Classification models using deep learning can speed up the process to create chat bots.
One of the biggest banks in Latin America needed to automate its customer service processes through chat channels. So it decided to generate some bots that could attend customers using the Google Dialogflow platform.
The standard process of creating a bot involves that “someone” begins to select groups of questions or topics that customers ask the bank and from them to make the trees of attention of a bot with their respective answers.
The bots require “examples” of questions to work, but people ask in hundreds of different ways to get the same answer, so the systems must be trained with a set of phrases, hopefully broad, to allow AI systems to respond more accurately.
A challenge then is to know “what customers ask” and “how customers ask,” the workers of a Bank who are in the contact center know some frequently asked questions, but the diversity is enormous.
Here begins the first problem to create a bot in large companies. “What customers ask” and “How customers ask,” doing this work by hand can take months, and the teams in charge start to hate the bot before beginning to build it.
The bank then made a radically different decision, to use the data contained in the hundreds of thousands of real conversations with customers, where a human was answering and use Deep Talk’s deep learning models to know from real conversational data and not hypotheses, “What customers ask” and “How they ask.”
In just 1 hour, the bank already had the answers to the above questions and the training sentences needed for each question. But it also had the solutions delivered by the customer service executives to those questions!
Deep Talk separated and classified the customer messages into clusters up to 3 levels deep.
Example: Level 1: Account, card, credit transfer.
Level 2 of “Account”: Checking account, current account, demand account, savings account
Level 3 of “Current account”: Open account, close account, account security, an account password, etc.
The next step of building the bot took only a few days to get the first productive version working. Each organized cluster had the necessary training sentences to be loaded. The company also knew perfectly well how the conversations with customers were distributed, privileging to make a bot that covered at least 80% of the questions in its first version.
Something that could take 3–6 months in traditional ways of creating bots took just a couple of weeks using real conversational data with customers and using Deep Talk for analysis.