自然语言处理学习日记14

1.PolicyTrainer类
解析：Trains a policy on a domain using training data from a file.
[1]param augmentation_factor: how many stories should be created by randomly concatenating stories
[2]param filename: story file containing the training conversations
[3]param max_history: number of past actions to consider for the prediction of the next action
[4]param max_training_samples: specifies how many training samples to train on
[5]param max_number_of_trackers: limits the tracker generation during story file parsing
[6]param remove_duplicates: remove duplicates from the training set before training
[7]param kwargs: additional arguments passed to the underlying ML trainer

2.命名实体识别评估指标
解析：命名实体识别通常采用的评估指标主要有正确率、召回率和F值，如下所示：
[1]正确率 = 识别出的正确实体数 / 识别出的实体数
[2]召回率 = 识别出的正确实体数 / 样本的实体数
[3]F1值 = [2*正确率*召回率] / [正确率+召回率]

3.FallbackPolicy
解析：

class ActionDefaultFallback(Action):
    def name(self):
        return "action_default_fallback"
    def run(self, dispatcher, tracker, domain):
        text = tracker.latest_message.get('text')
        message = get_response(text)
        if message['code'] == 100000:
            dispatcher.utter_message("{}".format(message['text']))
        else:
            dispatcher.utter_template('utter_default', tracker)
        return [UserUtteranceReverted()]

如果意图和预测的action分别低于这两个阈值的时候，便会调用action_default_fallback这个自定义的action。action_default_fallback在rasa-core中的实现：

class ActionDefaultFallback(Action):
    def name(self):
        return "action_default_fallback"
    def run(self, dispatcher, tracker, domain):
        text = tracker.latest_message.get('text')
        message = get_response(text)
        if message['code'] == 100000:
            dispatcher.utter_message("{}".format(message['text']))
        else:
            dispatcher.utter_template('utter_default', tracker)
        return [UserUtteranceReverted()]

如果调用图灵机器人api，那么get_response实现：

def get_response(msg):
    key = apikey
    api = 'http://www.tuling123.com/openapi/api?key={}&info={}'.format(key, msg)
    return requests.get(api).json()

4.Tracker Stores
解析：
[1]InMemoryTrackerStore [default]
[2]SQLTrackerStore
[3]RedisTrackerStore
[4]MongoTrackerStore
[5]Custom Tracker Store

5.default_actions()
解析：
[1]ActionListen()
[2]ActionRestart()
[3]ActionDefaultFallback()
[4]ActionDeactivateForm()
[5]ActionRevertFallbackEvents()
[6]ActionDefaultAskAffirmation()
[7]ActionDefaultAskRephrase()
[8]ActionBack()

6.Fallback Actions
解析：action_default_fallback is a default action in Rasa Core which sends the utter_default template message to the user.

7.Actions
解析：
[1]Utterance actions: start with utter_ and send a specific message to the user
[2]Retrieval actions: start with respond_ and send a message selected by a retrieval model
[3]Custom actions: run arbitrary code and send any number of messages (or none).
[4]Default actions: e.g. action_listen, action_restart, action_default_fallback
说明：rasa_core_sdk已经改名为rasa_sdk。

8.rasa run actions
解析：使用Rasa SDK启动操作服务。

9.custom actions
解析：custom actions在Rasa中的使用方式并不是直接调用，而是采用服务的形式，所以如果想使用自定义的action，还需要定义一个endpoints.yml文件，文件内容如下：

action_endpoint:
  url: 'http://localhost:5055/webhook'

在启动会话的时候添加额外的命令–endpoints endpoints.yml，该命令会在5055端口启动一个服务，这个服务就是自定义的action。

10.Rasa词槽类型
解析：
[1]Text Slot：文本类型
[2]Boolean Slot：布尔类型
[3]Categorical Slot：分类类型
[4]Float Slot：浮点类型
[5]List Slot：列表类型
[6]Unfeaturized Slot：想保存下来的信息

