KJCKorea
Journal Central

Laser welding applies a high-energy-density beam to the material, involving complex physical phenomena, and the measured optical signals exhibit non-stationary characteristics. Defect-induced signal variations appear not only as amplitude changes but also as changes in frequency composition and baseline shifts, and they rarely form fixed patterns corresponding one-to-one with defect types. Accordingly, a model is required that can be trained using only normal data while jointly learning local time-domain patterns and time–frequency characteristics. The proposed model represents both local temporal patterns and time–frequency variations through a hybrid time–frequency embedding combining a multi-scale dilated 1D convolution branch with a Morlet wavelet-based wavelet transform branch, and learns global dependencies among time steps via temporal self-attention. The anomaly detection performance was evaluated through synthetic anomaly injection experiments using various condition combinations and varying anomaly intensities. For actual defects, the model achieved AUROC of 0.925 and 0.941, AUPRC of 0.923 and 0.931, and recall of 0.901 and 0.912 for localized over-welding, corresponding to localized signal variation, and gap defects, corresponding to signal variation over a wide interval, respectively.

The rapid growth of online news and social media has increased the importance of multimodal sentiment analysis that jointly leverages textual and visual information. However, existing studies have mainly focused on social media, online reviews, and video data, and have not sufficiently addressed modality-level sentiment inconsistency, image noise, and uncertainty in weakly supervised labels in news articles. To address these issues, this study constructs CNN-MMSA, a weakly supervised multimodal news sentiment dataset consisting of 8,273 CNN news samples, and proposes a PMIG-Net-based sentiment analysis framework. PMIG-Net utilizes BERT-based text representations and CLIP-based image representations, selectively adjusts the contribution of visual information through prior-guided attention fusion and image gating, and jointly learns modality-specific and final multimodal representations through hierarchical multi-task learning. On CNN-MMSA, PMIG-Net achieved an Accuracy of 0.7164, a Macro-F1 of 0.6732, and a Macro-Recall of 0.6828. Compared with adapted baselines that reimplement the core fusion concepts of TFN, MulT, MAG-BERT, and Self-MM, the proposed model showed relatively stronger performance, particularly in Macro-F1 and Macro-Recall. Ablation experiments showed clear decreases in Macro-F1 and Macro-Recall when the image gate was removed, while the other components exhibited different trade-offs across evaluation metrics. In an additional benchmark evaluation on MVSA-Single, PMIG-Single-Pass also outperformed CLIP-only, BERT-only, and BERT+CLIP Concat under a controlled experimental setting, indicating its applicability to image–text sentiment analysis in social media data. By integrating weak-label construction and selective multimodal fusion within a unified framework, this study provides a practical multimodal sentiment analysis approach for news opinion analysis and intelligent media content analysis.

Sango is an extremely low-resource language with limited machine translation resources. This study compares six English-to-Sango translation systems and examines the out-of-domain effects of LoRA-based adaptation. The systems were evaluated on the complete FLORES-200 devtest set using BLEU, chrF, and TER, supplemented by native-speaker judgments on a subset. Google Translate achieved the strongest overall performance, followed by zero-shot NLLB-200. All three LoRA-adapted conditions showed significantly lower BLEU scores than the zero-shot baseline in paired-bootstrap comparisons. A quantity- and step-matched comparison further showed that combining pivot-generated data, back-translation, and orthographic normalization outperformed an equal-sized condition using additional directly mined data, but did not recover zero-shot performance. Exploratory native-speaker evaluations also identified substantial semantic distortions, including unrelated religious-domain content, in some adapted outputs. Importantly, because the present experiments were conducted under a relatively small-scale adaptation setting, the observed pattern should not be assumed to generalize to substantially larger fine-tuning datasets. Overall, the results indicate that fine-tuning on automatically mined data may reduce out-of-domain performance in extremely low-resource MT, underscoring the importance of strong zero-shot baselines, independent evaluation benchmarks, and native-speaker validation.