Browsing by Author "Bakal, Gokhan"

Now showing 1 - 17 of 17

Multi-Method Text Summarization: Evaluating Extractive and BART-Based Approaches on CNN/Daily Mail
(Institute of Electrical and Electronics Engineers Inc., 2025) Inal, Yasin; Bakal, Gokhan; Esit, Muhammed
With the exponential growth of digital content, efficient text summarization has become increasingly crucial for managing information overload. This paper presents a comprehensive approach to text summarization using both extractive and abstractive methods, implemented on the CNN/Daily Mail dataset. We leverage pre-trained BART (Bidirectional and AutoRegressive Transformers) models and fine-tuning techniques to generate high-quality summaries. Our approach demonstrates significant improvements, with our best model trained on 287 k samples achieving ROUGE-1 F1 scores of 0.4174, ROUGE-2 F1 scores of 0.1932, and ROUGE-L F1 scores of 0.2910. We provide detailed comparisons between extractive methods and various BART model configurations, analyzing the impact of training dataset size and model architecture on summarization quality. Additionally, we share our implementation through an opensource NLP toolkit to facilitate further research and practical applications in the field. © 2025 Elsevier B.V., All rights reserved.
Citation - Scopus: 1
Improving Salary Offer Processes With Classification Based Machine Learning Models
(Institute of Electrical and Electronics Engineers Inc., 2024) Kaya, Rukiye; Saatci, Mehtap; Bakal, Gokhan
In job applications, salary is major motivational factor for employees and making accurate salary prediction is crucial for both employers and employees. Utilizing advanced technologies can significantly enhance the accuracy and efficiency of salary prediction process. In this study, we explore Machine Learning (ML) methods to enhance salary prediction process. We evaluated seven classification models for predicting salary categories, with the Artificial Neural Network (ANN) model achieving the highest accuracy at 58.2% on the test dataset, followed by the K-Nearest Neighbors (KNN) model with an accuracy of 56.8%. Additionally, we employed ensemble models to further enhance prediction accuracy. Among these, the Majority Voting Classifier using Hard Voting achieved the highest accuracy at 59.3%, demonstrating the potential of ensemble techniques in refining salary predictions. The developed salary prediction tool estimates the most appropriate salary category for each candidate and help mitigate potential biases in manual salary assessments, hence enables a more objective and consistent compensation system. ∗CRITICAL: Do Not Use Symbols, Special Characters, or Math in Paper Title or Abstract, and do not cite other papers in the abstract. © 2024 Elsevier B.V., All rights reserved.
Citation - WoS: 5
Citation - Scopus: 11
Enhancing Sentiment Analysis in Stock Market Tweets Through Bert-Based Knowledge Transfer
(Springer, 2025) Cicekyurt, Emre; Bakal, Gokhan
One of the widely studied text classification efforts is sentiment analysis. It is a specific examination involving natural language processing and machine learning methods to understand semantic orientation from textual data. Working social media posts, such as tweets, for sentiment analysis, is quite common among researchers due to the speed of information dissemination. In this regard, forecasting stock market tweets is a widely studied research topic. Some studies have revealed a strong connection between sentiment and stock market performance, while others have not found any notable associations. The proposed work shows two distinct approaches to sentiment analysis over the stock market tweets. The first approach employs traditional machine learning algorithms, including logistic regression, random forest, and XGBoost. The second approach constructs deep learning (as a subfield of machine learning) models using LSTM and CNN algorithms to classify the test instances into positive, negative, or neutral classes through ten randomly shuffled data splits. In this study, the labeled data size is gradually increased utilizing a pre-trained model, FinBERT. It is exclusively employed to label unlabeled data instances to integrate them into the experiments. The goal is to monitor the effect of the additional newly-labeled examples on the sentiment analysis performance. The experiments showed that the average F1-score improved by 20% for the deep learning models and 17% for the machine learning models. In the end, the paper reveals a strong positive correlation between training data size and the classification performance of the experimental approaches.