11.MemoizationPolicy
解析：MemoizationPolicy只记住训练数据中的对话。如果训练数据中存在这样的对话，那么它将以置信度为1.0预测下一个动作，否则将预测为None，此时置信度为0.0。其中，max_history决定了模型查看多少个对话历史以决定下一个执行的action。

12.EmbeddingPolicy
解析：循环嵌入式对话策略[Recurrent Embedding Dialogue Policy，REDP]，它通过将actions和对话状态嵌入到相同的向量空间[vector space]能够获得较好的效果，REDP包含一个基于改进的Neural Turing Machine的记忆组件和注意机制，在该任务上显著优于基线LSTM分类器。

13.FormPolicy
解析：FormPolicy是MemoizationPolicy的扩展，用于处理表单的填充事项。当一个FormAction被调用时，FormPolicy将持续预测表单动作，直到表单中的所有槽都被填满，然后再执行对应的FormAction。

14.MappingPolicy
解析：MappingPolicy可用于直接将意图映射到要执行的action，从而实现被映射的action总会被执行，其中，这种映射是通过triggers属性实现的。[domain.yml文件]如下所示：

intents:
 - greet: {triggers: utter_goodbye}

其中，greet是意图，utter_goodbye是action。一个意图最多只能映射到一个action，机器人一旦收到映射意图的消息，它将执行对应的action。然后继续监听下一条message。需要注意的是，对于上述映射，还需要要在story.md文件中添加如下样本，否则任何机器学习策略都可能被预测的action_greet在dialouge历史中突然出现而混淆。

15.credentials.yml
解析：使用credentials.yml来存储对应的访问权限信息。如下所示：

twilio:
account_sid: "ACbc2dxxxxxxxxxxxx19d54bdcd6e41186"
auth_token: "e231c197493a7122d475b4xxxxxxxxxx"
twilio_number: "+440123456789"

slack:
slack_token: "xoxb-286425452756-safjasdf7sl38KLls"
slack_channel: "@my_channel"

telegram:
access_token: "490161424:AAGlRxinBRtKGb21_rlOEMtDFZMXBl6EC0o"
verify: "your_bot"
webhook_url: "your_url.com/webhook"

mattermost:
url: "https://chat.example.com/api/v4"
team: "community"
user: "user@user.com"
pw: "password"

facebook:
verify: "rasa-bot"
secret: "3e34709d01ea89032asdebfe5a74518"
page - access - token: "EAAbHPa7H9rEBAAuFk4Q3gPKbDedQnx4djJJ1JmQ7CAqO4iJKrQcNT0wtD"

webexteams:
access_token: "ADD-YOUR-BOT-ACCESS-TOKEN"
room: "YOUR-WEBEXTEAMS-ROOM-ID"

rocketchat:
user: "yourbotname"
password: "YOUR_PASSWORD"
server_url: "https://demo.rocket.chat"

# socket通道
# 前两个配置值定义Rasa Core在通过socket.io发送或接收消息时使用的事件名称
socketio:
user_message_evt: user_uttered
bot_message_evt: bot_uttered
session_persistence: true / false

# rest通道
rest:
# you don't need to provide anything here - this channel doesn't
# require any credentials

# CallbackInput通道
callback:
# URL to which Core will send the bot responses
url: "http://localhost:5034/bot"

当需要从开发的客户端访问人机对话系统，可以通过使用socket和http通道来实现，它们分别对应socketio输入通道和rest输入通道，而callback与rest通道类似，均是走HTTP协议，但是它不会直接将bot消息返回给发送消息的HTTP请求，而是调用一个URL，可以指定该URL来发送bot消息。

16.endpoints.yml
解析：在Rasa Core项目创建endpoint.yml文件，该文件用于指定RasaCore将要访问的CustomeAction web和nlu web，当rasa core需要执行意图、实体提取和执行action时，就会根据url找到对应的web进行执行。

