2.3 So, what does Allison do in the tech industry?

As Allison explained, the tech industry has multiple career paths in Spanish. What, exactly, does her job involve?

Transcript: “In my full-time job I’m a conversation design engineer–I work with a tech company as a contractor. I work on their chatbots, making sure that the chatbots perform appropriately and correctly in Spanish, primarily, and in English, when they need me to.”

Most of us have probably interacted with chatbots, like the kind Allison trains, without even knowing it. Let’s look at some examples to get a clearer sense of what her job entails.

Have you ever logged onto a website—perhaps your bank, or cell service provider—and been greeted by someone in a little pop-up chat bubble? It is likely that this customer service agent is not human and has been created by a conversation design engineer.

There are two sides to every well-functioning chatbot: the dialog design itself, and the language model that uses language data to interpret human utterances.

The dialog design is exactly what it sounds like; it is the behind-the-scenes configuration of the conversations a chatbot can have. Generally speaking, they look like flowcharts that a user traverses depending on their responses to the bot. Consider the following chat example from an imaginary sock retailer:

To the untrained eye, this conversation looks like a simple dialog between a user and a virtual assistant. In reality, there is likely a massive flow chart operating in the background that interprets the user’s language and determines how the customer should proceed through the conversational flow. It might look something like this:

This is an oversimplified visualization of a dialog design, but you can see that the bot delivers different content depending on the customer request. How the customer traverses the dialog is determined by rules, logic, synonym sets, timers, and other entities that provide conversational context.

Now, how in the world does the bot know exactly how to route a customer through a dialog flow chart?

This is where we rely on the other side of the chatbot—the language model. While many companies are beginning to rely on Generative AI, which involves unsupervised machine learning (in which we feed a bunch of data into a computer and allow it to detect its own patterns and draw its own conclusions from what it finds) many language models are still using supervised learning. In supervised learning, models are fed data in a way that is already organized and labeled. For example, we might provide a set of 100 sample utterances that a customer might use when they want to purchase something:

• “I’d like to buy some new socks”
• “I wanna buy new socks”
• “Buy socks”
• “My grandma’s birthday is next week. I want to buy her some fun socks”

With enough sample utterances, the model should recognize the utterance in the same chat “I want to buy new socks” as one that should be sent down the PURCHASE dialog flow. There would be a set of training data for order troubleshoot_order in our example, too.

In order to ensure the model is predicting customer utterances correctly, the conversation designers perform routine evaluations of its predictions to identify any gaps in understanding.

One example of a bot misunderstanding from Allison’s current job involved the Spanish language model. In many regional varieties of Spanish, customers use the word “renovar” or “renovación” to mean “to trade in” or “trade-in,” respectively. However, when people would chat in and use the word “renovar” or any derivative of it, the bot was not providing information about the trade-in program. Rather, it was providing information on how to buy refurbished products, or productos renovados.

Allison did an analysis of the existing training data and found that there was not enough data that included the words “renovar” or “renovación” to signal that the customer was asking about the trade-in process. Her team quickly added several phrases with which to train the model, and the gap in understanding disappeared.