Overview / Usage

Consider the scenario of emergency responders. They receive a large number of distress calls and have to respond to them simultaneously. It would be very beneficial if they had a system which would receive a large number of audio requests and generate 'scores' in real time, ranking them based on 'importance'. Apart from training a text classifier to predict an intent and then a score, other parameters may also be considered such as geo, time of calling.,etc which would all be used to compute a score which would be useful in the case when multiple such requests are received. My system does precisely all the above listed actions. An additional purpose served is the screening of 'crank' calls, for which the score generated would be lesser than a user-defined threshold. Such a system would also serve useful where large number of request calls are received at call centres, to allow users to respond to the most 'urgent' calls first.

Methodology / Approach

First we train a text classifier to predict an intent. Although I used data specific to our company, other datasets are also available at Kaggle, which I have used (will elaborate soon). The number of intents are 30+. The intents are mapped to a certain score (denoting user defined importance to the intents). Then I built an Alexa skill kit using an AWS instance, to facilitate implementing a small server program on an AWS machine and routing any voice/audio queries to the API hosted on the AWS machine. The request is received by the API and invokes a function calling the text classifier (as a pickle file) to generate a score based on the parameters received, to compute a score.

Technologies Used

SciKit learn used to build a text classifier.
API hosted using Flask-Ask library built by Amazon.
Alexa Skill Kit used.
Python 3.5 used.

Other links

• Intelligent Messaging Assistant

Collaborators