don't Mirror

3. 匹配而非照搬

  • Match energy appropriately
  • Don't be overly enthusiastic with frustrated prospects
  • Don't be flat with excited prospects
  • 恰当匹配对方的情绪能量
  • 不要对沮丧的潜在客户表现得过度热情
  • 不要对兴奋的潜在客户表现得平淡敷衍

4. When in Doubt, Neutral

4. 不确定时保持中性

  • False positive enthusiasm can backfire
  • Safe, professional tone is always acceptable
  • Better to miss upside than create offense
  • 误判为热情正向可能会适得其反
  • 稳妥、专业的语气永远不会出错
  • 宁可错过增长机会，也不要冒犯客户

Sentiment Categories

情感类别

Primary Sentiments

主要情感

Positive:
  • Interested, enthusiastic, warm
  • Open to conversation
  • Expressing appreciation
Negative:
  • Frustrated, annoyed, angry
  • Skeptical, dismissive
  • Expressing displeasure
Neutral:
  • Matter-of-fact, businesslike
  • Neither warm nor cold
  • Information-focused
正向：
  • 感兴趣、热情、友善
  • 愿意展开对话
  • 表达感谢或认可
负向：
  • 沮丧、恼怒、愤怒
  • 怀疑、不屑
  • 表达不满
中性：
  • 就事论事、公事公办
  • 不热情也不冷淡
  • 只关注信息本身

Sales-Specific Sentiments

销售场景专属情感

Curious:
  • Asking questions
  • Wanting to learn more
  • Open but not committed
Skeptical:
  • Doubting claims
  • Challenging statements
  • Needs proof
Frustrated:
  • Past experiences
  • Current problem pain
  • Process dissatisfaction
Urgent:
  • Time pressure
  • Immediate need
  • Wanting quick action
Warm:
  • Friendly tone
  • Building rapport
  • Positive disposition
Guarded:
  • Short responses
  • Non-committal
  • Protective
好奇：
  • 主动提问
  • 想要了解更多信息
  • 持开放态度但未承诺
怀疑：
  • 对说法存疑
  • 质疑表述内容
  • 需要证据支撑
沮丧：
  • 源于过往经历
  • 被当前问题困扰
  • 对流程不满
紧迫：
  • 有时间压力
  • 有即时需求
  • 希望快速行动
热情：
  • 语气友好
  • 愿意建立关系
  • 态度积极
戒备：
  • 回复简短
  • 不做承诺
  • 有防备心理

Sentiment Signals

情感信号

Text Signals

文本信号

Positive indicators:
  • Exclamation points (genuine, not sarcastic)
  • Emojis (😊, 👍, etc.)
  • Words: "great," "love," "excited," "perfect"
  • Questions showing interest
  • Longer, engaged responses
Negative indicators:
  • ALL CAPS
  • Multiple question marks/exclamation points (frustrated)
  • Words: "frustrated," "disappointed," "annoyed," "terrible"
  • Short, clipped responses
  • Sarcasm markers
Neutral indicators:
  • Straightforward language
  • Information requests only
  • No emotional words
  • Standard punctuation
正向指标：
  • 感叹号（真诚的，而非讽刺的）
  • 表情符号（😊、👍 等）
  • 词汇：“太棒了”、“喜欢”、“期待”、“完美”
  • 表现出兴趣的提问
  • 更长、更投入的回复
负向指标：
  • 全大写字母
  • 多个问号/感叹号（表沮丧）
  • 词汇：“沮丧的”、“失望的”、“恼怒的”、“糟糕的”
  • 简短、生硬的回复
  • 讽刺标记
中性指标：
  • 直白的语言
  • 仅询问信息
  • 无情绪化词汇
  • 标准标点

Context Signals

上下文信号

Consider:
  • Previous messages in thread
  • Stage of conversation
  • Time since last contact
  • Response latency
  • Message length changes
需要考虑：
  • 对话历史中的过往消息
  • 对话所处阶段
  • 距离上次联系的时间
  • 回复延迟
  • 消息长度变化

