Myself and a friend are working on a startup that is initially focused on the Amazon Echo. We integrate with Salesforce to enable executives at a company to get info about their sales pipeline by asking questions such as
"Alexa, ask Cricket about Vegan Delight restaurants"
"Alexa, ask Cricket about Big Steak Grill restaurants"
"Alexa, ask Cricket how many sales we've made to Taco Town during the last month"
But Alexa is stumbling on a lot of real-world company names. For instance, we just did a demo for a company that might want to use our service, but they have customers such as:
Avon
Avon Brazil
Unilever
I could not get Alexa to see that "Avon Brazil" was different from "Avon".
My startup can not possibly know what companies our customers will have as customers (that is, we can not know the names of our customers customers). So, taking into account the limitations that we face, is there anything I can do to improve our chances at getting the company names correct? That is, what can we do to improve the chance that Alexa will correctly identify them?
If I was only struggling against typos, then it would be trivial (using modern open source libraries such as the Stanford Natural Language Processing library) to calculate the Levenshtein Distance to the names of any company name that we find in the Salesforce database of our customers. But, of course, with Alexa we face a deeper problem. Instead of "Unilever" we get "You need liver". How do I calculate equivalent of a Levenshtein Distance to the phonetic sounds, especially when they've been morphed by Alexa?
If anyone is interested in background info on the startup, the thesis of the startup was stated here:
"The Amazon Echo will primarily succeed in business settings, not in home settings"
http://www.smashcompany.com/business/the-amazon-echo-will-primarily-succeed-in-business-settings-not-in-home-settings
1 Replies
15469 Views