当前位置：首页 > news >正文

基于Transformer的心理健康对话系统：从零构建AI心理咨询助手

news 2025/7/21 14:25:58

基于Transformer的心理健康对话系统：从零构建AI心理咨询助手

本文将详细介绍如何构建一个专业的心理健康支持系统，使用Transformer模型提供初步心理咨询和情绪支持，结合FastAPI后端和Streamlit前端实现完整解决方案。

引言：AI在心理健康领域的应用价值

随着心理健康问题日益普遍，AI心理咨询助手展现出巨大潜力：

24/7可访问性：随时提供支持
无偏见环境：用户可自由表达
早期干预：识别危机情况
资源引导：连接专业帮助
辅助治疗：补充传统心理咨询

本文将构建一个具备专业心理健康支持能力的AI系统，包含情感分析、危机检测和专业资源推荐功能。

系统架构设计

用户界面(Streamlit) ↔ REST API(FastAPI) ↔ 心理健康专用模型 ↔ 专业资源数据库

技术栈增强：

核心模型：mentalhealthai/counseling-gpt（心理咨询微调模型）
情感分析：bhadresh-savani/distilbert-base-uncased-emotion
危机检测：自定义关键词+情感分析组合
资源数据库：SQLite存储专业资源信息
安全机制：内容过滤和危机响应协议

环境准备与安装

# 创建虚拟环境
python -m venv mentalhealth-env
source mentalhealth-env/bin/activate# 安装核心依赖
pip install transformers torch
pip install fastapi "uvicorn[standard]" sqlalchemy
pip install streamlit streamlit-chat textacy
pip install pandas numpy scikit-learn

实现步骤详解

1. 心理健康专用模型与工具

创建 mentalhealth_models.py：

from transformers import pipeline, AutoTokenizer, AutoModelForCausalLM
import reclass MentalHealthAssistant:def __init__(self):# 心理咨询专用模型self.counseling_model = AutoModelForCausalLM.from_pretrained("mentalhealthai/counseling-gpt", device_map="auto")self.counseling_tokenizer = AutoTokenizer.from_pretrained("mentalhealthai/counseling-gpt")# 情感分析模型self.emotion_analyzer = pipeline("text-classification", model="bhadresh-savani/distilbert-base-uncased-emotion")# 危机关键词库self.crisis_keywords = ['自杀', '想死', '结束生命', '自残', '自杀念头','suicide', 'kill myself', 'end my life', 'self-harm','abuse', '虐待', '暴力', '不想活了']def generate_response(self, input_text, history=None, max_length=200):# 组合对话历史prompt = "作为心理咨询师，我理解你正在经历困难。让我们谈谈你的感受。\n"if history:for exchange in history[-3:]:  # 使用最近3轮对话prompt += f"用户: {exchange['user']}\n"prompt += f"咨询师: {exchange['bot']}\n"prompt += f"用户: {input_text}\n咨询师:"# 生成响应inputs = self.counseling_tokenizer.encode(prompt, return_tensors="pt", max_length=512, truncation=True).to(self.counseling_model.device)outputs = self.counseling_model.generate(inputs,max_length=max_length,temperature=0.85,top_p=0.92,repetition_penalty=1.2,num_return_sequences=1,pad_token_id=self.counseling_tokenizer.eos_token_id)response = self.counseling_tokenizer.decode(outputs[0], skip_special_tokens=True)# 提取咨询师回复部分return response.split("咨询师:")[-1].strip()def analyze_emotion(self, text):"""分析文本情感"""result = self.emotion_analyzer(text)return {"emotion": result[0]['label'],"confidence": result[0]['score']}def detect_crisis(self, text):"""检测危机信号"""text_lower = text.lower()# 关键词检测keyword_detected = any(keyword in text_lower for keyword in self.crisis_keywords)# 情感分析辅助emotion_result = self.analyze_emotion(text)intense_emotion = emotion_result['confidence'] > 0.9 and emotion_result['emotion'] in ['sadness', 'anger', 'fear']return keyword_detected or intense_emotion

2. 专业资源数据库

创建 resources_database.py：

from sqlalchemy import create_engine, Column, Integer, String, Text
from sqlalchemy.ext.declarative import declarative_base
from sqlalchemy.orm import sessionmakerBase = declarative_base()class Resource(Base):__tablename__ = 'mental_health_resources'id = Column(Integer, primary_key=True)category = Column(String(50), nullable=False)name = Column(String(100), nullable=False)description = Column(Text)contact = Column(String(200))website = Column(String(200))language = Column(String(20), default='zh')emergency = Column(Integer, default=0)  # 1=紧急服务def init_db():engine = create_engine('sqlite:///mentalhealth_resources.db')Base.metadata.create_all(engine)return enginedef seed_database():engine = init_db()Session = sessionmaker(bind=engine)session = Session()# 添加示例资源（实际应用需填充真实数据）resources = [Resource(category="危机干预",name="全国心理援助热线",description="24小时心理危机干预热线",contact="010-82951332",website="http://www.bjxl.org.cn",emergency=1),Resource(category="抑郁症",name="阳光心理健康中心",description="专业抑郁症治疗与支持",contact="400-161-9995",website="https://www.sunshinecenter.org"),Resource(category="焦虑症",name="安心心理诊所",description="焦虑症认知行为疗法专家",contact="021-64688888",website="https://www.anxin-clinic.com"),# 可添加更多资源...]session.add_all(resources)session.commit()print("数据库已初始化并填充示例资源")if __name__ == "__main__":seed_database()

