Response Approaches

Simple Rules-based Approach

(defineRule startsWith "what"
  and() contains "bot|virtual agent"
  and() contains "is"
  caseInsensitive()
  build(),
  """
    |A Bot, otherwise known as a chatbot, is an application that operates as a user in a messaging app. It can
    |receive and respond to messages using natural language.
  """.stripMargin
  )

NLP-based Information Retrieval Approach

Deep Learning Approach

• Word embeddings are a geometric way to capture the “meaning” of a word via a low-dimensional vector.
• A property of embeddings is cosine similarity.
• Analogy relationships can be solved using linear algebra, e.g. man is-to woman, as king is to ___ ?

Future State - Generative Approach

For example.

Human: connection refused or something like that

Machine: may i know the version of network connect you connect?

Human: i am not sure i know that

Machine: is the network connect prompting for any user certificate?

Human: yes

Machine: are you at home or at work at the moment? office?

Human: i am at home

Machine: try this. goto page and enter your name and password

Human: done

Machine: try logging in with and check if its working fine now

Human: yes, now it works!

Machine: great. anything else that i can help?

Taken from Deep Learning model self-trained on chat logs between Google employees and Google’s Tech Support team

• The machine doesn’t actually understand anything about networking. It is pattern matching on sequence of words in the training responses that match patterns in the input.
• Does it matter? As long as issues can be solved without having to create pre-canned responses for every possible combination of user question.
• The acceptance threshold will be crossed when responses have a 99% accuracy in content and grammar.
• The acceptance threshold will be exceeded when responses can change in tone and personality to match the customer’s preferences and sentiment.