Citation - Scopus: 6
Building a Challenging Medical Dataset for Comparative Evaluation of Classifier Capabilities
(Elsevier Ltd, 2024) Bozkurt, Berat; Coskun, Kerem; Bakal, Gokhan
Since the 2000s, digitalization has been a crucial transformation in our lives. Nevertheless, digitalization brings a bulk of unstructured textual data to be processed, including articles, clinical records, web pages, and shared social media posts. As a critical analysis, the classification task classifies the given textual entities into correct categories. Categorizing documents from different domains is straightforward since the instances are unlikely to contain similar contexts. However, document classification in a single domain is more complicated due to sharing the same context. Thus, we aim to classify medical articles about four common cancer types (Leukemia, Non-Hodgkin Lymphoma, Bladder Cancer, and Thyroid Cancer) by constructing machine learning and deep learning models. We used 383,914 medical articles about four common cancer types collected by the PubMed API. To build classification models, we split the dataset into 70% as training, 20% as testing, and 10% as validation. We built widely used machine-learning (Logistic Regression, XGBoost, CatBoost, and Random Forest Classifiers) and modern deep-learning (convolutional neural networks - CNN, long short-term memory - LSTM, and gated recurrent unit - GRU) models. We computed the average classification performances (precision, recall, F-score) to evaluate the models over ten distinct dataset splits. The best-performing deep learning model(s) yielded a superior F1 score of 98%. However, traditional machine learning models also achieved reasonably high F1 scores, 95% for the worst-performing case. Ultimately, we constructed multiple models to classify articles, which compose a hard-to-classify dataset in the medical domain. © 2024 Elsevier B.V., All rights reserved.
Citation - Scopus: 7
A Comparative Analysis on Medical Article Classification Using Text Mining & Machine Learning Algorithms
(Institute of Electrical and Electronics Engineers Inc., 2021) Kolukisa, Burak; Dedeturk, Bilge Kagan; Dedeturk, Beyhan Adanur; Gulsen, Abdulkadir; Bakal, Gokhan
The document classification task is one of the widely studied research fields on multiple domains. The core motivation of the classification task is that the manual classification efforts are impractical due to the exponentially growing document volumes. Thus, we densely need to exploit automated computational approaches, such as machine learning models along with data & text mining techniques. In this study, we concentrated on the classification of medical articles specifically on common cancer types, due to the significance of the field and the decent number of available documents of interest. We deliberately targeted MEDLINE articles about common cancer types because most cancer types share a similar literature composition. Therefore, this situation makes the classification effort relatively more complicated. To this end, we built multiple machine learning models, including both traditional and deep learning architectures. We achieved the best performance (R¿82% F score) by the LSTM model. Overall, our results demonstrate a strong effect of exploiting both text mining and machine learning methods to distinguish medical articles on common cancer types. © 2022 Elsevier B.V., All rights reserved.
Text Classification Experiments on Contextual Graphs Built by N-Gram Series
(Springer International Publishing AG, 2025) Sen, Tarik Uveys; Yakit, Mehmet Can; Gumus, Mehmet Semih; Abar, Orhan; Bakal, Gokhan
Traditional n-gram textual features, commonly employed in conventional machine learning models, offer lower performance rates on high-volume datasets compared to modern deep learning algorithms, which have been intensively studied for the past decade. The main reason for this performance disparity is that deep learning approaches handle textual data through the word vector space representation by catching the contextually hidden information in a better way. Nonetheless, the potential of the n-gram feature set to reflect the context is open to further investigation. In this sense, creating graphs using discriminative ngram series with high classification power has never been fully exploited by researchers. Hence, the main goal of this study is to contribute to the classification power by including the long-range neighborhood relationships for each word in the word embedding representations. To achieve this goal, we transformed the textual data by employing n-gram series into a graph structure and then trained a graph convolution network model. Consequently, we obtained contextually enriched word embeddings and observed F1-score performance improvements from 0.78 to 0.80 when we integrated those convolution-based word embeddings into an LSTM model. This research contributes to improving classification capabilities by leveraging graph structures derived from discriminative n-gram series.
