Optimal usage of conversations received in a chatbot

Posted by axay on Monday March 20, 2017 @05:35AM
from the self-post dept.

We have created a bot in house without using any platforms like wit.ai or api.ai. We treat is as a text classification problem. We are using word2vec to vectorize the input documents and a linear svc to classify it into one of our pre defined classes.

Also we have a feedback dashboard where the admin can tag and validate incoming user queries with correct class names (similar to what existing platforms provide). This is fine until a certain level but becomes impossible later.

So now we have a few (5k) tagged queries and a lot (350k) untagged user queries. Is there any way we can use these to enhance our system ?

( 2 comments | on HN )

Re:Optimal usage of conversations received in a chatbot by artminister on Thursday March 23, 2017 @07:54AM (#13938964)

We used to use svc for classification in the early days of chatbots, but switched to tensorflow

We are using a bidirectional lstm model for sequence classification.

[ Parent | on HN ]

Re:Optimal usage of conversations received in a chatbot by sprobertson on Monday March 20, 2017 @01:12PM (#13915847)

Retrain your model with the new data, or use it to build a better model.

[ Parent | on HN ]