Suppawong Tuarob

This study introduces a novel approach to enhance acne severity classification by leveraging cross-dataset training within a Label Distribution Learning (LDL) framework. Traditional acne grading methods are often subjective and inconsistent, while automated systems struggle to generalize across diverse datasets, limiting their real-world applicability. This research addresses this challenge by integrating data from multiple sources (Acne04 and AcneSCU) to train a robust model capable of accurately classifying acne severity despite variations in image quality, lighting, and skin tone. We explore the impact of different backbone architectures (ResNet50, EfficientNet B7, and BEiT) within the LDL framework, which is specifically designed to handle the inherent ambiguity in acne severity labeling. Our results demonstrate that cross-dataset training within the LDL framework significantly improves generalization performance compared to models trained on individual datasets. Additionally, results demonstrate that integrating EfficientNet B7 and BEiT enhances the model’s accuracy and robustness. This novel approach offers a promising solution for developing scalable and reliable automated acne grading systems suitable for telemedicine and dermatology applications, ultimately aiding dermatologists in providing more precise and personalized treatment plans.

Citation function analysis is crucial for understanding how cited literature affects the narrative in scientific publications, as citations serve multiple purposes that need accurate distinction and classification. The field faces challenges due to insufficient labeled data and the complexity of defining and categorizing citation functions, requiring specialized knowledge and a deep understanding of scholarly literature. This limitation leads to the imprecise identification and classification of citation functions. To mitigate this challenge, we propose a meta-learning strategy that utilizes prompt learning with pre-trained language models, also known as prompt-based tuning, for the task of few-shot learning in citation function classification. Our findings demonstrate that prompt-based tuning with SciBERT surpasses state-of-the-art pre-trained language models with the conventional fine-tuning approach when labeled data is scarce. Furthermore, we present an analysis that sheds light on the impact of template selection on the prompt-based tuning methodology.

Citation function analysis is crucial to understanding how cited literature contributes to the overall discourse and meaning conveyed in scientific publications. Citation functions serve diverse roles that must be accurately identified and categorized. Still, the field of citation function analysis faces challenges due to limited labeled data and the complexity of defining and categorizing citation functions, which require expertise and a deep understanding of scientific literature. This limitation results in imprecise identification and categorization of citation functions, emphasizing the need for further advancements to improve the accuracy and reliability of citation function analysis. This paper proposes GAN-CITE, a novel framework employing semi-supervised learning techniques to address these limitations. Its primary objective is to efficiently leverage available unlabeled data by combining generative adversarial networks (GANs) and the language model to incorporate substantial data representations from unlabeled data sources. Our study demonstrates that GAN-CITE outperforms both supervised and semi-supervised state-of-the-art models in limited data settings, namely 10%, 20%, and 30% of the total labeled data. We also examine its performance in insufficient and imbalanced labeled data situations, as well as the potential of unlabeled data utilization. These findings highlight the success of generative adversarial networks in enhancing citation function classification and their applications in digital libraries that require precise citation function categorization, such as trend analysis and impact quantification, under limited annotated data.

Iterative approaches like Agile Scrum are commonly adopted to enhance the software development process. However, challenges such as schedule and budget overruns still persist in many software projects. Several approaches employ machine learning techniques, particularly classification, to facilitate decision-making in iterative software development. Existing approaches often concentrate on characterizing a sprint to predict solely productivity. We introduce Sprint2Vec, which leverages three aspects of sprint information – sprint attributes, issue attributes, and the developers involved in a sprint, to comprehensively characterize it for predicting both productivity and quality outcomes of the sprints. Our approach combines traditional feature extraction techniques with automated deep learning-based unsupervised feature learning techniques. We utilize methods like Long Short-Term Memory (LSTM) to enhance our feature learning process. This enables us to learn features from unstructured data, such as textual descriptions of issues and sequences of developer activities. We conducted an evaluation of our approach on two regression tasks: predicting the deliverability (i.e., the amount of work delivered from a sprint) and quality of a sprint (i.e., the amount of delivered work that requires rework). The evaluation results on five well-known open-source projects (Apache, Atlassian, Jenkins, Spring, and Talendforge) demonstrate our approach’s superior performance compared to baseline and alternative approaches.

The present work re-evaluates the evaluation method for text summarization tasks. Two state-of-the-art assessment measures e.g., Recall-Oriented Understudy for Gisting Evaluation (ROUGE) and Bilingual Evaluation Understudy (BLEU) are discussed along with their limitations before presenting a novel evaluation metric. The evaluation scores are significantly different because of the length and vocabulary of the sentences, this suggests that the primary restriction is its inability to preserve the semantics and meaning of the sentences and consistent weight distribution over the whole sentence. To address this, the present work organizes the phrases into six different groups and to evaluate “text summarization” problems, a new hybrid approach (HybridEval) is proposed. Our approach uses a weighted sum of cosine scores from InferSent’s SentEval algorithms combined with original scores, achieving high accuracy. HybridEval outperforms existing state-of-the-art models by 10-15% in evaluation scores.

Comics, particularly Japanese manga, are a powerful medium that blends images and text to convey ideas and encapsulate a unique cultural heritage. Going beyond mere entertainment, manga merges diverse styles and content deeply rooted in Japanese cultural heritage. This study utilizes computer vision analysis, with a specific focus on facial landmark detection, acknowledging the growing significance of technology in analyzing manga images. Through a comprehensive exploration of various methods, the research identifies the extended version of Bidirectional Encoder Representations from Transformers (BERT), BERT Pre-Training of Image Transformers (BEiT), model as a standout performer due to its efficiency and effectiveness. The BEiT model’s success lies in its ability to extract facial features, consequently establishing itself as a go-to solution for landmark detection on manga faces. The outcomes achieved the lowest Failure Rate compared to other landmark detection networks, with a Failure Rate of approximately 9.4% and a Mean Average Error of about 4.6 pixels. Beyond its technical accomplishments, this study carries a cultural significance, contributing to the ongoing narrative of manga in Japan.

Computer scientists often publish their source code accompanying their publications, prominently using code repositories across various domains. Despite the concurrent existence of scholarly articles and their associated official code repositories, explicit references linking the two are often missing. Traditionally, identifying whether scholarly content and code repositories pertain to the same research project requires manual inspection, a time-consuming task. This paper proposes a deep learning-based algorithm for automatically matching scholarly articles with their corresponding official code repositories. Our findings indicate that the most common linking information includes the paper title and BibTeX entries, typically found in the repository’s readme document. In this study, we employed SPECTER for vector embedding of paper and repository metadata. Utilizing these embedding representations with the Light Gradient Boosting Machine (LGBM), our method achieved an F1 score of 0.94. Moreover, combining our best model with a rule-based approach improved performance by 5.31%. This study successfully delineates a connection between academic papers and associated official code repositories, minimizing reliance on explicit bibliographic information in repositories.