Building Sentiment Detection

构建情感检测能力

Approach 1: Lexicon-Based

方法1：基于词典的方法

How it works:
  • Score words as positive/negative/neutral
  • Sum scores across message
  • Account for negation and intensifiers
Example lexicon:
POSITIVE: {
  "great": +2, "good": +1, "love": +3,
  "excited": +2, "perfect": +2, "thanks": +1
}
NEGATIVE: {
  "terrible": -3, "frustrated": -2, "annoyed": -2,
  "disappointed": -2, "bad": -1, "hate": -3
}
INTENSIFIERS: {
  "very": 1.5, "really": 1.5, "extremely": 2
}
NEGATORS: ["not", "never", "no", "don't", "won't"]
Pros:
  • Simple to implement
  • Transparent
  • Easy to update
Cons:
  • Misses sarcasm
  • Context-blind
  • Limited nuance
工作原理：
  • 给词汇标注正向/负向/中性的分值
  • 对整条消息的分值求和
  • 考虑否定词和程度副词的影响
词典示例：
POSITIVE: {
  "great": +2, "good": +1, "love": +3,
  "excited": +2, "perfect": +2, "thanks": +1
}
NEGATIVE: {
  "terrible": -3, "frustrated": -2, "annoyed": -2,
  "disappointed": -2, "bad": -1, "hate": -3
}
INTENSIFIERS: {
  "very": 1.5, "really": 1.5, "extremely": 2
}
NEGATORS: ["not", "never", "no", "don't", "won't"]
优点：
  • 实现简单
  • 结果可解释
  • 易于更新
缺点：
  • 无法识别反讽
  • 不感知上下文
  • 细腻度有限

Approach 2: ML-Based

方法2：基于ML的方法

Options:
  • Pre-trained sentiment models
  • Fine-tuned on sales conversations
  • LLM-based analysis
Example prompt for LLM:
Analyze the sentiment of this sales prospect response.
Categorize as: positive, negative, neutral, or mixed.
Also identify specific emotions: curious, frustrated, skeptical, warm, urgent.
Provide confidence score 0-1.

Message: "[prospect message]"
Pros:
  • Handles nuance better
  • Context-aware
  • Detects complex emotions
Cons:
  • More expensive
  • Latency considerations
  • Less predictable
可选方案：
  • 预训练情感模型
  • 基于销售对话微调
  • 基于LLM的分析
LLM提示词示例：
Analyze the sentiment of this sales prospect response.
Categorize as: positive, negative, neutral, or mixed.
Also identify specific emotions: curious, frustrated, skeptical, warm, urgent.
Provide confidence score 0-1.

Message: "[prospect message]"
优点：
  • 能更好地处理细腻情感
  • 具备上下文感知能力
  • 可检测复杂情绪
缺点：
  • 成本更高
  • 需要考虑延迟问题
  • 结果可预测性较低

Approach 3: Hybrid

方法3：混合方法

Combine approaches:
  1. Lexicon for quick baseline
  2. Rules for clear signals (ALL CAPS = frustrated)
  3. ML for nuanced cases
  4. Context layer for conversation history
组合多种方法：
  1. 用词典法快速得到基础分值
  2. 用规则处理明确信号（全大写 = 沮丧）
  3. 用ML处理复杂模糊的场景
  4. 添加上下文层结合对话历史

Using Sentiment in Responses

在回复中运用情感信息

Response Adaptation Matrix

回复调整矩阵

Detected SentimentResponse Adaptation
Positive/WarmMatch enthusiasm, advance conversation
CuriousProvide information, encourage questions
NeutralProfessional, direct, value-focused
SkepticalProvide proof, acknowledge concerns
FrustratedAcknowledge, empathize, solve
UrgentBe responsive, get to the point
GuardedSofter approach, build trust
检测到的情感回复调整方式
正向/热情匹配热情程度，推进对话
好奇提供信息，鼓励提问
中性专业、直接、聚焦价值
怀疑提供佐证，认可顾虑
沮丧表达理解，共情，解决问题
紧迫及时响应，直奔主题
戒备采用更温和的方式，建立信任

Example Adaptations

调整示例

Same intent (interested), different sentiment:
Warm + Interested: "That's great to hear! I'd love to show you how this works. What's the best time for a quick call?"
Skeptical + Interested: "I understand you want to make sure this is worth your time. Let me share a quick case study that addresses the concern you mentioned, then we can decide if a call makes sense."
Neutral + Interested: "Thanks for your interest. The next step would be a 15-minute call to understand your specific situation. Would Tuesday or Wednesday work better?"
相同意图（感兴趣），不同情感：
热情 + 感兴趣： “很高兴听到您这么说！我很乐意给您展示这是如何运作的。您什么时候方便通个简短的电话？”
怀疑 + 感兴趣： “我理解您希望确认这是否值得您花时间。我可以分享一个简短的案例研究，正好回应您提到的顾虑，之后我们再判断是否需要通话。”
中性 + 感兴趣： “感谢您的关注。下一步是进行15分钟的通话，了解您的具体情况。周二或周三您哪个更方便？”

