AI Recommendations - How it works
In our previous blog in the series, we had discussed how AI Recommendations leverages Artificial Intelligence to help businesses understand and learn from conversational data. These actionable insights help in -
1. Battling Industry-wide decline in feedback and survey collection
2. Reducing human errors and biases
3. Improving Product Offerings
4. Discovering Out of Scope Intents
5. Reducing manual effort and time required
In today’s blog, we will dive deeper into the technology that goes behind the success of AI Recommendations.
The AI Recommendations module is based on multiple Deep Learning Algorithms. The purpose of the module is to enable faster and more effective maintenance of the IVA, once it has been released to real customers. The AI Recommendations module primarily analyses those customer messages, that the IVA answered with very low confidence or couldn’t answer at all. Since the AI recommendation module is trained on the Training Data present in the IVA, the module is effectively able to categorise the customer messages into 3 categories -
1. Messages that can be catered from existing intents in the IVA
The module applies deep learning-based sentence pair classification to determine the existing intents in the IVA, to which the above messages can be added. Sentence pair classification allows the module to find the closest training utterances that match the above messages.
In this section of AI Recommendations, accuracy in the range of 80 to 85% has been seen, while mapping low confidence customer queries to existing intents in the IVA.
2. Messages that can be catered by creating new intents in the IVA
The module applies deep learning-based unsupervised clustering to group messages, which are similar in meaning. An analyst can have a look at such clusters, and proceed to create new intents on the IVA. The above messages can be used as training data for the newly created intent as well.
In this section of AI Recommendations, an accuracy of 70% has been seen in the new clusters suggested by the module i.e 7 out of 10 new intent clusters suggested by AI recommendations can be added as intents to the IVA.
3. Messages that are Discarded - duplicate, profane, or are Gibberish, etc.
These messages are discarded by the module, as these messages don’t reflect any actionable customer intent. Example:
This becomes even more crucial as brands keep growing over time with the no. of services, facilities, offers, products and services that get added to its suite of company offerings. As per the changing needs of customers and services provided by the company, even the customer services have to be optimized.
With AI Recommendations, you will be able to maintain and improve the IVA experience more efficiently by replacing hours of manual efforts with AI-driven insights.
We cannot wait for you to use the feature! Reach out to me at email@example.com in case of any questions or feedback! =)