You are on page 1of 1

POSTER ID: 109

Analyzing Bangla Text Sentiment through


Convolutional Neural Network
Imamul Hasan MD. Farhad Hossain
Department of Electronics and Telecommunication Engineering
Chittagong University of Engineering and Technology
Chittagong University of
Engineering and Technology

Abstract: RESULTS AND DISCUSSIONS:


This study aims to develop a Convolutional Neural Network (CNN) model for • The proposed evaluation metrics for the CNN model include accuracy, precision,
sentiment analysis in Bangla text. The proposed model is trained on a dataset of recall, and F1-score. Experimental findings show that the CNN model outperforms
Bangla text, which is preprocessed using tokenization, stopword removal, and traditional techniques, achieving higher accuracy (84.84%) and F1-score (73%). SVM,
stemming techniques. The CNN architecture is designed with multiple RF, KNN, and previous CNN implementations scored 19%, 25%, 22%, and 81%,
convolutional layers and pooling layers to extract features from the input text. The respectively, indicating the superiority of the proposed approach in sentiment
model is trained and evaluated using various performance metrics such as analysis for the dataset.
TABLE 02 : EXPERIMENT RESULT USING THE DATASET
accuracy, precision, recall, and F1-score. In this paper, we propose method of a TABLE 01 : HYPER PARAMETERS OF OUR PROPOSED DEEP CNN MODEL

lowcost Deep Learning Framework based on Convolutional Neural Network (CNN)


that analyzes sentiments from texts written in Bangla. In this paper, I propose
method of a low-cost Deep Learning Framework based on Convolutional Neural
Network (CNN) that analyzes sentiments from texts written in Bangla. The
TABLE 03 : OUTPUT SHAPE AND PARAMETER SIZE OF EACH LAYER OF THE
classifier model obtains a classification accuracy of 84.84%, which is almost 2% PROPOSED MODEL

better and a F1-Score of 84% which is also 13% better than the available state-of-
the art Bangla sentiment classifier having a low amount of parameters in our • BNLP utilizes the dataset introduced by Rahman et al,
proposed approach. focusing on positive, negative, and neutral aspects. It
consists of 2900 cricket-related comments from online
news sites and Facebook pages of BBC Bangla News and
Prothom Alo Newspaper in Bangladesh. The dataset is
split into 80% training and 20% testing, with 15% of
INTRODUCTION/BACKGROUND: training data used for validation.
• Tensorflow, Keras, Pillow, and OpenCV Python libraries
• Sentiment analysis, a crucial subfield of NLP, involves classifying the emotional tone are used to make the training and testing protocols for
this Deep CNN model.
of text, aiding businesses in understanding customer sentiments. Despite being a • In Table 01 hyper parameter of our proposed deep CNN
widely spoken language, Bangla is low-resource, limiting sentiment analysis. Model is showed. Here the size of dataset is showed.
However, with the emergence of digital Bangla data, automated sentiment analysis • In table 02, here comparison is shown with other model
, and the highest accuracy is obtained by this CCN
becomes essential for understanding market trends, consumer preferences, and model.
movie reviews. We propose a dependency-free framework for Bangla sentiment • In figure 6 and 7 , here shows the loss and accuracy
analysis, addressing challenges and classifying emotions as positive, negative, or obtained by this model.
• In figure 05 , the polarity of the dataset is shown.
neutral. • analyzes sentiments from texts written in Bangla. The
classifier model obtains a classification accuracy of
84.84%, which is almost 2% better than the available
METHODOLOGY: state-of-the art Bangla sentiment classifier
Figure 05 : HYPER PARAMETERS OF OUR PROPOSED DEEP CNN MODEL

• Proposed methodology focuses on a CNN-based approach to improve accuracy and


computation speed in sentiment analysis for Bengali text data.
• Data preprocessing involves cleaning and tokenization. Word embeddings capture semantic
meaning, enabling a compact representation.
• CNN with Conv1D layers and 1015 activation prevents overfitting. Experimental evaluation
compares the model to other approaches, aiming for a smaller yet accurate sentiment analysis
system. The workflow includes data preprocessing, word embedding, and CNN architecture
building, followed by training and evaluation on diverse datasets.