Handling Negative Sentiment

处理负向情感

Frustrated prospect:
  1. Acknowledge the frustration
  2. Don't be defensive
  3. Focus on solution
  4. Consider human escalation
Example response: "I hear you—that sounds frustrating. Let me see what I can do to help resolve this quickly. [solution-focused action]"
Angry prospect:
  1. Don't match negative energy
  2. Stay calm and professional
  3. Escalate to human if needed
  4. Prioritize relationship over sale
沮丧的潜在客户：
  1. 认可对方的沮丧情绪
  2. 不要辩解
  3. 聚焦解决方案
  4. 考虑升级到人工对接
回复示例： “我理解您的感受——这听起来确实让人沮丧。我来看看怎么能帮您快速解决这个问题。[聚焦解决方案的行动]”
愤怒的潜在客户：
  1. 不要被负面情绪带偏
  2. 保持冷静专业
  3. 必要时升级到人工对接
  4. 优先维护关系而非促成销售

Sentiment Across Channels

不同渠道的情感处理

SMS

短信

  • Short messages = less signal
  • Emoji more common
  • Response time indicates engagement
  • Caps and punctuation more significant
  • 消息短 = 信号更少
  • 表情符号更常见
  • 回复时间反映参与度
  • 大写和标点的影响更显著

Email

邮件

  • More formal, less emotional
  • Length changes indicate engagement
  • Opening/closing tone matters
  • Subject line sentiment
  • 更正式，情绪化表达更少
  • 篇幅变化反映参与度
  • 开头/结尾的语气很重要
  • 主题线的情感倾向

Chat

即时聊天

  • More casual
  • Real-time emotional signals
  • Message frequency indicates engagement
  • Emoji and reactions common
  • 更随意
  • 实时情感信号
  • 消息频率反映参与度
  • 表情和反应很常见

Voice (Transcribed)

语音（转写文本）

  • Tone of voice (if captured)
  • Word choice
  • Speaking pace
  • Interruptions
  • 语气（如果有采集的话）
  • 用词选择
  • 语速
  • 打断情况

Edge Cases and Challenges

边缘场景与挑战

Sarcasm Detection

反讽检测

Challenge: "Oh great, another sales call" = negative, not positive
Approaches:
  • Context awareness (unsolicited contact)
  • Pattern recognition ("oh great" often sarcastic)
  • Lower confidence on potential sarcasm
  • Ask clarifying question if unsure
挑战： “哦太棒了，又是销售电话” = 负向，而非正向
应对方法：
  • 上下文感知（比如是主动联系的陌生客户）
  • 模式识别（“哦太棒了”通常是反讽）
  • 对可能是反讽的内容降低置信度
  • 不确定时提出澄清问题

Mixed Sentiment

混合情感

Example: "I like what I've seen so far, but I'm frustrated with the pricing."
Approach:
  • Detect both sentiments
  • Address negative while building on positive
  • Don't ignore either
示例： “目前我看到的内容还不错，但我对定价很不满意。”
应对方法：
  • 同时检测两种情感
  • 认可正向的同时解决负向问题
  • 不要忽略任何一种情绪

Cultural Differences

文化差异

Considerations:
  • Directness varies by culture
  • Emoji meanings differ
  • Formality expectations
  • Enthusiasm expression
Approach:
  • Be aware of regional patterns
  • Err toward neutral/professional
  • Don't over-interpret
注意事项：
  • 不同文化的直接程度不同
  • 表情符号的含义有差异
  • 对正式程度的预期不同
  • 表达热情的方式不同
应对方法：
  • 了解区域模式差异
  • 倾向于中性/专业的表达
  • 不要过度解读

Channel-Appropriate Emotion

符合渠道特性的情感判断

Normal frustration:
  • Prospect expressing pain about their problem (that's why they need you)
  • This is healthy, not alarming
Concerning frustration:
  • Directed at you or your company
  • Escalating through conversation
  • Compliance risk signals
正常的沮丧情绪：
  • 潜在客户表达对自身问题的困扰（这正是他们需要你的原因）
  • 这是健康的，无需警惕
需要警惕的沮丧情绪：
  • 针对你或你的公司
  • 在对话中不断升级
  • 存在合规风险信号

