Text2topic Leverage reviews data for multi-label topics classification in Booking.com

NLP.IL

Author

Oren Bochman

Published

Tuesday, February 28, 2023

Text2topic Leverage reviews data for multi-label topics classification in Booking.com - Moran Beladev & Elina Frayerman

Abstract:

Having millions of customer reviews, we would like to better understand them and leverage this data for different use cases. For example, finding popular activities per destination, detecting popular facilities per property, allowing the users to filter reviews by specific topics, detecting violence in reviews and summarizing most discussed topics per property.

In this talk, we will present how we build a multilingual multi-label topic classification model that supports zero-shot, to match reviews with unseen users’ search topics.

We will show how fine-tuning BERT-like models on the tourism domain with a small dataset can outperform other pre-trained models and will share experiment results of different architectures.

Furthermore, we will present how we collected the data using an active learning approach and AWS Sagemaker ground truth tool, and we will show a short demo of the model with explainability using Streamlit.

Moran Beladev Bio:

Moran is a machine learning manager at booking.com, researching and developing computer vision and NLP models for the tourism domain. Moran is a Ph.D candidate in information systems engineering at Ben Gurion University, researching NLP aspects in temporal graphs. Previously worked as a Data Science Team Leader at Diagnostic Robotics, building ML solutions for the medical domain and NLP algorithms to extract clinical entities from medical visit summaries.

Slides

slide

slide

slide

slide

slide

slide

slide

slide

What is CIP

What is CIP

What is CIP

What is CIP

Text2Topic

Text2Topic

Overview

Overview

Data Sources

Data Sources

Data Sources

Data Sources

Data Sources

Data Sources

Data Sources

Data Sources

Data Sources

Data Sources

Motivation/Goals

Motivation/Goals

slide

slide

How it Works?

How it Works?

Cross Encoder architecture

Cross Encoder architecture

Cross Encoder architecture

Cross Encoder architecture

Bi-Encoder architecture

Bi-Encoder architecture

Bi-Encoder architecture

Bi-Encoder architecture

Bi-Encoder architecture

Bi-Encoder architecture

Bi-Encoder architecture

Bi-Encoder architecture

Bi-Encoder self-supervised

Bi-Encoder self-supervised

Main Differences

Main Differences

Dynamic Padding

Dynamic Padding

Dynamic Padding

Dynamic Padding

Dynamic Padding

Dynamic Padding

Evaluation

Evaluation

Results

Results

Metrics

Metrics

Results

Results
  • note Muse-large used as a baseline!

slide

slide

slide

slide

slide

slide

Well done! They did the experiment way past the point where the effects maxed. The main takeaway here is that 100 docs suffice for getting good results on a new topic.

slide

slide

slide

slide

slide

slide

Great talk - the padding tip is probably worth the price of admission :-)

Citation

BibTeX citation:
@online{bochman2023,
  author = {Bochman, Oren},
  title = {Text2topic {Leverage} Reviews Data for Multi-Label Topics
    Classification in {Booking.com}},
  date = {2023-02-28},
  url = {https://orenbochman.github.io/posts/2023/2023-02-28-NLP.IL-Booking.com/NLP-IL-Booking Text2Topic.html},
  langid = {en}
}
For attribution, please cite this work as:
Bochman, Oren. 2023. “Text2topic Leverage Reviews Data for Multi-Label Topics Classification in Booking.com.” February 28, 2023. https://orenbochman.github.io/posts/2023/2023-02-28-NLP.IL-Booking.com/NLP-IL-Booking Text2Topic.html.