Citation - WoS: 13
Citation - Scopus: 9
An Empirical Study of Sentiment Analysis Utilizing Machine Learning and Deep Learning Algorithms
(Springernature, 2024) Erkantarci, Betul; Bakal, Gokhan
Among text-mining studies, one of the most studied topics is the text classification task applied in various domains, including medicine, social media, and academia. As a sub-problem in text classification, sentiment analysis has been widely investigated to classify often opinion-based textual elements. Specifically, user reviews and experiential feedback for products or services have been employed as fundamental data sources for sentiment analysis efforts. As a result of rapidly emerging technological advancements, social media platforms such as Twitter, Facebook, and Reddit, have become central opinion-sharing mediums since the early 2000s. In this sense, we build various machine-learning models to solve the sentiment analysis problem on the Reddit comments dataset in this work. The experimental models we constructed achieve F1 scores within intervals of 73-76%. Consequently, we present comparative performance scores obtained by traditional machine learning and deep learning models and discuss the results.
Citation - Scopus: 4
A Transfer Learning Application on the Reliability of Psychological Drugs' Comments
(Institute of Electrical and Electronics Engineers Inc., 2023) Sen, Tarik Uveys; Bakal, Gokhan
As digitalization and the Internet stay emerging concepts by gaining popularity, the accuracy of personal reviews/opinions will be a critical issue. This circumstance also particularly applies to patients taking psychological drugs, where accurate information is crucial for other patients and medical professionals. In this study, we analyze drug reviews from drugs.com to determine the effectiveness of reviews for psychological drugs. Our dataset includes over 200,000 drug reviews, which we labeled as positive, negative, or neutral according to their rating scores. We apply machine learning (ML) models, including Logistic Regression, Recurrent Neural Network (RNN), and Long Short-Term Memory (LSTM) algorithms, to predict the sentiment class of each review. Our results demonstrate an F1-Weighted score of 85.3% for the LSTM model. However, by applying the transfer learning technique, we further improved the F1 score (nearly 3% increase) obtained by the LSTM model. Our findings proved that there is no contextual difference between the comments made by the patients suffering from psychological or other diseases. © 2023 Elsevier B.V., All rights reserved.
Citation - WoS: 4
Citation - Scopus: 4
Beyond Visual Cues: Emotion Recognition in Images With Text-Aware Fusion
(Elsevier, 2025) Sungur, Kerim Serdar; Bakal, Gokhan
Sentiment analysis is a widely studied problem for understanding human emotions and potential outcomes. As it can be performed over textual data, working on visual data elements is also critically substantial to examining the current emotional status. In this effort, the aim is to investigate any potential enhancements in sentiment analysis predictions through visual instances by integrating textual data as additional knowledge reflecting the contextual information of the images. Thus, two separate models have been developed as image-processing and text-processing models in which both models were trained on distinct datasets comprising the same five human emotions. Following, the outputs of the individual models' last dense layers are combined to construct the hybrid multimodel empowered by visual and textual components. The fundamental focus is to evaluate the performance of the hybrid model in which the textual knowledge is concatenated with visual data. Essentially, the hybrid model achieved nearly a 3% F1-score improvement compared to the plain image classification model utilizing convolutional neural network architecture. In essence, this research underscores the potency of fusing textual context with visual information to refine sentiment analysis predictions. The findings not only emphasize the potential of a multi-modal approach but also spotlight a promising avenue for future advancements in emotion analysis and understanding.
NLP-Driven Fake News Detection: A Machine Learning Perspective
(IEEE, 2025) Coban, Mert Korkut; Bakal, Gokhan
The rapid spread of fake news poses a significant challenge, impacting public opinion, decision-making, and societal trust. This study explores the application of Natural Language Processing (NLP) and Machine Learning (ML) techniques for robust fake news detection. Using datasets such as ISOT Fake News, WELFake, and Football Fake News, the project employs advanced preprocessing methods and feature extraction techniques, including TF-IDF, Word2Vec, and GloVe. A comprehensive evaluation of machine learning models-Random Forest, Support Vector Machines (SVM), and Neural Networks-was conducted to identify the optimal configuration. Results demonstrate that Random Forest with TF-IDF excels in in-domain detection, achieving an F1-score of 99.70%, while Neural Networks paired with Word2Vec and GloVe embeddings outperform in cross-dataset scenarios. The study highlights the importance of dataset size, domain relevance, and feature representation in achieving high generalizability. These findings provide a scalable framework for combating misinformation on digital platforms.
