Chapter 11. Sequence-to-sequence models for chatbots

 

This chapter covers

  • Examining sequence-to-sequence architecture
  • Vector embedding of words
  • Implementing a chatbot by using real-world data

Talking to customer service over the phone is a burden for both the customer and the company. Service providers pay a good chunk of money to hire these customer service representatives, but what if it’s possible to automate most of this effort? Can we develop software to interface with customers through natural language?

The idea isn’t as farfetched as you might think. Chatbots are getting a lot of hype because of unprecedented developments in natural language processing using deep-learning techniques. Perhaps, given enough training data, a chatbot could learn to navigate the most commonly addressed customer problems through natural conversations. If the chatbot were truly efficient, it could not only save the company money by eliminating the need to hire representatives, but even accelerate the customer’s search for an answer.

In this chapter, you’ll build a chatbot by feeding a neural network thousands of examples of input and output sentences. Your training dataset is a pair of English utterances; for example, if you ask, “How are you?” the chatbot should respond, “Fine, thank you.”

11.1. Building on classification and RNNs

11.2. Seq2seq architecture

11.3. Vector representation of symbols

11.4. Putting it all together

11.5. Gathering dialogue data

11.6. Summary

sitemap