CASE STUDY 01
- Professional
- NLP
- Machine Learning
- Data Engineering
Turning customer reviews into insights
A multi-label NLP pipeline, from model to production.
Tressol-Chabrier · Work-study placement
The project, step by step
38,000customer reviews processed
- 01
Customer reviews
Google My Business feedback
- 02
Classify reviews
CamemBERT adapted to the business taxonomy
- 03
Assign topics
Multiple topics per review and a vehicle type
Context
At Groupe Tressol-Chabrier, Google My Business reviews provide a source of customer feedback to analyse.
Problem
Classify reviews using a business taxonomy with two dimensions: ten topics and vehicle type. A review can cover multiple topics.
Data
38,000 Google My Business customer reviews processed by the production pipeline.
Approach
I designed the multi-label classification pipeline and fine-tuned CamemBERT to adapt it to the business taxonomy.
Method
Fine-tuning, F1 evaluation on the main topics, then production deployment of the processing pipeline. Dataset splits and training parameters remain to be documented.
Results
The F1 score increased from 0.83 to 0.93 on the main topics. The pipeline was put into production to process 38,000 customer reviews.
Technologies
CamemBERT and multi-label NLP classification.
Lessons and limitations
The result applies to the main topics and does not describe the performance of every class. Analysis of remaining errors and generalisation limits is still to be documented.