鏈接:https://pan.baidu.com/s/1ZiMzKulcsEt2xo_a2XK1nw
提取碼:9umt
import pandas as pd
import fool
import re
import random
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.linear_model import LogisticRegression
# -----------------------------------------------------
# 加載停用詞詞典
stopwords = {}
with open(r'stopword.txt', 'r', encoding='utf-8') as fr:
for word in fr:
stopwords[word.strip()] = 0
# -----------------------------------------------------
# 定義類
class clf_model:
"""
該類將所有模型訓練、預測、數據預處理、意圖識別的函數包括其中
"""
# 初始化模塊
def __init__(self):
self.model = "" # 成員變量,用於存儲模型
self.vectorizer = "" # 成員變量,用於存儲tfidf統計值
# 訓練模塊
def train(self):
"""
訓練結果存儲在成員變量中,沒有return
"""
# 從excel文件讀取訓練樣本
d_train = pd.read_excel("data_train.xlsx")
# 對訓練數據進行預處理
d_train.sentence_train = d_train.sentence_train.apply(self.fun_clean)
print("訓練樣本 = %d" % len(d_train))
# 利用sklearn中的函數進行tifidf訓練
self.vectorizer = TfidfVectorizer(analyzer="word",
token_pattern=r"(?u)\b\w+\b") # 注意,這裏自己指定token_pattern,否則sklearn會自動將一個字長度的單詞過濾篩除
features = self.vectorizer.fit_transform(d_train.sentence_train)
print("訓練樣本特徵表長度爲 " + str(features.shape))
# 使用邏輯迴歸進行訓練和預測
self.model = LogisticRegression(C=10)
self.model.fit(features, d_train.label)
# 預測模塊(使用模型預測)
def predict_model(self, sentence):
# --------------
# 對樣本中沒有點特殊情況做特別判斷
if sentence in ["好的", "需要", "是的", "要的", "好", "要", "是"]:
return 1, 0.8
# --------------
sent_features = self.vectorizer.transform([sentence])
pre_test = self.model.predict_proba(sent_features).tolist()[0]
clf_result = pre_test.index(max(pre_test))
score = max(pre_test)
return clf_result, score
# 預測模塊(使用規則)
def predict_rule(self, sentence):
"""
如果模型訓練出現異常,可以使用規則進行預測,同時也可以讓學員融合"模型"及"規則"的預測方式
:param sentence:
:return 預測結果:
"""
sentence = sentence.replace(' ', '')
if re.findall(r'不需要|不要|停止|終止|退出|不買|不定|不訂', sentence):
return 2, 0.8
elif re.findall(r'訂|定|預定|買|購', sentence) or sentence in ["好的", "需要", "是的", "要的", "好", "要", "是"]:
return 1, 0.8
else:
return 0, 0.8
# 預處理函數
def fun_clean(self, sentence):
"""
預處理函數
:輸入 用戶輸入語句:
:輸出 預處理結果:
"""
# 使用foolnltk進行實體識別
words, ners = fool.analysis(sentence)
# 對識別結果按長度倒序排序
ners = ners[0].sort(key=lambda x: len(x[-1]), reverse=True)
# 如果有實體被識別出來,就將實體的字符串替換成實體類別的字符串(目的是看成一類單詞,看成一種共同的特徵)
if ners:
for ner in ners:
sentence = sentence.replace(ner[-1], ' ' + ner[2] + ' ')
# 分詞,並去除停用詞
word_lst = [w for w in fool.cut(sentence)[0] if w not in stopwords]
output_str = ' '.join(word_lst)
output_str = re.sub(r'\s+', ' ', output_str)
return output_str.strip()
# 分類主函數
def fun_clf(self, sentence):
"""
意圖識別函數
:輸入 用戶輸入語句:
:輸出 意圖類別,分數:
"""
# 對用戶輸入進行預處理
sentence = self.fun_clean(sentence)
# 得到意圖分類結果(0爲“查詢”類別,1爲“訂票”類別,2爲“終止服務”類別)
clf_result, score = self.predict_model(sentence) # 使用訓練的模型進行意圖預測
# clf_result, score = self.predict_rule(sentence) # 使用規則進行意圖預測(可與用模型進行意圖識別的方法二選一)
return clf_result, score
def fun_replace_num(sentence):
"""
替換時間中的數字(目的是便於實體識別包fool對實體的識別)
:param sentence:
:return sentence:
"""
# 定義要替換的數字
time_num = {"一": "1", "二": "2", "三": "3", "四": "4", "五": "5", "六": "6", "七": "7", "八": "8", "九": "9", "十": "10",
"十一": "11", "十二": "12"}
for k, v in time_num.items():
sentence = sentence.replace(k, v)
return sentence
def slot_fill(sentence, key=None):
"""
填槽函數(該函數從sentence中尋找需要的內容,完成填槽工作)
:param sentence:
:return slot: (填槽的結果)
"""
slot = {}
# 進行實體識別
words, ners = fool.analysis(sentence)
to_city_flag = 0 # flag爲1代表找到到達城市(作用:當找到到達城市時,默認句子中另一個城市信息是出發城市)
for ner in ners[0]:
# 首先對time類別的實體進行信息抽取填槽工作
if ner[2] == 'time':
# --------------------
# 尋找日期的關鍵詞
date_content = re.findall(
r'後天|明天|今天|大後天|週末|週一|週二|週三|週四|週五|週六|週日|本週一|本週二|本週三|本週四|本週五|本週六|本週日|下週一|下週二|下週三|下週四|下週五|下週六|下週日|這週一|這週二|這週三|這週四|這週五|這週六|這週日|\d{,2}月\d{,2}號|\d{,2}月\d{,2}日',
ner[-1])
slot["date"] = date_content[0] if date_content else ""