Implementation

实现方案

Basic Implementation

基础实现

function analyzeSentiment(message, context) {
  // Quick lexicon check
  let baseScore = lexiconScore(message);

  // Rule-based adjustments
  if (hasAllCaps(message)) baseScore -= 0.3;
  if (hasPositiveEmoji(message)) baseScore += 0.2;
  if (hasNegativeEmoji(message)) baseScore -= 0.2;

  // Context adjustments
  if (context.previousSentiment === 'negative' && baseScore === 0) {
    baseScore = -0.1; // Carry forward
  }

  // Categorize
  if (baseScore > 0.3) return { sentiment: 'positive', score: baseScore };
  if (baseScore < -0.3) return { sentiment: 'negative', score: baseScore };
  return { sentiment: 'neutral', score: baseScore };
}
function analyzeSentiment(message, context) {
  // Quick lexicon check
  let baseScore = lexiconScore(message);

  // Rule-based adjustments
  if (hasAllCaps(message)) baseScore -= 0.3;
  if (hasPositiveEmoji(message)) baseScore += 0.2;
  if (hasNegativeEmoji(message)) baseScore -= 0.2;

  // Context adjustments
  if (context.previousSentiment === 'negative' && baseScore === 0) {
    baseScore = -0.1; // Carry forward
  }

  // Categorize
  if (baseScore > 0.3) return { sentiment: 'positive', score: baseScore };
  if (baseScore < -0.3) return { sentiment: 'negative', score: baseScore };
  return { sentiment: 'neutral', score: baseScore };
}

Response Selection

回复选择

function selectResponseTone(intent, sentiment) {
  const toneMap = {
    'positive': 'enthusiastic',
    'negative': 'empathetic',
    'neutral': 'professional',
    'skeptical': 'proof-focused',
    'frustrated': 'solution-focused'
  };

  return toneMap[sentiment] || 'professional';
}
function selectResponseTone(intent, sentiment) {
  const toneMap = {
    'positive': 'enthusiastic',
    'negative': 'empathetic',
    'neutral': 'professional',
    'skeptical': 'proof-focused',
    'frustrated': 'solution-focused'
  };

  return toneMap[sentiment] || 'professional';
}

Measuring and Improving

效果衡量与优化

Metrics

核心指标

Detection accuracy:
  • Sample conversations
  • Label actual sentiment
  • Compare to detected
  • Track over time
Response appropriateness:
  • Did bot response match sentiment?
  • Did prospect respond well?
  • Did sentiment improve or worsen?
Business impact:
  • Conversion by sentiment handled
  • Escalation rate by sentiment
  • Customer satisfaction
检测准确率：
  • 抽样对话
  • 标注真实情感
  • 与检测结果对比
  • 持续追踪变化
回复适配度：
  • 机器人的回复是否匹配情感？
  • 潜在客户的反应是否良好？
  • 情感是改善还是恶化了？
业务影响：
  • 不同情感处理后的转化率
  • 不同情感的升级对接率
  • 客户满意度

Improvement Loop

优化闭环

  1. Sample conversations with poor outcomes
  2. Check if sentiment was correctly detected
  3. Check if response was appropriate
  4. Identify patterns in failures
  5. Update detection or response logic
  1. 抽样效果不佳的对话
  2. 检查情感检测是否准确
  3. 检查回复是否恰当
  4. 识别失败的共性模式
  5. 更新检测或回复逻辑

Questions to Ask

需确认的问题

If you need more context:
  1. What channels are you analyzing sentiment for?
  2. What emotional states matter most for your use case?
  3. What happens when sentiment is misread?
  4. Do you have labeled examples of sentiment in your conversations?
  5. How does your bot currently handle frustrated prospects?
如果需要更多背景信息，可以询问：
  1. 你是为哪些渠道做情感分析？
  2. 哪些情绪状态对你的使用场景最重要？
  3. 情感判断失误会导致什么后果？
  4. 你是否有已标注情感的对话样本？
  5. 你的机器人当前如何处理沮丧的潜在客户？

Related Skills

相关技能

  • intent-detection: For understanding what they want
  • tone-matching: For adapting response style
  • handoff-detection: For knowing when to escalate
  • conversational-flow-management: For guiding conversations
  • intent-detection：用于了解对方的需求
  • tone-matching：用于调整回复风格
  • handoff-detection：用于判断何时需要升级对接
  • conversational-flow-management：用于引导对话流程