Abstract:
Sentiment analysis is a computational method that uses preliminary emotion analysis to
retrieve feelings and key phrases from various texts (e.g., Positive, Negative, Neutral). It’s
necessary to extract useful information from big data, categorize it, and predict end-user
behavior or emotions. Text classification is a research area of Natural Language Processing
(NLP). Which is converted from unstructured data to meaningful categorical classes. All
previous work is most likely based on traditional different classifiers such as KNN, SVM,
and so on. In this study, we propose a method which is combined in two familiar deep
learning models: Convolutional Neural Networks (CNN) and Bidirectional Long Short
Term Memory (BiLSTM). CNN method retrieves greater characteristics by convolutional
layers and max pooling layers and BiLSTM can capture long term dependencies by lexical
items and it’s better for text classification. Our own built datasets collected from various
Bengali newspapers, such as Bangla Tribune, The Daily Sun etc. generate massive amounts
of data. Our dataset size is 5996. The accuracy of the proposed model differs optimizer
wise. We use three different optimizers: Adam, Adamax, RMSProp. With these three
optimizers, the highest outcomes come from Adam optimizer. Accuracy of our first
proposed model (BiLSTM CNN) with word2vec and Glove word embedding is 91.59%
and 99.40%. In (CNN-BiLSTM) methods obtained outcomes are 92.60% and 94% and the
final model (BiLSTM-CNN-BiLSTM) got their results 82.79% and 98.47%. The
experimental results represent how the deep learning models effectively work. We say that
the amount and quality of training examples have a significant impact on models
performance.