3. 增强型FastAPI后端服务

创建 mentalhealth_api.py：

from fastapi import FastAPI, HTTPException
from pydantic import BaseModel
from mentalhealth_models import MentalHealthAssistant
from resources_database import Resource, init_db
from sqlalchemy.orm import sessionmaker
import uvicorn
import datetimeapp = FastAPI(title="心理健康支持系统API",description="提供心理咨询、情感分析和危机干预服务",version="1.0.0"
)# 初始化模型和数据库
assistant = MentalHealthAssistant()
engine = init_db()
Session = sessionmaker(bind=engine)class ChatRequest(BaseModel):text: struser_id: str = "anonymous"session_id: str = "default"max_length: int = 250class ResourceRequest(BaseModel):category: str = Noneemergency: bool = Falselanguage: str = "zh"@app.on_event("startup")
async def startup_event():# 初始化日志open("chat_log.csv", "a").write("timestamp,user_id,session_id,message,emotion,crisis_detected\n")@app.post("/chat")
async def chat_endpoint(request: ChatRequest):try:# 检测危机信号crisis_detected = assistant.detect_crisis(request.text)# 分析情绪emotion_result = assistant.analyze_emotion(request.text)# 生成响应response = assistant.generate_response(request.text)# 记录对话（实际应用中需匿名化处理）timestamp = datetime.datetime.now().isoformat()log_entry = f"{timestamp},{request.user_id},{request.session_id},{request.text},{emotion_result['emotion']},{crisis_detected}\n"with open("chat_log.csv", "a") as log_file:log_file.write(log_entry)# 构建响应result = {"response": response,"emotion": emotion_result,"crisis_detected": crisis_detected}# 添加危机资源（如果检测到危机）if crisis_detected:result["crisis_resources"] = get_resources(category="危机干预", emergency=True)return resultexcept Exception as e:raise HTTPException(status_code=500, detail=str(e))@app.get("/resources")
def get_resources(category: str = None, emergency: bool = False, language: str = "zh"):try:session = Session()query = session.query(Resource).filter(Resource.language == language)if category:query = query.filter(Resource.category == category)if emergency:query = query.filter(Resource.emergency == 1)resources = query.all()return [{"id": r.id,"category": r.category,"name": r.name,"description": r.description,"contact": r.contact,"website": r.website,"emergency": bool(r.emergency)} for r in resources]except Exception as e:raise HTTPException(status_code=500, detail=str(e))if __name__ == "__main__":uvicorn.run(app, host="0.0.0.0", port=8000)

