Source-linked AI summary
Enhancing Bangla Fake News Detection Using Bidirectional Gated Recurrent Units and Deep Learning Techniques
Utsha Roy, Mst. Sazia Tahosin, Md. Mahedi Hassan, Taminul Islam, Fahim Imtiaz, Md Rezwane Sadik, Yassine Maleh, Rejwan Bin Sulaiman, Md. Simul Hasan Talukder
TL;DR
The paper addresses limited research attention to fake-news detection in Bangla, a language with distinct linguistic challenges. It proposes a dataset and evaluates several deep-learning models, including bidirectional GRU. The balanced-data results report 99.16% accuracy for the two-stage GRU model.
Problem
Fake-news research has focused largely on major languages, while Bangla's linguistic diversity, idioms, and cultural nuances present distinct detection challenges.
Source: §1 · p. 1Method
The study uses Bangladeshi news articles to evaluate bidirectional GRU, CNN, LSTM, and hybrid deep-learning models, and examines class inequality.
Source: §1 · pp. 1–2Results
99.16% accuracy was reported for the two-stage GRU model on balanced data.
Source: §4.1 · p. 6Takeaways & Limitations
The authors identify dataset balancing and expansion with more pseudo-events as priorities for a more representative training environment.
Source: §6 · p. 7
Abstract
from arXiv · showhide
The rise of fake news has made the need for effective detection methods, including in languages other than English, increasingly important. The study aims to address the challenges of Bangla which is considered a less important language. To this end, a complete dataset containing about 50,000 news items is proposed. Several deep learning models have been tested on this dataset, including the bidirectional gated recurrent unit (GRU), the long short-term memory (LSTM), the 1D convolutional neural network (CNN), and hybrid architectures. For this research, we assessed the efficacy of the model utilizing a range of useful measures, including recall, precision, F1 score, and accuracy. This was done by employing a big application. We carry out comprehensive trials to show the effectiveness of these models in identifying bogus news in Bangla, with the Bidirectional GRU model having a stunning accuracy of 99.16%. Our analysis highlights the importance of dataset balance and the need for continual improvement efforts to a substantial degree. This study makes a major contribution to the creation of Bangla fake news detecting systems with limited resources, thereby setting the stage for future improvements in the detection process.
1 Introduction
The paper addresses the limited attention to Bangla fake-news detection despite the language's linguistic and cultural challenges. It proposes a dataset and evaluates deep-learning models, including bidirectional GRU, while also examining class inequality.
1 Introduction
- Prior fake-news research has focused mainly on major languages, while Bangla presents linguistic diversity, idiomatic expressions, and cultural nuances.The authors position their study as applying deep learning to this setting.Source: §1 · p. 1
- The study uses Bangladeshi news articles and describes a dataset intended to support analysis of false news in Bangladesh.The authors also describe processing the data and using WordClouds to visualize frequent words.Source: §1 · p. 1
- The work examines 1D CNN, LSTM, and hybrid deep-learning models, as well as regression methods to address class inequality and efforts to reduce false positives and false negatives.Source: §1 · p. 2
- The study evaluates bidirectional GRU for determining the accuracy of Bangla news articles.This is identified as a principal contribution of the work.Source: §1 · p. 2
- The paper aims to improve authentic-versus-fake classification accuracy while addressing class inequality through alternative modeling techniques.Source: §1 · p. 2
- The paper frames its contribution as bridging advanced learning methods and Bangla-specific challenges in fake-news detection.Source: §1 · p. 2
2 Literature Review
The reviewed work spans Bangla-specific datasets and models, imbalance treatments, and multilingual or broader-domain fake-news detection. Reported results vary across datasets and approaches.
2 Literature Review
- A Bangla dataset study reported 91% F1 for SVM, while its neural models performed less effectively [7].Source: §2 · p. 2
- A study combining fresh and earlier Bangla fake-news data reported 95.9% accuracy for CNN on 4,678 distinct news items.Source: §2 · p. 2
- On BanFakeNews, which had 97% majority instances, SMOTE yielded 93.1% F1 and Stacked Generalization 79.1%, compared with 67.6% for baseline models without these techniques [9].Source: §2 · p. 2
- Other reviewed work includes a Bengali-and-English detection approach reporting 93.29% accuracy and 0.93 F1 with Linear Support Vector Classification.The review also describes BERT-based Bengali detection and broader-domain English fake-news studies.Source: §2 · p. 2
3 Proposed Methodology
The methodology combines data collection and cleaning with word visualization, deep-learning model training, and permutation methods intended to address class heterogeneity. The described dataset comprises four files with article information and labels, plus additional context fields in two files.
