Machine Reading Comprehension Model Based on Fusion of Mixed Attention DOI Creative Commons
Yanfeng Wang,

Ning Ma,

Zechen Guo

и другие.

Applied Sciences, Год журнала: 2024, Номер 14(17), С. 7794 - 7794

Опубликована: Сен. 3, 2024

To address the problems of insufficient semantic fusion between text and questions lack consideration global information encountered in machine reading comprehension models, we proposed a model called BERT_hybrid based on BERT hybrid attention mechanism. In this model, is utilized to separately map into feature space. Through integration Bi-LSTM, an mechanism, self-attention achieves comprehensive questions. The probability distribution answers computed using Softmax. experimental results public dataset DuReader demonstrate that improvements BLEU-4 ROUGE-L scores compared existing models. Furthermore, validate effectiveness design, analyze factors influencing model’s performance.

Язык: Английский

Enhancing Chinese abbreviation prediction with LLM generation and contrastive evaluation DOI
Jingping Liu, Xianyang Tian, Hanwen Tong

и другие.

Information Processing & Management, Год журнала: 2024, Номер 61(4), С. 103768 - 103768

Опубликована: Май 10, 2024

Язык: Английский

Процитировано

3
Research on crime motivation identification and quantitative analysis methods based on EEG signals DOI Creative Commons
Dongli Ma

Frontiers in Psychology, Год журнала: 2025, Номер 16

Опубликована: Март 18, 2025

Introduction Understanding and quantifying crime motivation is essential for developing effective interventions in criminology psychology. This research, closely aligned with quantitative psychology measurement, presents a novel approach to identifying analyzing motivations using EEG signals. Traditional methods often fail capture the intricate interplay of individual, social, environmental factors due data sparsity absence real-time adaptability. Methods In this study, we introduce Hierarchical Crime Motivation Network (HCM-Net), multi-layered framework that integrates signal analysis social temporal modeling. HCM-Net employs neural network-based individual feature encoders, graph networks interaction analysis, predictors evolution motivations. To enhance practical applicability, Dynamic Risk-Adaptive Strategy (DRAS) complements by incorporating adaptation, scenario-based simulations, targeted interventions. addresses challenges such as ethical considerations interpretability employing Shapley values attribution bias mitigation techniques. Results Experiments datasets demonstrate superior performance proposed classifying high-risk individuals compared state-of-the-art Discussion These findings highlight potential integrating advanced computational prevention psychological research.

Язык: Английский

Процитировано

0
RealExp: Decoupling correlation bias in Shapley values for faithful model interpretations DOI

Wendong Jiang,

Chih‐Yung Chang, Show-Jane Yen

и другие.

Information Processing & Management, Год журнала: 2025, Номер 62(4), С. 104153 - 104153

Опубликована: Март 23, 2025

Язык: Английский

Процитировано

0
Multimodal emotion recognition method in complex dynamic scenes DOI Creative Commons
Long Liu, Qingquan Luo, Wenbo Zhang

и другие.

Journal of Information and Intelligence, Год журнала: 2025, Номер unknown

Опубликована: Март 1, 2025

Язык: Английский

Процитировано

0
TransField: Improving transformer efficiency and performance through conditional random fields DOI Creative Commons
Nguyễn Văn Hiệu, N. Minh,

Anh Ho Quoc Thien

и другие.

ETRI Journal, Год журнала: 2025, Номер unknown

Опубликована: Май 8, 2025

Abstract Transformer architectures have become a dominant method in natural language processing, though their high parameter requirements remain challenge. This paper presents the TransField encoder, novel architecture integrating conditional random fields with transformer mechanisms to mitigate this issue. By enhancing ability capture semantic nuances sentences, encoder was evaluated across five tasks: masked modeling, machine translation, text classification, named entity recognition, and automatic speech recognition. The experimental results consistently demonstrate that transformer‐based models incorporating maintain robust performance key evaluation metrics achieve significant reduction number of parameters compared standard encoder. Although tested on smaller datasets, these findings suggest promising potential for broader applications, warranting further investigation larger‐scale datasets.

Язык: Английский

Процитировано

0
A novel CTGAN-ENN hybrid approach to enhance the performance and interpretability of machine learning black-box models in intrusion detection and IoT DOI
Houssam Zouhri, Ali Idri

Future Generation Computer Systems, Год журнала: 2025, Номер unknown, С. 107882 - 107882

Опубликована: Май 1, 2025

Язык: Английский

Процитировано

0
Deep learning-based image classification for AI-assisted integration of pathology and radiology in medical imaging DOI Creative Commons

L. He,

Lan Luan, Dan Hu

и другие.

Frontiers in Medicine, Год журнала: 2025, Номер 12

Опубликована: Июнь 2, 2025

Introduction The integration of pathology and radiology through artificial intelligence (AI) represents a groundbreaking advancement in medical imaging, providing powerful tool for accurate diagnostics the optimization clinical workflows. Traditional image classification methods encounter substantial challenges due to inherent complexity heterogeneity imaging datasets, which include multi-modal data sources, imbalanced class distributions, critical need interpretability decision-making. Methods Addressing these limitations, this study introduces an innovative deep learning-based framework tailored AI-assisted tasks. It incorporates two novel components: Adaptive Multi-Resolution Imaging Network (AMRI-Net) Explainable Domain-Adaptive Learning (EDAL) strategy. AMRI-Net enhances diagnostic accuracy by leveraging multi-resolution feature extraction, attention-guided fusion mechanisms, task-specific decoders, allowing model accurately identify both detailed overarching patterns across various techniques, such as X-rays, CT, MRI scans. EDAL significantly improves domain generalizability advanced alignment techniques while integrating uncertainty-aware learning prioritize high-confidence predictions. employs attention-based tools highlight regions, improving transparency trust AI-driven diagnoses. Results Experimental results on datasets underscore framework's superior performance, with accuracies reaching up 94.95% F1-Scores 94.85%, thereby enhancing Discussion This research bridges gap between radiology, offering comprehensive solution that aligns evolving demands modern healthcare ensuring precision, reliability, imaging.

Язык: Английский

Процитировано

0
Assessing the effectiveness of dimensionality reduction on the interpretability of opaque machine learning-based attack detection systems DOI
Houssam Zouhri, Ali Idri, Hajar Hakkoum

и другие.

Computers & Electrical Engineering, Год журнала: 2024, Номер 120, С. 109627 - 109627

Опубликована: Сен. 19, 2024

Язык: Английский

Процитировано

1
Machine Reading Comprehension Model Based on Fusion of Mixed Attention DOI Creative Commons
Yanfeng Wang,

Ning Ma,

Zechen Guo

и другие.

Applied Sciences, Год журнала: 2024, Номер 14(17), С. 7794 - 7794

Опубликована: Сен. 3, 2024

To address the problems of insufficient semantic fusion between text and questions lack consideration global information encountered in machine reading comprehension models, we proposed a model called BERT_hybrid based on BERT hybrid attention mechanism. In this model, is utilized to separately map into feature space. Through integration Bi-LSTM, an mechanism, self-attention achieves comprehensive questions. The probability distribution answers computed using Softmax. experimental results public dataset DuReader demonstrate that improvements BLEU-4 ROUGE-L scores compared existing models. Furthermore, validate effectiveness design, analyze factors influencing model’s performance.

Язык: Английский

Процитировано

0