Enhancing your home assistant: we make Jarvis work!

The market for smart speakers is heating up: in 2014, Amazon quietly released the Echo, a 9-inch tall cylinder speaker controlled by a cloud-based voice assistant that goes by the name Alexa. Following the Amazon Echo’s popularity, Alphabet released the Google Home in late 2015. At the end of December 2017, Apple will also be joining the fray with the Siri-powered HomePod. And there is one more to come, Microsoft also announced Invoke, a Cortana-powered speaker (August 2017).

As a home device, one of its main purposes will be automating typical domestic tasks, like turning on lights. While at first sight this may look like a handful of operations, at Bitext we have identified more than 500 types of actions, more than 100 types of devices, plus a number of contextual features like time or space: from toasting bread at 7 am to stopping sprinklers at noon. Of course, the list can only be expected to grow.

Anticipating this widespread adoption, at Bitext we have been working on a piece of middleware to help tackle this situation and allow home assistants to understand user requests with high accuracy. The purpose of this middleware is to automatically detect intent (as actions are also referred to) and entities (or devices). It's a full implementation of the natural language processing component used in Jarvis. Let’s see a practical example:

For a command like: turn on the lights in the living room we can find different variations like:

  • turn on the living room lights
  • I’d like to turn on the lights in the living room?
  • can you turn on the living room lights?
  • please, turn on the living room lights

Bitext normalizes the examples above into two types of output to solve two specific problems in the assistant lifecycle: training and live use.

Output 1. Intent rewriting for bot or assistant training:

intent rewriting for bot training.png

This type of output shows the relevant entities that any assistant should be able to identify in order to properly understand its users’ requests. Intent rewriting allows to automatically tag sentences, avoiding the long and boring job of manually tagging the millions of sentences of your training dataset.

Output 2. Sentence rewriting for bot or assistant live use:

sentence rewriting for bot live use.png 

Sentence rewriting shows a simplified version of the original sentence. There are unlimited combinations of words and phrases to ask for something. The assistant, to work well, must be able to understand all of them and return to a common concept, but it’s a very complex task. The Bitext sentence rewriting analysis bypasses this process by removing all the superfluous information and mapping all those variants into a single, simple sentence.

Both solutions will help automate and speed up the production of intelligent and reliable home assistants.

Another key feature is that this middleware is compatible with any of the existing platforms: Amazon Echo, Google Home, Apple HomePod, Microsoft Invoke.

Also, as a home device, working in the language of the home owners is critical in this market.

Since the middleware is based on Bitext Framework for Text Analysis, the portability of the middleware to different languages (and language variants like UK English or Norwegian Bokmal and Nynorsk) is granted.




Subscribe Here!