Skip to main content

Chatbot / chatops

https://www.freecodecamp.org/news/how-to-build-an-ai-chatbot-with-redis-python-and-gpt

Rasa​

Open source machine learning tools for developers to build, improve, and deploy text-and voice-based chatbots and assistants

  • The core of building a Rasa assistant is providing examples that your system learns from
    • How do people say things?
    • How do conversations go?

Types of chatbot​

  1. Rule-based Chatbots / Workflow chatbots
  2. Intellectually Independent Chatbots
  3. AI-powered Chatbots

Types of product chatbots​

Customer Service Bots​

  1. Welcome Bots
  2. Resolution Bots
  3. Navigation Bots

E-Commerce Bots​

  1. Sales Bots
  2. Product Recommendation Bots
  3. Billing Bots
  4. Order Management Bots

https://www.freshworks.com/live-chat-software/chatbots/chatbot-ideas-blog

Task Oriented Dialogue System​

  • Task Oriented - Users want to accomplish something (task they want to achieve)
  • Dialogue system - Talking with an automated system in a two-way conversation
  • Chit chat bot - Goal is to have an ongoing conversation without necessarily doing anything

image

State machines vs neural methods​

  1. Understanding Text

    • NLU - raw text in, machine-readable information out
    • Rule based
      • Example: A regular rexpression that finds and extracts email addresses
      • You don't need much/any data for these approaches
      • They can run very fast
      • But they're not very good at handling things they haven't seen before
    • Neural
      • Example: A transformer-based model (e.g. DIET) that sorts text into intents based on examples it's been provided
      • Require training examples (in general, the more the better)
      • But they're very good at handling things they haven't seen before, making "informed guesses"

    A Rasa assistant might use both, but the neural methods are the core of the framework

  2. Deciding what do do next

    • Dialog policy - Given the conversation so far, what should your assistant say or do next?
    • Rule-based
      • Example: A big dialog tree of all possible paths a conversation can take
      • This is a traditional approach to dialog and it can work well in some cases
      • But they can't hand digressions and are a bit of a pain to extend/maintain
    • Neural
      • A transformer-based model (e.g. TED) that picks the next best turn based on the conversation so far and all the conversations it's been trained on
      • Require training examples (in general, the more the better)
      • But this approach let's users have a more natural conversations, even if they say things in a different order every time

Both 1 and 2 is recommended to use in tandem

How to make sure conversations work (and improve over time)

  • The Rasa approach is
    • Flexible, every conversation might be unique
    • Works better the more high-quality training data you have
  • How do you make sure conversations work?
    • Manually reviewing and annotating conversations
  • How do you make sure conversations improve over time?
    • Correct any errors your assistant made in a conversation, then add it to the training data, retrain and redeploy
    • A Rasa conversational AI ystem itsn't static, It changes and adapts over time
  • We call this process "conversation-driven development"

Files​

  • domain.yml
  • config.yml
  • Data files
    • nlu.yml
    • stories.yml
    • rules.yml

Commands​

pip3 install --upgrade rasa
rasa init
rasa train
rasa shell
rasa shell --debug
rasa -h

Domain.yml​

The domain file is a directory of everything our assistant "knows"

  • Responses: These are the things the assistant can say to users
  • Intents: These are categories of things users say
    • affirm
    • deny
    • greet
    • thankyou
    • goodbye
    • search_concerts
    • search_venues
    • compare_reviews
    • bot_challenge
    • how_to_get_started
  • Slots: These are variables remembered over the course of a conversation
  • Entities: These are pieces of information extracted from incoming text
  • Forms & actions: These add application logic and extend what your assistant can do

Data​

  • The text data used to pretrain any models or features you're using (e.g. language models,, word embeddings, etc.)
  • User-generated text
  • Patterns of conversations
  • Examples:
    • Customer support logs
    • User conversations with your assistant
  • How should conversations with your chatbot go?
    • Stories
    • Rules
  • How do users say things?
    • Intents
  • If you have data
    • Modified content analysis
      • Go through data (or sample) by hand and assign each datapoint to a group
      • If no existing group fits, add a new one
      • At given intervals, go through your groups and combine or separate them as needed
      • Start with 2-3 passes through your dataset
    • Can't you just automate this i.e. intent discovery and classification
      • It's easiest to use interactive learning to create stories
  • If you don't have data
    • Start with the most common intent
    • Start with the smallest possible number of intents (that conver your core use case)
    • Everything else goes in and out of scope intent
      • If your assistant can't handle something, give users an escape hatch right away
    • Additional intents will come from user data
  • Have fewer intents
    • Rasa style CDD
      • You only need to start with the most popular, important intents & a way to handle things outside them
      • Continue to build from there if that's what users need
    • Human reasons
      • More intents = more training data, maintenance, documentation
      • More intents = annotation more difficult
    • ML reasons
      • Transformer classifiers scale linearly with the # of classes
      • Entity extraction (esp with very lightweight rule-baed system like Duckling) is often faster
  • Paring down intents
    • Don't use intents as a way to store information (use slots instead)
    • Consider to combine, if a lot of same tokens show up in training data for two intents
    • Example

image

  • Training data for an intent
    • User-generated > synthetic
    • Each utterance should unambigously match to a single intent
      • You can verify this using human sorting & inter-rater reliability
    • Is an utterance ambiguous?
      • Use end-to-end instead (the raw text as training data w/out classifying it)

Stories​

Training data to teach your assistant what it should do next

  • If you have conversational data, start with those patterns
  • Generate your own conversational patterns
    • It's easiest to use interactive learning to create stories
    • Start with common flows, happy paths
    • Then add common errors/digressions
  • Once your model is trained, add more data from user conversations

image

Or statements​

image

Checkpoints​

image

image

Rules​

A way to describe short pieces of conversations that always go the same way

  • User rules for one-off interactions (checking account balance, checking if this is a bot)
  • Don't use rules for multi-turn interactions
  • Don't use OR statements and checkpointing often
  • Don't write out every possible conversation flow start to finish
  • Don't delay user testing

Entities​

Entities are structured pieces of information inside of a user message

An entity can be any important detail that your assitant could use later in a conversation:

  • Numbers
  • Dates
  • Country names
  • Product names, etc

There are 3 ways entities can be extracted in Rasa

  1. Using pre-build models

    • Duckling for extracting numbers, dates, url, email addresses
    • SpaCy - for extracting names, product names, locations, etc

  2. Using Regex

    • For entities that match a specific pattern (e.g. phone numbers, postcodes, etc.)

  3. Using ML

    • For extracting custom entities

image

The output of the entity extraction is a snippet of JSON which contains the details of

  • Entity category ("city")
  • Entity value ("New York City")
  • Confidence levels
  • The component that extracted the entity

Synonyms​

Synonyms can be used to map the extracted values to a single standardized value

image

OR during entity extraction

image

Others​

Learning​

Demo​

http://financial-demo.rasa.com/guest/conversations/production/d379406177e14dcfa67ed47ebfdd9106

https://rasa.com/showcase/albert-heijn

Responses​

Rasa 3.0​

Solutions​

Chatbot strategy​

image