17.Event对象
解析：Event对象是Rasa Core描述会话中发生的所有事件和明确rasa.core.trackers.DialogueStateTracker该如何更新其状态的基类，因此不能被直接使用，而是通过它包含的具体事件[event]实现。

18.rasa run actions和rasa x --endpoints ENDPOINTS
解析：
[1]rasa run actions --actions ACTIONS：name of action package to be load [default: None]。
[2]rasa x --endpoints ENDPOINTS：configuration file for the model server and the connectors as a yml file [default: None]。

19.class Dispatcher(sender_id, output_channel, domain)
解析：Send messages back to user，如下所示：
[1]utter_response(message)：Send a message to the client.
[2]utter_message(text)：Send a text to the output channel
[3]utter_custom_message(*elements)：Sends a message with custom elements to the output channel.
[4]utter_button_message(text, buttons, **kwargs)：Sends a message with buttons to the output channel.
[5]utter_attachment(attachment)：Send a message to the client with attachments.
[6]utter_button_template(template, buttons, filled_slots=None, **kwargs)：Sends a message template with buttons to the output channel.
[7]utter_template(template, filled_slots=None, **kwargs)：Send a message to the client based on a template.
[8]_template_variables(kwargs)：Combine slot values and key word arguments to fill templates.
[9]_fill_template_text(template, filled_slots=None, **kwargs)
[10]retrieve_template(template_name, filled_slots=None, **kwargs)：Retrieve a named template from the domain.

20.Python Logging Options
解析：
[1]-v[–verbose]：Be verbose. Sets logging level to INFO. [default: None]
[2]-vv[–debug]：Print lots of debugging statements. Sets logging level to DEBUG. [default: None]
[3]–quiet：Be quiet! Sets logging level to WARNING. [default: None]

21.rasa-ui
解析：Rasa UI is a frontend for the Rasa Framework。如下所示：
[1]UI for creating and managing training data - Examples, Intents, Entities, Synonyms, Regex, Stories, Actions, Responses
[2]Manage multiple bots from a single UI/instance of Rasa UI
[3]Create, manage and load different versions of your models for testing and optimizing
[4]Log requests for usage tracking, history and improvements to your models
[5]Easily execute intent parsing using different models
[6]Data is stored in a SQLite DB for backing up/sharing
[7]Can be used with or without a Rasa backend to manage your training data

22.get_slot(key)
解析：retrieves the value of a slot, the return type is Optional[Any].

23.ActionCheckRestaurants
解析：

from rasa_sdk import Action
from rasa_sdk.events import SlotSet

class ActionCheckRestaurants(Action):
   def name(self) -> Text:
      return "action_check_restaurants"
   def run(self,
           dispatcher: CollectingDispatcher,
           tracker: Tracker,
           domain: Dict[Text, Any]) -> List[Dict[Text, Any]]:
      cuisine = tracker.get_slot('cuisine')
      q = "select * from restaurants where cuisine='{0}' limit 1".format(cuisine)
      result = db.query(q)
      return [SlotSet("matches", result if result is not None else [])]

24.dispatcher对象
解析：
[1]dispatcher.utter_template()
[2]dispatcher.utter_message()
[3]rasa_sdk.executor.CollectingDispatcher()

25.typing常用类型
解析：
[1]int，long，float：整型，长整形，浮点型
[2]bool，str：布尔型，字符串类型
[3]List，Tuple，Dict，Set：列表，元组，字典，集合
[4]Iterable，Iterator：可迭代类型，迭代器类型
[5]Generator：生成器类型

参考文献：
[1]rasa对话系统踩坑记（四）：https://www.jianshu.com/p/9393d319e698
[2]Fallback Actions：https://rasa.com/docs/rasa/core/fallback-actions/#fallback-actions
[3]Rasa命令行：https://panchuang.net/2019/08/31/rasa_command_line_interface/
[4]dispatcher：https://www.pydoc.io/pypi/rasa-core-0.9.0a1/autoapi/dispatcher/index.html