Paper Review

The document discusses two papers related to artificial intelligence. The first paper investigates using GANs for Bengali fake review detection. The second addresses glaucoma detection using transfer learning and model explanations.

Uploaded by

Dipta Biswas

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PDF, TXT or read online on Scribd

0% found this document useful (0 votes)

7 views3 pages

Paper Review

Uploaded by

Dipta Biswas

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PDF, TXT or read online on Scribd

You are on page 1/ 3

Ahsanullah University of Science & Technology

Department of Computer Science & Engineering

Paper Review
Submitted To
Md. Tanvir Rouf Shawon
Lecturer,Cse
Mr. Raihan Tanvir
Lecturer,CSE

Course Number & Name:

CSE-4138 : Soft Computing Lab
Submitted By

Name : Dipta Biswas

Id : 20200104013
Section :A
Date of Submission : 26-11-2023
Paper 1: Bengali Fake Review Detection using Semi-supervised Generative Adversarial
Networks.

Paper Content:
1. The paper investigates the application of semi-supervised Generative Adversarial Networks (GANs) for
fine-tuning pretrained language models to classify Bengali fake reviews.
2. It addresses the challenges in detecting deceptive reviews in low-resource languages like Bengali,
emphasizing the importance of protecting consumers from misleading information.

Ideas:
1. The primary idea revolves around leveraging GANs in conjunction with pretrained language models to
enhance the accuracy of fake review detection.
2. The study explores the effectiveness of a semi-supervised GAN-LM architecture for this purpose.

Structure:
1. The paper follows a structured format, including sections such as Introduction, Related Works, Background
Study, Proposed Methodology, Experimental Results & Evaluations, and Conclusion.
2. Each section contributes to building a comprehensive understanding of the research, its context, and
outcomes.

Writings:
1. The writing style is technical and research-oriented, focusing on presenting the problem, methodology, and
experimental results in a clear and concise manner.
2. Specific attention is given to detailing the GAN-LM architecture, pretrained language models, and the challenges
associated with fake review detection.

Methodology:
1. The proposed methodology involves a four-step process: Text Normalization, Data Distribution,
Fine-tuning Language Models with Semi-supervised GAN, and Evaluation using performance metrics.
2. The GAN-LM architecture includes a generator and discriminator, utilizing pretrained language models as
encoders.

Datasets:
1. The dataset comprises 6014 Bengali reviews collected from publicly accessible Facebook groups.
2. Reviews are manually labeled as authentic or fake, and the dataset serves as a benchmark for training and
evaluating the proposed methodology.
3. Class distribution indicates 5015 authentic reviews and 871 fake reviews.

Result Analysis:
1. BanglaBERT achieves the highest accuracy (83.59%) and F1-score (84.89%) with 1024 annotated samples,
outperforming other language models in the proposed GAN-LM architecture.
2. Comparative analysis reveals the stability of BanglaBERT's accuracy across different sample sizes, while
BanglaBERT generators with fewer parameters closely follow in performance.
3. Despite variations, Bangla-Electra and SahajBERT demonstrate reasonable accuracy, emphasizing the
potential of these models in the context of limited labeled data.

1
Paper 2: Explainable AI based Glaucoma Detection usingTransfer Learning and LIME

Paper Content:
1. The paper addresses the problem of glaucoma detection, emphasizing the limitations of traditional
screening methods.
2. It highlights the significance of automating the classification process and introduces a transfer learning
model with high accuracy.
3. The content covers challenges in the existing models and proposes solutions for more reliable glaucoma
detection.

Ideas:
1. The key idea revolves around using a transfer learning model for glaucoma detection and the importance of
explainability in artificial intelligence.
2. The paper explores the potential of various pre-trained models, emphasizing the need for interpretable
systems in the medical field.

Structure:
1. The paper follows a structured format, beginning with an introduction to glaucoma and its impact. It
discusses related work in a literature review, presents the methodology, and analyzes the results.
2. The structure ensures a logical flow of information, leading to a conclusive interpretation.

Writings:
1. The writing is clear and focused, effectively conveying complex concepts related to glaucoma detection and
artificial intelligence.
2. The authors use technical terms appropriately, making the content accessible to both medical professionals
and those with a background in artificial intelligence.

Methodology:
1. The methodology involves the use of a transfer learning model and Local Interpretable Model-Agnostic
Explanations (LIME) for explainability.
2. The authors compare the performance of various pre-trained models, addressing challenges such as
overfitting and computing resources.

Datasets:
1. The dataset consists of fundus images, providing a real-world basis for glaucoma detection.
2. The paper specifies the size of the dataset, its division into training, testing, and validation sets, and the
labeling criteria for glaucoma and non-glaucoma cases.

Result Analysis:
1. Compares proposed model performance with pre-trained models.
2. Includes accuracy metrics, misclassification counts, and a validation curve.
3. Identifies ResNet50 as the top-performing model, achieving 94.7% accuracy.
4. Acknowledges VGG-19 as another well-fitted model.
5. Applies LIME to explain model predictions, addressing the black box issue in medical deep learning.