4. 专业级Streamlit前端界面

创建 mentalhealth_app.py：

import streamlit as st
import requests
import pandas as pd
import time
import matplotlib.pyplot as plt
import json# 设置页面
st.set_page_config(page_title="心理健康支持助手",page_icon="❤️",layout="wide",initial_sidebar_state="expanded"
)# 自定义CSS样式
st.markdown("""
<style>/* 主容器 */.main-container {background-color: #f9f7f7;border-radius: 15px;padding: 25px;box-shadow: 0 4px 8px rgba(0,0,0,0.05);}/* 聊天框 */.user-msg {background-color: #e3f2fd;border-radius: 15px 15px 0 15px;padding: 15px;margin: 10px 0 10px 20%;border-left: 4px solid #2196f3;}.bot-msg {background-color: #e8f5e9;border-radius: 15px 15px 15px 0;padding: 15px;margin: 10px 20% 10px 0;border-left: 4px solid #4caf50;}.crisis-msg {background-color: #ffebee;border-radius: 15px;padding: 15px;margin: 15px 0;border-left: 4px solid #f44336;animation: pulse 2s infinite;}@keyframes pulse {0% { box-shadow: 0 0 0 0 rgba(244, 67, 54, 0.4); }70% { box-shadow: 0 0 0 10px rgba(244, 67, 54, 0); }100% { box-shadow: 0 0 0 0 rgba(244, 67, 54, 0); }}/* 情绪指示器 */.emotion-indicator {display: inline-block;padding: 3px 10px;border-radius: 12px;font-size: 0.8em;margin-right: 5px;}
</style>
""", unsafe_allow_html=True)# 情绪颜色映射
EMOTION_COLORS = {'love': '#e91e63', 'joy': '#ffc107', 'surprise': '#ff9800','anger': '#f44336', 'sadness': '#2196f3', 'fear': '#673ab7','neutral': '#9e9e9e'
}# API配置
API_URL = "http://localhost:8000/chat"
RESOURCES_URL = "http://localhost:8000/resources"# 初始化会话状态
def init_session():if "history" not in st.session_state:st.session_state.history = []if "session_start" not in st.session_state:st.session_state.session_start = time.time()if "user_id" not in st.session_state:st.session_state.user_id = f"user_{int(time.time())}"if "session_id" not in st.session_state:st.session_state.session_id = f"session_{int(time.time())}"if "emotion_data" not in st.session_state:st.session_state.emotion_data = []if "crisis_detected" not in st.session_state:st.session_state.crisis_detected = False# 显示聊天记录
def display_chat():st.markdown("<div class='main-container'>", unsafe_allow_html=True)# 显示危机警告（如果有）if st.session_state.crisis_detected:st.markdown("""<div class='crisis-msg'><h3>⚠️ 重要提示：检测到危机信号</h3><p>我们注意到您可能在经历极度困难的时刻。请记住：</p><ul><li>您并不孤单</li><li>您的感受是重要的</li><li>专业帮助触手可及</li></ul><p>在下方查看紧急资源 →</p></div>""", unsafe_allow_html=True)# 显示对话历史for i, msg in enumerate(st.session_state.history):if msg["role"] == "user":# 显示用户消息和情绪标签emotion_html = ""if 'emotion' in msg:emotion = msg['emotion']['emotion']confidence = msg['emotion']['confidence']color = EMOTION_COLORS.get(emotion, '#9e9e9e')emotion_html = f"""<div style="text-align:right; margin-bottom:5px;"><span class="emotion-indicator" style="background:{color};color:white;">{emotion} ({confidence:.0%})</span></div>"""st.markdown(f"""{emotion_html}<div class="user-msg"><b>👤 您:</b> {msg["content"]}</div>""", unsafe_allow_html=True)else:st.markdown(f"""<div class="bot-msg"><b>🤖 心理咨询助手:</b> {msg["content"]}</div>""", unsafe_allow_html=True)st.markdown("</div>", unsafe_allow_html=True)# 显示资源面板
def display_resources():st.sidebar.header("心理健康资源")category = st.sidebar.selectbox("选择资源类型", ["全部", "危机干预", "抑郁症", "焦虑症", "创伤治疗", "青少年心理"])emergency_only = st.sidebar.checkbox("仅显示紧急服务")try:resources = requests.get(RESOURCES_URL, params={"category": category if category != "全部" else None,"emergency": emergency_only}).json()if not resources:st.sidebar.info("未找到匹配的资源")returnfor resource in resources:with st.sidebar.expander(f"🔗 {resource['name']}"):st.markdown(f"**类别**: {resource['category']}")if resource['emergency']:st.markdown("**紧急服务**: ⚠️ 是")st.markdown(f"**描述**: {resource['description']}")if resource['contact']:st.markdown(f"**联系方式**: `{resource['contact']}`")if resource['website']:st.markdown(f"**网站**: [{resource['website']}]({resource['website']})")except Exception as e:st.sidebar.error(f"获取资源失败: {str(e)}")# 显示情绪分析图表
def display_emotion_chart():if not st.session_state.emotion_data:returnst.subheader("情绪变化分析")# 准备图表数据df = pd.DataFrame(st.session_state.emotion_data)df['timestamp'] = pd.to_datetime(df['timestamp'])df.set_index('timestamp', inplace=True)# 情绪分布饼图fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(14, 6))emotion_counts = df['emotion'].value_counts()colors = [EMOTION_COLORS.get(e, '#9e9e9e') for e in emotion_counts.index]ax1.pie(emotion_counts, labels=emotion_counts.index, autopct='%1.1f%%', colors=colors, startangle=90)ax1.set_title('情绪分布')# 情绪趋势图emotion_over_time = df.groupby([pd.Grouper(freq='5T'), 'emotion']).size().unstack().fillna(0)emotion_over_time.plot(kind='area', stacked=True, ax=ax2, color=[EMOTION_COLORS.get(e, '#9e9e9e') for e in emotion_over_time.columns])ax2.set_title('情绪随时间变化')ax2.set_xlabel('时间')ax2.set_ylabel('消息数量')ax2.legend(title='情绪', loc='upper left')st.pyplot(fig)# 主应用
def main():init_session()# 标题和介绍st.title("❤️ 心理健康支持助手")st.markdown("""<div style="background-color:#e3f2fd; padding:15px; border-radius:10px; margin-bottom:20px;"><p>欢迎使用心理健康支持助手。这是一个安全的倾诉空间，您可以随时表达您的感受和想法。</p><p><b>重要提示：</b>本系统不能替代专业心理咨询。如果您处于危机中，请立即联系紧急服务。</p></div>""", unsafe_allow_html=True)# 侧边栏资源display_resources()# 聊天界面display_chat()# 用户输入user_input = st.chat_input("分享您的感受或想法...")if user_input:# 添加用户消息到历史user_message = {"role": "user", "content": user_input}st.session_state.history.append(user_message)# 调用API获取响应try:response = requests.post(API_URL,json={"text": user_input,"user_id": st.session_state.user_id,"session_id": st.session_state.session_id}).json()if "response" in response:# 添加情绪数据if "emotion" in response:user_message["emotion"] = response["emotion"]st.session_state.emotion_data.append({"timestamp": pd.Timestamp.now(),"emotion": response["emotion"]["emotion"],"confidence": response["emotion"]["confidence"]})# 添加助手响应bot_message = {"role": "bot", "content": response["response"]}st.session_state.history.append(bot_message)# 检查危机信号if response.get("crisis_detected", False):st.session_state.crisis_detected = Truest.experimental_rerun()else:st.error(f"API错误: {response.get('error', '未知错误')}")except Exception as e:st.error(f"连接API失败: {str(e)}")# 刷新显示st.experimental_rerun()# 显示情绪分析图表if st.session_state.emotion_data:display_emotion_chart()# 添加控制按钮st.sidebar.divider()if st.sidebar.button("清空对话历史"):st.session_state.history = []st.session_state.emotion_data = []st.session_state.crisis_detected = Falsest.experimental_rerun()if st.sidebar.button("导出对话记录 (匿名)"):history_text = "\n\n".join(f"{msg['role']}: {msg['content']}" for msg in st.session_state.history)st.sidebar.download_button(label="下载对话记录",data=history_text,file_name="mentalhealth_conversation.txt",mime="text/plain")# 页脚st.sidebar.divider()st.sidebar.markdown("""<div style="font-size:0.8em; color:#666;"><p><b>隐私声明:</b> 所有对话内容将被匿名记录，仅用于改善服务质量。</p><p><b>紧急情况:</b> 如果您或您认识的人处于危机中，请立即拨打当地紧急电话。</p></div>""", unsafe_allow_html=True)if __name__ == "__main__":main()

