Technology Talks | 15 Apr 2022

Text-Based Recommendation Engine to Improve Turnaround Time of Solving Quality Complaints

TVS Motor Company Ltd. (TVSM) successfully adopted the Total Productive Maintenance (TPM) Policy in June 1999 and it resulted in significant improvement in business performance & key performance indicators for Quality, Cost, Delivery, Productivity, Morale, and Safety. Product Performance Feedback Report (PPFR) falls under the pillar of “Quality Maintenance” where TVSM establishes a robust system to arrest product-related customer complaints.

Problem Scenario:

PPFR is a complaint report raised by the Dealer Mechanics of TVSM Dealerships in the event of the occurrence of any failure in the product to take immediate robust actions to solve the complaint and prevent the occurrence of the same in the future across all products.

Initially, Complaints are raised in the portal after manual consolidation where the inflow was around 80 complaints/month. In the process, registration itself took a long time due to manual consolidation. Hence, Dealer mechanics are allowed to raise the complaints so that the responsiveness will improve, and just as any new process change that makes work easier, eventually faces its roadblocks, the below issues act as a hurdle for the executives handling the complaints.

  1.   1. Multiple complaints are raised for the same incident and the inflow of complaints is peaking at 400 complaints/month and complaints are raised in regional languages. This becomes a bottleneck and caused a delay in providing solutions within the stipulated time. Hence not able to devote time to genuine complaints that need more attention.

  2.  2. For every complaint that is raised, an executive must manually search for a similar complaint in their database and suggest an appropriate solution for it within 24 hours. And with the large influx of complaints, this becomes exceedingly difficult to do.

Solution:

PPFR Solution is a recommender engine that has introduced automation to aid the executives who are working hard to ensure that dealers get solutions to the complaint as quickly as possible.

To solve the issue of duplicate complaints, the NLP model is used to identify duplicates and map them to the original complaints. Once the dealer mechanic enters the details to the fields, the top 10 recommendations are selected by a statistical approach and then filtered by TF-IDF (Elastic search) based retrieval. Finally, the text similarity-based model matches the duplicate complaint.

To solve the issue of manual searching of related complaints in the database and provide the solution, the Deep Learning model of Huggingface-finetuned-distilbert-base-uncased has been used on the complaint & resolution data to display the top 10 most probable complaints and their resolutions to the complaint entered by a mechanic.

 

The text data is broken into 4 segments and ranking is provided by the augmentation process. A REST API is created using Flask and integrated into the system to display the top 10 recommendations.

This gives dealer mechanics the ability to select the complaint from the recommendations that match their complaint & record it as an incident rather as a new complaint and it resolves the issue of creating duplicate entries into the system, this gives executive perspective on what the problem is and limits the potential solutions to the top 10 as opposed to 100s of solutions that are there in the system. Also, the mechanics are now aware of the stage of resolution to the complaint if it is already in the system and proceed with the actions to be taken as mentioned in it immediately rather than waiting for instruction. When the system receives a new complaint that is not available in the database immediately triggers the action to the complaint and eliminates the delay in registration.

The model has a performance accuracy of a whopping 96% in the top 10 recommendations which is reassuring for the executives because it gives them the confidence that the top 10 suggested resolutions are pretty much all they need to investigate. So, the recommendation engine has brought a perfect balance of automation and human judgment into decision-making in a system.

Name is required field Name should be Alphabet only Name can not be morethan 30 characters!
Email is required field. Email is required field. Email can not be morethan 50 characters!
Comment is required field. Comment can not be morethan 300 characters!