Inspiration

Why spending a lot of time in resolving the tickets??? When Machine Learning is here:) One of the major place where the agents spend a lot of time relies in reading the tickets. What if we use our Machine Learning model to summarise all the text present in the ticket body?? We strongly believe that it would reduce the resolution time of the agents.

What it does

Our machine learning technique would roll out a summary of each conversation in the text(every note or reply) . The text could be from any one - either the agent, or the end-user. No worries- our machine learning algorithm will roll out a summary for every conversation. We would also give out the overall summary of the entire conversation. If the agent likes the answer by our algorithm ,he can use the summary and update the ticket, or he could write out his own summary. Remember humans are intelligent than machines . :)

How I built it

We have trained an tensorflow model using CNN/Daily mail dataset- which contains millions of data. Since we faced a lot of challenges in the loading of the final ML model.(which requires a lot of processing) in marketplace., We have created an flask API which would load the model and serve requests and give out predictions for the text. We have hosted this external service using Ngrok to serve requests from the market place app. Coming to UI and UX, We have used technologies like javascript,html and cssto build the app. We have added an exclusive timeline for agents to know when the conversation has started and when an new event is updated in the conversation.

Challenges I ran into

Training the model - Had to use CPU to finish the training, would be better if we had GPU.
Dataset - We didn't have freshdesk data to train the model and give the real time summary.

Accomplishments that I'm proud of

Was able to train a tensor-flow model without the support of GPU I am sure that if we fine tune the model with a better Dataset and with a support of GPU it would create a moment of WOW for everyone. Was able to build something that would add a lot of value to the business and save a lot of time. Cleaning the dataset.

What I learned

How word embeddings works. How Freshdesk marketplace works. Tensorflow serving

What's next for Text Summarization Using Machine Learning - Data Pirates

We are planning to fine tune the model using our freshdesk data which would result in better accuracy and results.

Share this project:
×

Updates