VaxVerdict

A ROBERTA BASED MULTILABEL CLASSIFIER

PICT CL Lab Group 1

Table of

Problem Statement

01

02

03

04

Approaches

Why roBERTa

Advantages

Disadvantages

05

06

Future Scope

CONTENTS

PROBLEM STATEMENT

The goal is to build an effective multi-label classifier to label a social media post (particularly, a tweet) according to the specific concern(s) towards vaccines as expressed by the author of the post.

​

Labels: Unnecessary, Mandatory, Pharma, Conspiracy, Political, Country, Rushed, Ingredients, Side-effect, Ineffective, Religious, None

Convolutional NN

macro f1 score: 0.55

BERT

macro f1 score: 0.6

roBERTa

macro f1 score: 0.65

Approaches

Why

• Bidirectional context modeling
• Pretrained on diverse multilingual datasets
• No Next Sentence Prediction
• Compatibility with CUDA architecture

ROBERTA

Effective Handling of Long Sequences

Dynamic Masking

Pretrained on Large Corpus

Large Batch Training

Advantages

Domain Specificity

Computational Resources

Large Memory Footprint

​

Lack of Incremental Learning

Disadvantages

FUTURE SCOPE

• Hyperparameter tuning
• Checkpoint incorporation and Dropout
• Data Augmentation
• Adversarial Training

PICT CL Lab Group 1

THANK YOU