十分鐘快速了解 Rasa Core

一個聰明的Chatbot 必須要能夠理解使用者說的話, 才能給予最精準的反饋, 要打造出這樣的功能, 可以使用諸如以下的線上語意識別服務 Google DialogFlow, Microsoft Luis, 以及Facebook Wit等等　　

這些語意識別的服務猶如Chatbot的大腦語言中樞, 用來了解每句對話背後的意圖

有了意圖之後, 程式可以根據意圖的種類來回答問題

if intent == 'MAKE_ORDER':

    print('okay total price is $100')
   
elif intent == 'CHECK_SHOPPING_CART':

    print('there are milk, beef, ...')

elif intent == 'CHANGE_PYMENT':

    print('change complete')

else:
    print('hi')

但光是了解自然語言是不夠的, 若要打造出更人性化的Chatbot, 還需要作到對話管理, 上下文理解等等...

Rasa Core

想要作到對話管理, 不妨試試Open Source Rasa Core


功能介紹

Rasa Core可以用神經網路去學習預先定義好的對話邏輯(順序), 學完之後就能根據先前對話的資訊, 來推測Chatbot接下來的最佳反饋應該是什麼


什麼是 Stories, Domain, Action, Policy


Stories

用來定義對話邏輯


一個適當的對話一定會有前後順序, 上下文關係, Story 就是用來描述這種上下文關係的資訊

## happy path               <!-- 故事名 - just for debugging -->

* _greet

  - utter_greet

* _mood_great               <!-- user utterance, in format _intent[entities] -->

  - utter_happy

## sad path 1               <!-- this is already the start of the next story -->

* _greet

  - utter_greet             <!-- action of the bot to execute -->

* _mood_unhappy

  - utter_cheer_up

  - utter_did_that_help

* _mood_affirm

  - utter_happy

## sad path 2

* _greet

  - utter_greet

* _mood_unhappy

  - utter_cheer_up

  - utter_did_that_help

* _mood_deny

  - utter_goodbye

## say goodbye

* _goodbye

  - utter_goodbye

Domains

用來描述定義在story中每個項目的細節, 包含用到的意圖, slot, action

intents:

- greet

- goodbye

- mood_affirm

- mood_deny

- mood_great

- mood_unhappy

- inform

slots:

  group:

    type: text

    

entities:

- group

actions:

- utter_greet

- utter_did_that_help

- utter_happy

- utter_goodbye

- utter_unclear

- utter_ask_picture

- __main__.ApiAction

templates:

  utter_greet:

  - text: "Hey! How are you?"

  utter_did_that_help:

  - text: "Did that help you?"

  utter_unclear:

  - text: "I am not sure what you are aiming for."

  

  utter_happy:

  - text: "Great carry on!"

  utter_goodbye:

  - text: "Bye"

  

  utter_ask_picture:

  - text: "To cheer you up, I can show you a cute picture of a dog, cat or a bird. Which one do you choose?"

Actions

Action可分兩種

1. Utterance Action: 單純回話
2. Custom Action: 可以執行特定動作, 如串接下訂單的API, 透過呼叫指定的URL來作動作, 若沒有特定的REST API 來服務這個動作, 也可以透過Rasa Core 起一個內建的Action server,　在裡面去實作客製化的Action讓 Rasa core去呼叫它

Policy


訓練的時候, 可以設定要使用哪種Policy 來訓練對話模型, 而產出的模型可以用來推測最佳的Action

from rasa_core.policies import FallbackPolicy, KerasPolicy, MemoizationPolicy

from rasa_core.agent import Agent

# this will catch predictions the model isn't very certain about

# there is a threshold for the NLU predictions as well as the action predictions

fallback = FallbackPolicy(fallback_action_name="utter_unclear",

                          core_threshold=0.2,

                          nlu_threshold=0.1)

agent = Agent('domain.yml', policies=[MemoizationPolicy(), KerasPolicy(), fallback])

# loading our neatly defined training dialogues

training_data = agent.load_data('stories.md')

agent.train(

    training_data,

    validation_split=0.0,

    epochs=200

)

agent.persist('models/dialogue')

目前預設是Keras Policy, 以LSTM的神經網路去學習對話情境(原始碼如下)

 # Build Model
 model = Sequential()
 model.add(Masking(mask_value=-1,input_shape=(None, input_shape[1])))
 model.add(LSTM(self.rnn_size, return_sequences=True, dropout=0.2))
 model.add(TimeDistributed(Dense(units=output_shape[-1])))
 model.add(Activation('softmax'))
 model.compile(loss='categorical_crossentropy',
                      optimizer='rmsprop',
                      metrics=['accuracy'])

另外其他種Policy還有MemoizationPolicy, Embedding Policy

結論

要打造人性化的Chatbot, 除了要能理解語意之外, 還要能作到對話管理, 上下文理解等等, 但是要自己去實作以上的功能其實真的很花時間, Rasa Core可以讓Chatbot的開發者省掉這個工, 進而更專注在商務邏輯上的開發, 提供更好的服務給前端使用者