Citation - WoS: 1
Machine Learning Based Network Intrusion Detection With Hybrid Frequent Item Set Mining
(Gazi Univ, 2024) Firat, Murat; Bakal, Gokhan; Akbas, Ayhan
With the development and expansion of computer networks day by day and the diversity of software developed, the damage that possible attacks can cause is increasing beyond the predictions. Intrusion Detection Systems (STS/IDS) are one of the practical defense tools against these potential attacks that are constantly growing and diversifying. Thus, one of the emerging methods among researchers is to train these systems with various artificial intelligence methods to detect subsequent attacks in real time and take the necessary precautions. However, the ultimate goal is to propose a hybrid feature selection approach to improve the classification performance. The raw dataset originally enclosed 85 descriptor features (attributes) for classification. These attributes are extracted using CICFlowMeter from a PCAP file where network traffic is recorded for data curation. In this study, classical feature selection methods and frequent item set mining approaches were employed in feature selection for constructing a hybrid model. We aimed to examine the effect of the proposed hybrid feature selection approach on the classification task for the network traffic data containing ordinary and attack records. The outcomes demonstrate that the proposed method gained nearly 3% improvement when applied with the Logistic Regression algorithm on classifying more than 225,000 records.
From Traditional to Deep: Evaluating Sentiment Analysis Models on a Large-Scale Tweet Dataset
(Institute of Electrical and Electronics Engineers Inc., 2024) Mammadov, Alisahib; Bakal, Gokhan
This study investigates the effectiveness of various machine learning (ML) and deep learning (DL) techniques for large-scale sentiment analysis on Twitter data. We leverage a publicly available dataset of one million tweets, annotated with four sentiment labels (positive, negative, uncertainty, and liti-gious), to train and evaluate a range of models. Our experiments demonstrate that traditional ML algorithms, particularly XG-Boost, achieve high performance, with the best F1 score reaching 95.81% using a combination of unigrams and bigrams. Among DL models, a hybrid CNN-BiGRU architecture yields the highest average F1 score of 95.42%. Our findings highlight the strengths of different approaches for sentiment analysis on Twitter data and emphasize the importance of data preprocessing and model selection for achieving optimal performance. © 2025 Elsevier B.V., All rights reserved.
Graph-Based Biomedical Knowledge Discovery
(IEEE, 2024) Altuner, Osman; Bakir-Gungor, Burcu; Bakal, Gokhan
The digitalization process is progressing at a very high speed all over the world. While this situation provides many conveniences in today's life, it also brings along a problem such as analyzing and processing the huge digital data. This also applies to published academic studies. In this sense, the process of evaluating each study to access previously unknown information within the studies requires a very laborious process. For this reason, in this study, the publications obtained for the target diseases were analyzed by text analysis processes and converted into a graph structure that enables the linking of meaningful terms through biomedical relationships. On the dense graph structure obtained, binary biomedical entities with important links such as treats, causes, associated_with were queried. The entity pairs obtained according to the query results were also confirmed by manual search method and proved to be real connections. In this study, retrieval of known biomedical entities with the proposed approach solved the time-consuming manual search problem. There is also the potential to obtain unknown/unexplored possible new relationships (e.g., therapeutic, causal, etc.) with multiple binary linking patterns.
Words Speak Louder Than Actions: Decoding Emotions Through NLP
(Institute of Electrical and Electronics Engineers Inc., 2024) Paksoy, Melda; Bakal, Gokhan
Emotion detection in text remains a significant challenge in Natural Language Processing due to human emotions' complexity and subtle nuances. This paper presents multiple experimental models for emotion classification using an up-to-date dataset curated to address 13 emotions implied in Twitter posts. We evaluated various machine learning (ML) models, including Logistic Regression, Random Forest, SVM, and XGBoost, alongside deep learning (DL) architectures such as LSTM and CNN. Our results demonstrate the efficacy of deep learning models, particularly the CNN model by achieving an impressive F1 score of 0.99. This study contributes to emotion detection capabilities, paving the way for more nuanced and accurate sentiment analysis (SA) in various text analysis applications. © 2025 Elsevier B.V., All rights reserved.