Figure 06 : Loss of this mode Figure 06 : Accuracy of this model

Data Data Pre- Channel


Raw Data Set
Analysis processing Data CONCLUSIONS:
There has been a push in recent years to increase NLP work in Bangla. In this study we present a model for
opinion extraction based on low-cost CNN architecture with a bit of preprocessing. We use traditional steps to
clean data and achieve better results comparing other state-of-the-art approaches. More detailed annotated
Evaluate the Split Dataset
dataset like SemEval for Bengali language can lead to impressive results. Sentiment analysis is becoming
Training and increasingly used for identifying spam reviews/comments and detecting fraudulent apps. As a result,
performance into Train
Testing
with other Validation
model
Model
Test
broadening research to include additional non-English languages and aspects can lead to a more accurate
knowledge of users and their assessments, which can aid in making better business decisions and increasing
Figure 02: Workflow of our Approach cyber security. We want to use Aspect Based Sentiment Analysis (ABSA) in the future, which is more
sophisticated than typical sentiment analysis. Aspect-based sentiment analysis improves opinion mining by
allowing users to categorize data by characteristic (aspect) and subsequently identify attributable opinions
The methodology begins with crucial data preprocessing, which involves cleaning (sentiment). It automates time-consuming tasks, works in real time, is easy to scale, and gives an unbiased
the text data and tokenizing it into a series of meaningful units. Word embeddings Bangla Text From Dataset
experience for a truly customercentric experience. In order to achieve the goals of Aspect Based Sentiment
are then generated to represent words in a numerical form, capturing their
semantic meaning and context. The CNN model is constructed using Conv1D layers
তারা ক াথায় যারা বলেছিে হাতু রাছ িংহ বনাম বািংোলেশ এর কেো Analysis, we will work on a specific dataset and technique that correlates sentiments with the associated
with a 3-kernel size to extract essential features from the text. To prevent aspects.
overfitting, the 1015 activation function is employed, and fully connected (FC)
layers are utilized to fine-tune the model's performance. The model's output layer
uses the Sigmoid activation function to determine sentiment polarity in complex Tokenization and Converting to Sequence
sentences. The experimental evaluation includes testing the proposed deep CNN [70, 1514, 141, 859, 1, 460, 4, 19, 23]
model against other machine learning approaches, aiming to create a smaller yet REFERENCES:
more accurate sentiment analysis system for Bengali text data. [1] B. Liu, “Sentiment analysis and opinion mining,” Synth. lect. hum. lang. technol., vol. 5, no. 1, pp. 1–167,
Padding Sequence 2012.
[
0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
[2] S. Hochreiter and J. Schmidhuber, “Long short-term memory,” Neural Comput., vol. 9, no. 8, pp. 1735–
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1780, 1997.
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 70 [3] A. Hassan, M. R. Amin, A. K. Al Azad, and N. Mohammed, “Sentiment analysis on bangla and romanized
1514 141 859 1 460 4 19 23]
bangla text using deep recurrent models,” in 2016 International Workshop on Computational Intelligence
(IWCI). IEEE, 2016, pp. 51–56
Figure 03: Text Pre-Processing
[4] R. A. Tuhin, B. K. Paul, F. Nawrine, M. Akter, and A. K. Das, “An automated system of sentiment analysis
from bangla text using supervised learning techniques,” in 2019 IEEE 4th International Conference on
Figure 04: CNN Model for 1D Text Classification Computer and Communication Systems (ICCCS). IEEE, 2019.
[5] M. H. Alam, M.-M. Rahoman, and M. A. K. Azad, “Sentiment analysis for bangla sentences using
convolutional neural network,” in 2017 20th International Conference of Computer and Information
Technology (ICCIT). IEEE, 2017, pp. 1–6.

IN-COLLABORATION WITH:

IEB IEEE IEOM Career ASCE CUET ASRRO ASME RMA CAMS SPE SCBA CCS
ACI
Chattogram Student Branch, CUET CUET Club Student Chapter Computer Club CUET CUET CUET CUET Student Chapter CUET CUET
Student Chapter

ORGANIZED BY: SPONSORED BY:

Directorate of Research and Extension (DRE)


Chittagong University of Engineering and Technology (CUET), Chattogram- 4349, Bangladesh.

You might also like