Desarrollo de Chatbots con Rasa

Rafa Haro Ramos

​

@rafa_haro

rafa_haro

​

Agenda

• Chatbots Overview
• What is Rasa?
• Demo
• Rasa Architecture
• Starting a Project with Rasa
• Rasa NLU
• Rasa Core
• Interactive Learning
• Channels

What’s a Chatbot?

Hype, Hype Everywhere

Chatbots Landscape

Disclaimer: Manage The Expectations

Rasa Framework

• Rasa is an open source machine learning framework for automated text and voice-based conversations. Understand messages, hold conversations, and connect to messaging channels and APIs.�
• Backed by a company of the same name�
• Mature Community:
• 500K+ Downloads
• 3500K+ Forum Members
• 300+ Contributors�
• Similar to vendors systems like DialogFlow (Google), Amazon Lex or Wit.ai (Facebook)

Why Rasa?

• Open Source
• Customizable
• Community
• Tune models for your use case�
• Own your data�
• Avoid Vendor Lock-In�
• Work locally

Demo

Rasa Architecture

Starting a Project with Rasa

• rasa init

​

Rasa NLU

• NLU stands for Natural Language Understanding�
• Understand User Request (Intent)�
• Converts user input text into a Data Structure, including extracted features�
• Entities: most common features, but can be more�
• NLU process can be rule based or ML based. Rasa is ML based

​

Rasa NLU: Select a Pipeline

• Rasa provides two “macros”

​

pipeline: "pretrained_embeddings_spacy"

pipeline: "supervised_embeddings"

Rasa NLU: Select a Pipeline

Rasa NLU: Intents

## intent:saludo

- hola

- buenas

- hola, como estas?

- buenos días

- buenas tardes

- buenas noches

- que pasa

- saludos

## intent:a_que_hora

- a qué hora es la charla de [Rafa Haro](speaker)

- cuando es la charla de [Antonio David Perez](speaker)

- a qué hora presenta [María Marcos](speaker)

- cuándo es la charla de [Guillem Duran](speaker)

- cuando habla [Alberto](speaker)

- a que hora es la charla de [antonio](speaker)

- a qué hora es la presentación de [Orange](speaker)

- cuando presenta [Ravenpack](speaker)

- a que hora es la charla de [irene](speaker)

- cuando es la charla de [david garcia](speaker)

## intent:que_hay_a_las

- que charlas hay a las [11:30](time)?

- que hay a las [12](time)?

- que charlas hay [hoy](day) a las [17:00](time)

- que hay [mañana](day) a las [14:30](time)

- que charlas hay el [sabado](day) a las [11:00](time)

- charlas el [domingo](day) a la [13:00](time)?

- quién presenta el [sabado](day) a las [17](time)?

- que charlas hay el [sabado](day)

- charlas del [domingo](day)

- [hoy](day)

- [sabado](day)

- el [domingo](day)

- [12:10](time)

- [17](time)

- [1](time)

## regex:time

- ([01]?[0-9]|2[0-3])(:[0-5][0-9])?

​

## lookup:day

- hoy

- mañana

- el sabado

- el domingo

- el sábado

- sábado

- domingo

​

## lookup:speaker

data/speakers.txt

​

## lookup:talk

data/talks.txt

Rasa NLU: Input Processing

Rasa NLU: Entities

• Entity Extraction depends on the extractor configured in the Pipeline
• By default it trains a CRF NER
• Training data support lookup tables, regex and synonyms�

​

Rasa NLU: Entities

• You can build your own Entity Extractor tailored to your use case
• For example: talks title extraction:
• Dummy Frequency based extraction
• Remove Stopwords
• Normalize
• Tokenize the input text
• Select the candidates with the higher number of tokens in common

​

{

‘Python para calentar tu casa’: [‘python’, ‘calentar’, ‘casa’],

​

‘MI CASA: venturas y desventuras de una radioastrónoma en Python’: [‘casa’, ‘venturas’, ‘desventuras’, ‘radioastronoma’, ‘python’]

​

}

charla sobre calentar mi casa

2

1

Rasa NLU Demo

Rasa Core: Stories

• Conversations Training data for the chatbot�
• Rasa will decide the next action depending of this training data. LSTM Models �
• Stories are made by a combination of Intents, Actions, Templates, Slots and Forms�
• Try to reflect as much conversations’ paths as possible (Interactive Learning can help)�

​

Rasa Core: Stories

## say hello

* saludo

- action_hello

​

## say goodbye

* despedida

- utter_despedida

- action_restart

​

## estoy bien path

* como_estas

- utter_estoy_bien

- utter_como_estas

* estoy_bien

- utter_me_alegro

- utter_te_puedo_ayudar

templates:

utter_saludo:

- text: "Qué pasa pisha?"

