Alexa, Let's Chat
This week James Vlahos published a piece in Wired offering a behind the scenes look at the Alexa Prize and those who participated. Inside the Alexa Prize is both a great read and a good non-techy introduction into the challenges of designing truly conversational machines.
When I say “truly conversational” I mean machines that you can talk to in nearly the same way that you would talk to another person. The goal of the Alexa Prize competition was to design an Alexa skill that would engage a typical Alexa user in chit chat for 20 minutes. Wow. That’s a really long time to keep someone’s interest--even in human-to-human chit chat.
I tried some of these “social bots” during the competition by saying “Alexa, let’s chat.” I found the resulting dialogue annoying and rarely stayed in the skill for more than a minute. But as the competition continued, the universities involved continued to tweak their skills to make them more engaging.
Vlahos describes the two primary approaches to building a conversational “chatbot.”
Handcrafted Rules-Based Dialogue
The standard way of building a conversational chatbot is by handscripting prompts and responses to common conversational topics, such as discussions about the weather, books, movies, and so forth. Ideally, the AI can gently nudge the person they’re talking to down a specific conversational path, for example by asking about their favorite movie, book genre, or hobby. That path is then structured to enable the AI to choose from a database of pre-scripted responses and further prompts, all of which more or less sound like they are on topic.
The weakness of this handcrafted approach is that it can very easily go off the rails. If the user doesn’t bite at the suggested topic, the AI has to exit ungracefully and try something else. It’s impossible it anticipate everything a person might say. Most handcrafted AI conversations quickly stumble.
Machine Learning Generated Dialogue
These days handcrafted dialogue is generally seen as old-fashioned and the preference is given to generating responses on the fly using machine learning algorithms. Just as speech recognition needs to be trained on actual human audio files, automatically generated dialogue needs to be trained on many samples of what humans actually say.
At the crux of machine learning is reliance on user-generated texts available on social media sites. A favorite goldmine of user exchanges are Reddit forums. But how do you piece all these random comments into convincing dialogue? It’s not so easy. The other problem is that, well, it turns out that people really aren’t that interesting in what they say. Plus, unedited user comments from the web often are not appropriate for general consumption.
According to Vlahos, the team that won the competition (University of Washington) used a combination of handcrafted responses and machine learning algorithms. Their socialbot was able to engage at least one judge in a conversation that nearly reached the 20 minute threshold.
What Does the Future Hold?
I often hear people say they have no desire for a future where people talk to machines. They would rather people continue to talk to people. I don’t think anyone in the conversational AI space is hoping for a world where people stop talking to one another and only talk to their voice assistants.
On the other hand, people are wired up to converse. It’s the way we learn about ourselves, form opinions, develop concepts, gain and share knowledge, and so much more. What if AI-enabled voice assistants were equipped with smart conversational abilities that helped them entertain and engage, offer insights on educational topics, pique our curiosity to learn, or maybe help us see a topic from a differing perspective? It seems like a goal worth pursuing.
The future is wide open. By experimenting with an open mind, we might be able to turn the next generation of AI machines into partners that enrich our lives. The Alexa Prize was a fun start. More is sure to come.