3 Proposed Methodology
- The proposed workflow collects Bangladeshi news data, cleans text, visualizes common words with WordClouds, trains several deep-learning models, and uses permutation methods to address class heterogeneity.The models include bidirectional GRU, 1D CNN, LSTM, and combinations of these approaches.Source: §3 · p. 3, Figure 1 · p. 3
3.1 Dataset Description
- The four dataset files include core article fields and labels, with two labeled files adding source, relation, and—in the fake-news file—fake-news type.The core files are Authentic-48K.csv and Fake-1K.csv; the additional files are LabeledAuthentic-7K.csv and LabeledFake-1K.csv.Source: §3.1 · p. 3, Table 1 · p. 3
3.2 Data Preprocessing
The preprocessing pipeline prepares news text for analysis and model training; WordClouds visualize words in the training and test datasets to help identify themes and patterns.
3.2 Data Preprocessing
- The pipeline extracts content and label columns, combines the real and fake news datasets, and applies typographical corrections before analysis and training.It also uses WordCloud visualization and tokenization as part of preprocessing.Source: §3.2 · p. 3
- Tokenization transforms words into mathematical vectors so they can be used in deep-learning models.The authors state that this processing supports analysis and identification of misinformation in Bangla.Source: §3.2 · p. 3
3.3 WordCloud Visualization
- WordClouds visualize words in both training and test datasets, helping identify themes and patterns in Bangla text.The authors describe these visualizations as a guide for data creation and false-positive analysis.Source: §3.3 · p. 4, Figure 2 · p. 4, Figure 3 · p. 4
3.4 Stopword Removal
The text-cleaning steps use a Bengali stopword library and remove unwanted text elements, while retaining news items that exceed a 100-word threshold.
3.4 Stopword Removal
- The Bengali stopword library is used to remove redundant characters and common punctuation during text processing.The authors identify the library as “Stopwords Bengali (BN)” [22].Source: §3.4 · p. 4
3.5 Text Cleaning
- News items are retained only when they exceed a 100-word limit.The authors describe this as a filtering step intended to meet information-quality needs.Source: §3.5 · p. 4
- Text correction removes line breaks, tabs, special marks, numeric characters, and specified Bengali characters.The stated aim is to refine the content for analysis.Source: §3.5 · p. 4
3.6 Data Preparation
The study reports separate training and testing datasets and uses tokenization and padding to prepare text sequences for model input.
3.6 Data Preparation
- The authors report using 49,000 articles for training and 8,000 articles for testing, without a random split.The training and testing datasets were assembled from authentic and fake news database files.Source: §3.6 · p. 4
3.7 Tokenization and Padding
- Tokenization converts text into numeric sequences, assigning numerical values to words for model processing.Figure 4 is identified as an example of tokenization.Source: §3.7 · p. 4, Figure 4 · p. 4
- Padding makes sequences a constant length, which the authors say supports convergence during training.The text presents tokenization and padding as preparation for models to learn from Bangla news content.Source: §3.7 · p. 4
3.8 Models
The study evaluates CNN, LSTM, bidirectional GRU, and hybrid deep-learning architectures for Bangla fake-news detection, with descriptions of each model’s layer structure.
3.8 Models
- The study evaluates CNN, LSTM, bidirectional GRU, CNN + LSTM, CNN + GRU, and CNN + LSTM + Global Max Pooling 1D models.The authors say they compare models to capture language variations, patterns, and context.Source: §3.8 · p. 5
3.8.1 Bidirectional Gated Recurrent Unit.
- The bidirectional GRU model combines information from both directions, followed by global average pooling, dense ReLU layers, dropout, and a sigmoid output for binary classification.The model includes an embedding layer before the bidirectional GRU layer.Source: §3.8.1 · p. 5
3.8.2 1D Convolutional Neural Network.
