Response generation is how virtual assistants construct a response based on the request given.

A response in relation to action performed can look like this:

Some traits for a virtual assistant:

Purpose for response:

VA or Chatbot

Virtual Assistants

This is software that can assist people with basic tasks, usually using natural language. Intelligent personal assistants can go online and search for an answer to a user's question. Either text or voice can trigger an action.

Chatbots

Text is the main way to get assistance from a chatbot. Chatbots can simulate a conversation with a human user. Many companies use them in the customer service sector to answer basic questions and connect with a live person if necessary

Conversational Design

8 principles:

• cooperative
  • Paul Grice believed that there's no conversation without cooperation. It means your chatbot can support a customer only if it cooperates and provides information the user is looking for.
• goal-oriented
  • 
• context-aware
  • Create personalized chatbot greetings. They'll pop up in various conditions, based on the time a user spends on a page, a page address, or a referring source.
  • You can create a different greeting for users who visit your website for the first time and for those who are recurring visitors.
  • Personalization boosts engagement and improves user experience.
• quick and clear
  • While creating your chatbot stories, try to avoid complex metaphors, idioms, and long ambiguous statements. They slow down the conversation and take users away from where they need to be.
  • Use simple language, don't ask a user to choose many things at once, and get to the point
• turn-based
• truthful
• polite
• error tolerant
  • To make your chatbot error-tolerant, try to anticipate common spelling mistakes first. Teach your chatbot these variations to improve its understanding

Gricean's Maxims

Text Generation from Structured Data

Applications of Natural Language Generation:

• Analysis of business intelligence dashboards
• Reporting on business data/data analysis
• IoT device status and maintenance reporting
• Individual client financial portfolio summaries and updates
• Personalized customer communication

Stages of NLG process

1. Document planning

• content planning, document outline

2. Microplanning

• referring expressions, word choice, aggregation

3. Realisation

• converting specifications to a real text

Current trend

Sequence-to-Sequence Models (SEQ2SEQ models)

Evaluation

NLG: Structure Data to Text

Input:

• Mona Lisa, is-in, Louvre

Output:

• Mona Lisa is located in le Louvre

• Le Louvre is home to the Mona Lisa

• The Mona Lisa can be found in Le Louvre

• Le Louvre contains the Mona Lisa

• Metrics (automatic evaluation)

  • Comparison between some references and the predictions
  • Zero-cost beyond the gathering of train-dev-test data

• Human evaluation

  • Analysis of predictions
  • Very costly
  • Requires InterAnnotator Agreement analysis as well

Word-Based Metrics

• TER, BLEU, ROUGE, NIST, LEPOR, CIDER, METEOR
• Grammar-Based metrics
  • Readability
    • Word per sentence, syllables per sentence
  • Grammaticality
    • Mispellings, parsing score

Human Evaluation

• Informativeness
  • Does the utterance provide all the useful information from the meaning representation?
• Naturalness
  • Could the utterance have been produced by a native speaker?
• Quality
  • How do you judge the overall quality of the utterance in terms of its grammatical correctness and fluency?

Virtual Assistants

Virtual Assistant (or Conversational Agent)

• It's capacity (what it can do)

• It's perceived value by customers/users

• Engagement

  • How the user is interested in the conversation?

• Coherence

  • A coherent response indicates a comprehensible and relevant response to a user's request.
  • Measure number of incoherent responses/total utterances

• Domain Coverage

  • Coverage of agent within a domain
  • Goal-oriented (chatbots)
  • General domain (VAs) - average on different domains

• Conversational Depth

  • Number of turns that are possible (on same topic)

• Topical Diversity / Conversational Breadth

  • For general domain agents, how the agent is able to chat on various domains