# 完成日期的填槽
# --------------------
# --------------------
# 尋找具體時間的關鍵詞
time_content = re.findall(r'\d{,2}點\d{,2}分|\d{,2}點鐘|\d{,2}點', ner[-1])
# 尋找上午下午的關鍵詞
pmam_content = re.findall(r'上午|下午|早上|晚上|中午|早晨', ner[-1])
slot["time"] = pmam_content[0] if pmam_content else "" + time_content[0] if time_content else ""
# 完成時間的填槽
# --------------------
# 對location類別對實體進行信息抽取填槽工作
if ner[2] == 'location':
# --------------------
# 開始對城市填槽
# 如果沒有指定槽位
if key is None:
if re.findall(r'(到|去|回|回去)%s' % (ner[-1]), sentence):
to_city_flag = 1
slot["to_city"] = ner[-1]
continue
if re.findall(r'從%s|%s出發' % (ner[-1], ner[-1]), sentence):
slot["from_city"] = ner[-1]
elif to_city_flag == 1:
slot["from_city"] = ner[-1]
# 如果指定了槽位
elif key in ["from_city", "to_city"]:
slot[key] = ner[-1]
# 完成出發城市、到達城市的填槽工作
# --------------------
return slot
def fun_wait(clf_obj):
"""
等待詢問函數
:輸入 None:
:輸出 用戶意圖類別:
"""
# 等待用戶輸入
print("\n\n\n")
print("-------------------------------------------------------------")
print("-------------------------------------------------------------")
print("Starting ...")
sentence = input("客服:請問需要什麼服務?(時間請用12小時製表示)\n")
# 對用戶輸入進行意圖識別
clf_result, score = clf_obj.fun_clf(sentence)
return clf_result, score, sentence
def fun_search(clf_result, sentence):
"""
爲用戶查詢餘票
:param clf_result:
:param sentence:
:return: 是否有票
"""
# 定義槽存儲空間
name = {"time": "出發時間", "date": "出發日期", "from_city": "出發城市", "to_city": "到達城市"}
slot = {"time": "", "date": "", "from_city": "", "to_city": ""}
# 使用用戶第一句話進行填槽
sentence = fun_replace_num(sentence)
slot_init = slot_fill(sentence)
for key in slot_init.keys():
slot[key] = slot_init[key]
# 對未填充對槽位,向用戶提問,進行鍼對性填槽
while "" in slot.values():
for key in slot.keys():
if slot[key] == "":
sentence = input("客服:請問%s是?\n" % (name[key]))
sentence = fun_replace_num(sentence)
slot_cur = slot_fill(sentence, key)
for key in slot_cur.keys():
if slot[key] == "":
slot[key] = slot_cur[key]
# 查詢是否有票,並答覆用戶(本次查詢是否有票使用隨機數完成)
if random.random() > 0.5:
print("客服:%s%s從%s到%s的票充足" % (slot["date"], slot["time"], slot["from_city"], slot["to_city"]))
# 返回1表示有票
return 1
else:
print("客服:%s%s從%s到%s無票" % (slot["date"], slot["time"], slot["from_city"], slot["to_city"]))
print("End !!!")
print("-------------------------------------------------------------")
print("-------------------------------------------------------------")
# 返回0表示無票
return 0
def fun_book():
"""
爲用戶訂票
"""
print("客服:已爲您完成訂票。\n\n\n")
print("End !!!")
print("-------------------------------------------------------------")
print("-------------------------------------------------------------")
if __name__ == "__main__":
# 實例化對象
clf_obj = clf_model()
clf_obj.train()
threshold = 0.55 # 用戶定義閾值(當分類器分類的分數大於閾值才採納本次意圖分類結果,目的是排除分數過低的意圖分類結果)
while 1:
clf_result, score, sentence = fun_wait(clf_obj)
# -------------------------------------------------------------------------------
# 狀態轉移條件(等待-->等待):用戶輸入未達到“查詢”、“訂票”類別的閾值 OR 被分類爲“終止服務”
# -------------------------------------------------------------------------------
if score < threshold or clf_result == 2:
continue
# -------------------------------------------------------------------------------
# 狀態轉移條件(等待-->查詢):用戶輸入分類爲“查詢” OR “訂票”
# -------------------------------------------------------------------------------
else:
search_result = fun_search(clf_result, sentence)
if search_result == 0:
continue
else:
# 等待用戶輸入
sentence = input("客服:需要爲您訂票嗎?\n")
# 對用戶輸入進行意圖識別
clf_result, score = clf_obj.fun_clf(sentence)
# -------------------------------------------------------------------------------
# 狀態轉移條件(查詢-->訂票):FUN_SEARCH返回有票 AND 用戶輸入分類爲“訂票”
# -------------------------------------------------------------------------------
if clf_result == 1:
fun_book()
continue
運行結果