- The 1D CNN architecture uses embedding, convolutional, and max-pooling layers for feature extraction, followed by dense layers and dropout.The authors state that dense layers and dropout improve generalizability and prevent overfitting.Source: §3.8.2 · p. 5
3.8.3 Long Short-Term Memory.
- The LSTM model connects embedding layers to LSTM layers, then uses dense layers with dropout and a sigmoid output for binary classification.Source: §3.8.3 · p. 5
3.8.4 CNN + LSTM.
- The CNN-LSTM hybrid combines embedding, convolutional, and LSTM layers, merging convolutional and LSTM outputs before the dense layer.Source: §3.8.4 · p. 5
3.8.5 CNN + GRU
- The CNN-GRU hybrid combines convolutional layers for feature extraction with GRU layers for sequential dependencies before dense layers.Source: §3.8.5 · p. 5
3.8.6 CNN + LSTM + Global Max Poling 1D.
- The CNN + LSTM + Global Max Pooling 1D hybrid combines convolutional feature extraction, LSTM sequence analysis, and global reference capture.Source: §3.8.6 · p. 5
3.8 Proposed Model
The proposed model uses a bidirectional GRU to capture context from both preceding and following terms. It combines text embeddings, recurrent processing, pooling, dense layers, dropout, and a sigmoid output for binary classification.
3.8 Proposed Model
- The proposed bidirectional GRU captures context from both preceding and following terms.Figure 5 presents the proposed architecture.Source: §3.8 · p. 5, Figure 5 · p. 5
- The architecture converts text to numerical embeddings, processes it with a bidirectional GRU, then applies GlobalAveragePooling1D and dense layers.The dense layers introduce nonlinearity.Source: §3.8 · p. 5
- Dropout is used to reduce overfitting and improve generalization.Source: §3.8 · p. 5
- A sigmoid output classifies Bangla news content as valid or deceptive in a binary classification task.Source: §3.8 · p. 5
4 Experimental Analysis
The experiments evaluate accuracy, precision, recall, and F1 across several individual and hybrid models, before and after dataset balancing. The reported balanced-set results include 99.16% accuracy for the two-stage GRU, and the authors report better model performance with balanced data.
4 Experimental Analysis
- The evaluation considers accuracy, precision, recall, and F1 score to assess model performance.The metrics are described as providing a more complete view of true and false positive and negative classifications.Source: §4 · pp. 5–6, §4.1 · p. 6
4.1 Result Analysis
- Before balancing, the bidirectional GRU achieved 99.12% accuracy, 99.28% precision, 99.85% recall, and 99.56% F1 score.Table 2 presents the pre-balancing results.Source: §4.1 · p. 6, Table 2 · p. 6
- 99.16% accuracy is reported for the two-stage GRU on the balanced dataset.The authors state that their models perform better with balanced data.Source: §4.1 · p. 6, Table 3 · p. 6
- The confusion matrices and ROC-AUC curves cover the bidirectional GRU, 1D CNN, LSTM, and CNN + LSTM models.The confusion matrices show true positives, true negatives, false positives, and false negatives; the ROC-AUC curves illustrate the true-positive versus false-positive rate trade-off.Source: §4.1 · p. 6, Figure 6 · p. 6, Figure 7 · p. 6
- The CNN + GRU and CNN + LSTM + GlobalMaxPooling1D models also have confusion-matrix and ROC-AUC figures.Source: §4.1 · p. 6, Figure 10 · p. 7, Figure 11 · p. 7
5 Discussion
The discussion reports favorable Bidirectional GRU performance across accuracy, precision, recall, and F1, alongside evaluation of several deep-learning architectures. It presents the work as relevant to fake-news detection in Bangla and to challenges in other languages with limited resources.
5 Discussion
- The Bidirectional GRU showed outstanding results across accuracy, precision, recall, and F1 scores.Source: §5 · p. 7
- The evaluation included 1D CNN, LSTM, hybrid architecture, and bidirectional GRU models.Source: §5 · p. 7
- The study offers insights and perspectives for addressing similar challenges in other languages with limited resources.Source: §5 · p. 7
6 Conclusion
The conclusion reports favorable model outcomes for Bangla fake-news detection and identifies dataset balance as a future priority. It also points to intended practical applications and work across a wider range of languages.
6 Conclusion
- The models had favorable outcomes in accuracy, precision, recall, and F1 scores for identifying false positives.Source: §6 · p. 7
- Future work intends to expand the dataset with more pseudo-events to create a more representative and balanced training environment.Source: §6 · p. 7
- The authors aim to support practical applications and fake-news detection across a wider range of languages.Source: §6 · p. 7