Previous studies have proposed semantic-based approaches for code search over large-scale codebases, which has bridged the gap in understanding the semantics between natural language and source-code language. However, these studies either failed to determine an effective method for semantic representations or lacked the distinction of semantic features. In this study, we propose a novel attention-based LSTM neural network known as CSDA (Code Search based on Description Attention), which can effectively improve the code search performance. The proposed model can focus on different parts of a semantic feature when numerous aspects of a source code snippet are used as input. As opposed to assigning the same weight to different parts of the semantic vector, CSDA takes the semantics of natural language descriptions into account, so that the subtle differences hidden in the code snippet can be discriminated and associated with the corresponding queries. We compare CSDA with the existing state-of-the-art approach CODEnn, which uses a jointly embedding technique for code search. Our experimental evaluation demonstrates that CSDA outperforms previous methods and achieves superior code search performance to CODEnn, with higher success rates and mean reciprocal ranks. This study provides significant insights into the use of semantic representation methods in deep learning-based code search approaches.

Automatically extracting entity-pair interactions from biomedical literature plays an important role in promoting the development of the biomedical field. For instance, the interactions between drugs can guide patients to take drugs correctly and avoid clinical adverse drug reactions; The interactions between proteins can help researchers design therapeutic drugs and discover disease mechanisms. However, it is found that relation instances of different classes generated from the same sentence are easily classified into the same class since their context information is almost the same. To address this issue, Triplet Loss based BERT (TL-BERT) approach is proposed in this paper, where the Triplet Loss training strategy is first introduced into the biomedical relation extraction field. Triplet Loss training strategy will increase the distances between these instances generated from the same sentence but belonging to different classes, and decrease the distances between these instances generated from different sentences but belonging to the same class. As a result, our approach can classify these instances generated from the same sentence more correctly. TL-BERT was evaluated on AIMed, BioInfer, and DDI Extraction-2013 corpus, the experimental results demonstrate that the Triplet Loss training strategy can improve the performance on both Protein-Protein interactions extraction tasks and Drug-Drug interactions detection tasks.

Entity alignment (EA), aiming to match entities with the same meaning across different knowledge graphs (KGs), is a critical step in knowledge fusion. Existing EA methods usually encode the multi-aspect features of entities as embeddings and learn to align the embeddings with supervised learning. Although these methods have achieved remarkable results, two issues have not been well addressed. Firstly, these methods require pre-aligned entity pairs to perform EA tasks, limiting their applicability in practice. Secondly, these methods overlook the unique contribution of digital attributes to EA tasks when utilising attribute information to enhance entity features. In this paper, we propose a self-supervised entity alignment framework via attribute correction. Specifically, we first design a highly effective seed pair generator based on multi-aspect features of entities to solve the labour-intensive problem of obtaining pre-aligned entity pairs. Then, a novel alignment mechanism via attribute correction is proposed to address the problem that different types of attributes have different contributions to the EA task. Extensive experiments on real-world datasets with semantic features demonstrate that our framework outperforms state-of-the-art (SOTA) EA tasks.

Underwater multi-class object detection in nearshore waters is essential for intelligent cleaning operations and ecological monitoring. However, strong reflection and scattering interference, color attenuation, frequent occlusion, and non-rigid deformation often cause fine-grained information loss and feature misalignment in conventional detectors, leading to missed and false detections. To address these challenges, we propose an enhanced YOLOv11 framework integrating FasterNet and attention mechanisms. Specifically, we include FasterNet to replace the YOLOv11 baseline backbone to improve fine-grained feature preservation while reducing computational redundancy. Furthermore, a Deformable Underwater Attention Module (DUAM) is introduced to capture local texture variations and deformation-aware features, enhancing discrimination among heterogeneous categories. Additionally, a Submerged Occlusion-Aware Head (SOAH) is designed to recalibrate features based on occlusion visibility, improving the detection of small-scale and partially occluded objects in the high-resolution P2 layer. Performance gains mainly stem from the recalibration strategy and its synergy with multi-scale optimization objectives. Experiments on a nearshore underwater multi-class dataset (8610 images across 40 classes) show that the proposed method increases mAP from 66.9% to 82.3%, achieving a 15.4-point improvement over baseline YOLOv11, with superior robustness under complex backgrounds.

In the field of marine ecosystem conservation and underwater garbage management, accurately identifying diverse debris is critical to improve cleanup efficiency and advance intelligent underwater operations. However, reflection, occlusion, color attenuation, object deformation in complex underwater environment render the traditional detection models impractical for dynamic scenarios. To address these issues, this paper proposes an underwater debris recognition model based on an enhanced YOLOv11 network. The improvements are as follows:a) FasterNet is adopted as the backbone network, which balances lightweight and high-fidelity feature extraction, effectively optimizing the retention of small-object features.b) SOAH, an occlusion-aware attention mechanism is introduced to reconstruct the detection head. This enables the module to strengthen the response to occluded targets during fusion process, thereby amplifying recognition performance in complex backgrounds.c) DAttention is utilized to replace the conventional spatial attention mechanism, which makes the DUAM module adaptable to partial texture deformation and increase the recognition accuracy of heterogeneous targets. The datasets comprises 9 major categories and 40 subcategories underwater debris images, collected from natural marine environment of varying depths and lighting conditions. It virtually contains typical challenges such as aquatic particle interference and uneven illumination. Experimental results show that the advanced network boosts from 66.9% to 77.3%, a 10.4% increase was observed. This proposed methodology demonstrates high precision and stability in complex conditions, providing robust support for underwater marine debris detection.

Referring image segmentation aims to segment a target object from an image by providing a natural language expression. While recent methods have made remarkable advancements, few have designed effective deep fusion processes for cross-model features or focused on the fine details of vision. In this paper, we propose AIUnet, an asymptotic inference method that uses U2-Net. The core of AIUnet is a Cross-model U2-Net (CMU) module, which integrates a Text guide vision (TGV) module into U2-Net, achieving efficient interaction of cross-model information at different scales. CMU focuses more on location information in high-level features and learns finer detail information in low-level features. Additionally, we propose a Features Enhance Decoder (FED) module to improve the recognition of fine details and decode cross-model features to binary masks. The FED module leverages a simple CNN-based approach to enhance multi-modal features. Our experiments show that AIUnet achieved competitive results on three standard datasets.Code is available at https://github.com/LJQbiu/AIUnet.