Orian Sharoni &
Nataly Shrits

Poster – Intent based search engine suggestions API using Tensorflow, clustering and kubernetes.
Soluto

avatar

Orian Sharoni &
Nataly Shrits

Poster – Intent based search engine suggestions API using Tensorflow, clustering and kubernetes.
Soluto

avatar

Bio

Nataly is a senior engineer @ Soluto
Orian is a DS @ Soluto

Bio

Nataly is a senior engineer @ Soluto
Orian is a DS @ Soluto

Abstract

At Soluto we have a chat platform for users that need technical support.

The project’s aim was to help the chat users to better express their needs. In order to do so, we used Tensorflow Universal Sentence Encoder to cluster our chat support applications into “families” of common issues. In each cluster, the question nearest to the center of the cluster was chosen as a representative of the entire cluster.

 

Nowadays, when a user first types a question into the service, the input is being processed into a vector, and the nearest cluster is located using cosine similarity. As a result, the first 3 cluster centers are being recommended for an opening sentence.

 

The feature is one of many micro-services the app uses, which are deployed using kubernetes, takes less than 0.5 seconds to process, and has caused more users to use the chat as they now can articulate their intent better.

Abstract

At Soluto we have a chat platform for users that need technical support.

The project’s aim was to help the chat users to better express their needs. In order to do so, we used Tensorflow Universal Sentence Encoder to cluster our chat support applications into “families” of common issues. In each cluster, the question nearest to the center of the cluster was chosen as a representative of the entire cluster.

 

Nowadays, when a user first types a question into the service, the input is being processed into a vector, and the nearest cluster is located using cosine similarity. As a result, the first 3 cluster centers are being recommended for an opening sentence.

 

The feature is one of many micro-services the app uses, which are deployed using kubernetes, takes less than 0.5 seconds to process, and has caused more users to use the chat as they now can articulate their intent better.