- text: "Qué pasa cabesa?"

- text: "Qué pasa quillo?"

​

utter_despedida:

- text: "venga nos vemos"

​

utter_estoy_bien:

- text: "yo estoy de arte"

- text: "aquí estamos chatboteando"

- text: "de categoría"

​

utter_como_estas:

- text: "¿y tu como andas?"

- text: "¿y tu que te cuentas?"

​

utter_me_alegro:

- text: "me alegro"

- text: "wena esa"

​

utter_te_puedo_ayudar:

- text: "¿qué te hace falta?"

- text: "¿qué necesitas?"

- text: "¿en qué te puedo ayudar?"

​

Rasa Core: Flow Diagram Example

Hola cómo estás?

RASA NLU

intent: como_estas, 95%

entities: []

​

RASA CORE

next_action: utter_estoy_bien, 100%

slots: []

​

Aquí estamos chatboteando

utter_estoy_bien:

- text: "yo estoy de arte"

- text: "aquí estamos chatboteando"

- text: "de categoría"

RASA CORE

## estoy bien path

* como_estas

- utter_estoy_bien

- utter_como_estas

* estoy_bien

- utter_me_alegro

- utter_te_puedo_ayudar

Rasa Core: Slots

• Key - Value store for your bot. It’s the bot memory�
• Can be used to store data gathered from the user (entities) or from the outside world (API, database…)�
• Slots also influence next action decision�
• Actions can fill slots and return Slot events

Rasa Core: Slots

## a que hora es la charla

* a_que_hora

- slot{"speaker":"Rafa Haro"}

- action_find_talk

- action_listen

## que hay a las sin hora

* que_hay_a_las

- action_find_talks_by_time{"day":"domingo"}

- form{"name": "talk_form"}

- action_find_talks_by_time{"time":"11:30","day":"domingo"}

Stories must reflect where the declared slots are filled

Rasa Core: Slots Example

Rasa Core: Actions

• Actions are the things your bot runs in response to user input�
• Four kinds:
• Utterance actions: start with utter_ and send a specific message to the user
• Retrieval actions: start with respond_ and send a message selected by a retrieval model
• Custom actions: run arbitrary code and send any number of messages (or none)
• Default actions: e.g. action_listen, action_restart, action_default_fallback�
• Custom Actions: backend query when the required slots are filled (Example: look for a talk when the bot knows time and day)�

​

Rasa Core: Actions

class ActionFindNextTalks(Action):

​

def name(self) -> Text:

return "action_find_next_talks"

​

def run(self, dispatcher: CollectingDispatcher,

tracker: Tracker,

domain: Dict[Text, Any]) -> List[Dict[Text, Any]]:

​

now = datetime.now()

search = Talk.search()

query = search.sort('start').query('range', start={'gt': now})

if query.count() == 0:

dispatcher.utter_message("Pues ahora mismo no encuentro ninguna")

else:

message = "Las próximas charlas que he encontrado son:\n"

message += range_query_to_message(query)

dispatcher.utter_message(message)

​

return []

Rasa Core: Events

• Actions are executed in a separated server. As a result of an action execution, a set of events are returned�
• Types:
• SlotSet (Fill a set of slots)
• Restarted (Restart a conversation)
• AllSlotsReset
• ReminderScheduled (Schedule another action….Ex: Are you there?)
• ConversationPaused
• FollowupAction (force next action, not predicted)
• Utter Message
• …..

​

Rasa Core: Forms

• Interactive Slot Filling Strategy�
• Single action which contains the logic to loop over the required slots and ask the user for information�
• The requested slots must be defined and the bot will keep going asking for them until all are filled. Unhappy paths can still be managed�
• The message to ask for the next slot can be customized using utter_ask_{slot_name} templates

​

Rasa Core: Forms

Rasa Core: Forms

• The way to fill the slot from user message can be customized:
• By default, from entities with the same name
• Directly from text (input text is the slot value)
• From an intent (if an intent is detected, fill the slot with a particular predefined value). This is useful for affirmations and negations
• From a particular entity with a particular name�
• Forms are active until filling all slots or programmatically deactivated. BE CAREFUL!�
• On one hand, it could give the impression of having “memory” but, on the other, it could enter into an annoying loop

​

Rasa Workflow

• rasa init
• rasa train
• rasa shell (--debug)
• rasa interactive�
• rasa run actions
• rasa run

​

Rasa Interactive Demo

Rasa Channels

• Rasa can ‘easily’ be used from several bot channels like Telegram, Slack, etc...

​

HTTPS

setWebhook(https://your_url.com/..)

ProxyPass

​

GRACIAS!

@adperezmorales

@rafa_haro

adperezmorales

rafa_haro