Polarity and Subjective classification with Multitask Learning and BERT Embedding

Multi task learning often helps to improve the perfor-mance of related tasks. Related tasks often have inter-dependence on each other and perform better whensolved in a joint framework. In this paper, we presenta deep multi-task learning framework that jointly performs sentiment and subjective analysis. We propose an attention based multi task model for predicting the sentiment and subjectivity. The input sentences are transformed using pretrained BERT and Glove embeddings and the results depict that BERT embedding based model works better than Glove based model. We com-pare our approach with state of the art models in bothsubjective and sentiment classification in both singletask and multi task frameworks. Our results suggest that multitask learning framework offers improvement over the single task framework even with two different datasets.

Folder Structure:

├── Codes                       # Contains Python Notebooks (*.ipynb)
|   ├── Data Preprocessing         
│   ├── Bert Embeddings         
|   ├── Bert Model 
|   ├── Glove Model 
├── Dataset                   
│   ├── Polarity_data
|   |   ├── Bert Embeddings      # bert embeddings 
|   |   ├── sentiment.csv        # csv file 
|   |   ├── sentiment.pkl        # Glove embeddings 
│   ├── Subjective_data 
|   |   ├── Bert Embeddings      # bert embeddings 
|   |   ├── subjective.csv       # csv file 
|   |   ├── subjective.pkl       # Glove embeddings      
└──  Results                     

Dataset

1. Polarity Dataset: The dataset contains 5331 positive and 5331 negative processed sentences.We take 5000 from each class randomly.
2. Subjectivity Dataset: The dataset contains 5000 subjective and 5000 objective processed sentences.

Dataset	Train	Dev	Test	Max Length	Avg. Length	Vocab
POL	7.2K	800	2K	40	15	16.5k
SUBJ	7.2K	800	2K	85	17	18.5k

Model Architecture

Results

	Framework	Subjective	Polarity
Baselines	SenticNet 5 Subjective detector Multi-task uniform layer Multi-task shared-layer	- 92% 93.4% 94.1%	92.8% - 87.1% 87.9%
BERT Embedding	BILSTMpol BILSTMsubj MTLsharedNTN	- 93.5% 95.1%	77.5% - 94.6%
GLOVE Embedding	BILSTMsubj BILSTMpol MTLsharedNTN	90.7% - 92.3%	- 75.9% 92.1%

Prerequisites

!pip install numpy
!pip install sklearn
!pip install pandas
!pip install csv
!pip install scipy
!pip install keras
!pip install keras-self-attention

How to run the code

1. Install the prequisites mentioned above.
2. Run '/codes/Bert_model.ipynb'
3. Run '/codes/Glove_model.ipynb'

Note: The glove and bert embeddings of both the datasets are provided. Make sure you check the relative paths before running the code and change the file names while saving the model.

Citation

MDPI and ACS Style

Satapathy, R.; Pardeshi, S.R.; Cambria, E. Polarity and Subjectivity Detection with Multitask Learning and BERT Embedding. Future Internet 2022, 14, 191. https://doi.org/10.3390/fi14070191

AMA Style

Satapathy R, Pardeshi SR, Cambria E. Polarity and Subjectivity Detection with Multitask Learning and BERT Embedding. Future Internet. 2022; 14(7):191. https://doi.org/10.3390/fi14070191

Chicago/Turabian Style

Satapathy, Ranjan, Shweta R. Pardeshi, and Erik Cambria. 2022. "Polarity and Subjectivity Detection with Multitask Learning and BERT Embedding" Future Internet 14, no. 7: 191. https://doi.org/10.3390/fi14070191

Find our paper here

Link: https://www.mdpi.com/1999-5903/14/7/191