系统部署与运行

启动步骤：

初始化资源数据库：
```
python resources_database.py
```

启动FastAPI后端：

uvicorn mentalhealth_api:app --reload --port 8000

启动Streamlit前端：
```
streamlit run mentalhealth_app.py
```

访问地址：

前端界面：http://localhost:8501
API文档：http://localhost:8000/docs

专业功能详解

1. 多层次危机检测系统

def detect_crisis(self, text):# 关键词匹配keyword_detected = any(keyword in text_lower for keyword in self.crisis_keywords)# 情感强度分析emotion_result = self.analyze_emotion(text)intense_emotion = emotion_result['confidence'] > 0.9 and emotion_result['emotion'] in ['sadness', 'anger', 'fear']# 语义分析（简化版）semantic_crisis = any(["不想活了" in text,"看不到希望" in text,"无法继续" in text])return keyword_detected or intense_emotion or semantic_crisis

2. 专业心理咨询响应模式

# 心理咨询专用提示工程
prompt = """
作为心理咨询师，我理解你正在经历困难。让我们谈谈你的感受。
请记住：
- 保持同理心和耐心
- 避免评判性语言
- 使用开放式问题
- 关注情绪而非解决方案对话历史:
{history}当前对话:
用户: {input_text}
咨询师:
"""# 生成参数优化
outputs = self.counseling_model.generate(inputs,max_length=max_length,temperature=0.85,  # 平衡创造性和一致性top_p=0.92,        # 提高多样性repetition_penalty=1.2,  # 减少重复num_beams=3,             # 提高响应质量
)

3. 情感追踪与分析

# 情感随时间变化分析
emotion_over_time = df.groupby([pd.Grouper(freq='5T'), 'emotion']).size().unstack().fillna(0)# 生成交互式图表
fig = px.area(emotion_over_time, labels={'value':'消息数量', 'timestamp':'时间'},title='情绪随时间变化')
fig.update_layout(legend_title='情绪')
st.plotly_chart(fig)

伦理与安全措施

1. 隐私保护机制

数据匿名化：用户ID使用哈希处理

敏感信息过滤：

def sanitize_input(text):# 移除个人信息text = re.sub(r'\b\d{11}\b', '[PHONE]', text)  # 电话号码text = re.sub(r'\b\d{18}\b', '[ID]', text)     # 身份证号text = re.sub(r'\b\w+@\w+\.\w+\b', '[EMAIL]', text)  # 邮箱return text