Citation - Scopus: 8
On Comparative Classification of Relevant COVID-19 Tweets
(Institute of Electrical and Electronics Engineers Inc., 2021) Bakal, Gokhan; Abar, Orhan
Due to the impressive information dissemination power of social networks such as Twitter, people tend to check social networks and Web pages more than other traditional news sources, including newspapers, TV news programs, or radio channels. In that sense, the information carried by the content of the shared social media posts becomes much more considerable. However, most of the posts are commonly either irrelevant or inaccurate. Besides, the more critical case than the correctness of the information is the diffusion speed on Twitter through the reply or retweet actions. These activities make the initial situation even more complicated than itself due to the unregulated nature of the social networks and the lack of an immediate verification mechanism for the correctness of the posts. When we consider the current Covid-19 pandemic period (causing the coronavirus disease), one of the most utilized information resources is Twitter except the official health administration institutions. Thereupon, examining the correctness of the information related to the Covid-19 pandemic by computational techniques (e.g., Data Mining, Machine Learning, and Deep Learning) has been gaining popularity and remains a substantial task. Hence, we mainly focused on analyzing the correctness of the posts related to the current pandemic shared on the Twitter platform. Therefore, the overall goal of this work is to classify the relevant tweets using linear and non-linear machine learning models. We achieved the best F1 performance score (99%) with the neural network model using the unigram features & threshold value of 50 among all model configurations. © 2022 Elsevier B.V., All rights reserved.
Citation - WoS: 3
Citation - Scopus: 4
Combining N-Grams and Graph Convolution for Text Classification
(Elsevier, 2025) Sen, Tarik Uveys; Yakit, Mehmet Can; Gumus, Mehmet Semih; Abar, Orhan; Bakal, Gokhan
Text classification, a cornerstone of natural language processing (NLP), finds applications in diverse areas, from sentiment analysis to topic categorization. While deep learning models have recently dominated the field, traditional n-gram-driven approaches often struggle to achieve comparable performance, particularly on large datasets. This gap largely stems from deep learning' s superior ability to capture contextual information through word embeddings. This paper explores a novel approach to leverage the often-overlooked power of n-gram features for enriching word representations and boosting text classification accuracy. We propose a method that transforms textual data into graph structures, utilizing discriminative n-gram series to establish long-range relationships between words. By training a graph convolution network on these graphs, we derive contextually enhanced word embeddings that encapsulate dependencies extending beyond local contexts. Our experiments demonstrate that integrating these enriched embeddings into an long-short term memory (LSTM) model for text classification leads to around 2% improvements in classification performance across diverse datasets. This achievement highlights the synergy of combining traditional n-gram features with graph-based deep learning techniques for building more powerful text classifiers.
Citation - Scopus: 2
A Computational Drug Repositioning Effort Using Patients' Reviews Dataset
(Institute of Electrical and Electronics Engineers Inc., 2023) Akkaya, Ali; Bakal, Gokhan
The drug discovery process is one of the core motivations in both medical and, specifically, pharmaceutical disciplines. Due to the nature of the process, it requires an excessive amount of time, clinical experiments, and budget to cover each discovery phase. In this sense, computational drug discovery efforts can shorten the discovery process by providing plausible candidates since many of the attempts fail for several reasons, such as a lack of participants, financial problems, or ineffective results. In this study, the goal is to identify plausible candidate drugs for diseases. To do that, we utilize a personal experience of drugs dataset generated by patients. Beyond the user-generated comments, the users also give a rate between 1 and 10. Since we want to ensure the dataset quality, we first performed sentiment analysis experiments to prove that the reviews/comments are consistent with the given rating score. Then, only the review pairs having an effectiveness rate of 6 or more are selected as pre-filtered drug-disease pairs. We also build a knowledge graph using treatment-related biomedical relations using predications from Semantic Medline Database to identify drug similarities utilizing the Simrank similarity algorithm. As a result, we reported a list of plausible drugs as repurposing/repositioning candidates for further experiments. © 2023 Elsevier B.V., All rights reserved.