Fusion Mechanism Research Articles

Dissecting the mechanism of atlastin-mediated homotypic membrane fusion at the single-molecule level

Homotypic membrane fusion of the endoplasmic reticulum (ER) is mediated by dynamin-like GTPase atlastin (ATL). This fundamental process relies on GTP-dependent domain rearrangements in the N-terminal region of ATL (ATLcyto), including the GTPase domain and three-helix bundle (3HB). However, its conformational dynamics during the GTPase cycle remain elusive. Here, we combine single-molecule FRET imaging and molecular dynamics simulations to address this conundrum. Different from the prevailing model, ATLcyto can form a loose crossover dimer upon GTP binding, which is tightened by GTP hydrolysis for membrane fusion. Furthermore, the α-helical motif between the 3HB and transmembrane domain, which is embedded in the surface of the lipid bilayer and self-associates in the crossover dimer, is required for ATL function. To recycle the proteins, Pi release, which disassembles the dimer, activates frequent relative movements between the GTPase domain and 3HB, and subsequent GDP dissociation alters the conformational preference of the ATLcyto monomer for entering the next reaction cycle. Finally, we found that two disease-causing mutations affect human ATL1 activity by destabilizing GTP binding-induced loose crossover dimer formation and the membrane-embedded helix, respectively. These results provide insights into ATL-mediated homotypic membrane fusion and the pathological mechanisms of related disease.

A high-precision fusion bathymetry of multi-channel waveform curvature for bathymetric LiDAR systems

Airborne LiDAR bathymetry (ALB) system is an attractive and efficient method for nearshore bathymetry and underwater topography mapping. To ensure and improve the measurement accuracy and reliability in various water environments, different receivers using a segmented field of view (FOV) have been designed and are implemented in ALB. These are used to obtain various echo data of multiple channels and perform bathymetry at various water depths. However, in echo waveform processing, detailed information about the water body described by the different echo waveforms is lacking, making it difficult to overcome severe waveform superposition in extremely shallow areas or to separate the weak echo of the water bottom from the noise in extremely deep areas. Therefore, we employed a novel method of multi-channel waveform fusion bathymetry (MWFB) to achieve information complementarity, gain multi-channel waveforms, and detect the robustness of water and bottom echo signals. This method uses an adaptive signal waveform extraction of a multichannel fusion mechanism to eliminate systematic and random noise, and the waveform curvature is used to reconstruct the energy–waveform curves. Iterative decomposition based on waveform curvature and energy curves was conducted to improve the correctness of waveform decomposition and the reliability of waveform components. Furthermore, the water surface and bottom peaks were detected based on the multichannel waveform curvature, which enabled acquisition of the accurate temporal positions of the peaks and achieve high-accuracy bathymetry. According to the experimental results and a comparison with reference datasets, the highest bathymetric accuracy of the MWFB method reached an RMSE of 0.22. For the different study areas, the bathymetric accuracy of the slope reached 95%, 92%, and 97% at Ganquan Island, Lingyang Reef, and Bei Island, respectively. Furthermore, the bathymetric point numbers were effectively increased and bathymetric accuracy was also improved in shallow and deep water, which illustrates the superior bathymetric performance of the MWFB method and its ability to provide a high-efficiency waveform dataset through multi-channel data fusion. The novel LiDAR bathymetric method proposed in this study can effectively achieve high-accuracy nearshore bathymetry and seafloor topographical mapping.

CASF-MNet: multi-scale network with cross attention mechanism and spatial dimension feature fusion for maize leaf disease detection

Maize is an important food crop, and accurate identification of maize leaf diseases is significant for disease control. Aiming at the complexity of the color and texture features of diseases with inter-class similarity and intra-class variability, a two-channel multi-scale network model with a cross-attention mechanism and spatial dimension feature fusion (CASF-MNet) was designed. Firstly, aiming at the complexity of color features, the HSV color Sub-network (CSNet) was designed to extract the color features of disease spots. Then, aiming at the complexity of texture features, Texture Sub-Network (TSNet) and Cross Attention mechanism (CAM) were proposed to enhance the texture features of disease spots. Finally, to better fuse color and texture features, a cross-spatial feature fusion mechanism (FFM) was proposed. In the five-fold cross-validation of the self-built dataset, the average F1-Score and average Accuracy of CASF-MNet reach 97.17% and 97.02%. In the test experiments of two public datasets, the F1-Score of CASF-MNet reaches 98.13% and 98.71%, and the Accuracy reaches 97.44% and 98.67%, which are better than the seven most advanced leaf disease identification methods. It is proved that CASF-MNet is advanced and stable, and can assist disease detection personnel in maize disease detection in field environments.

EC-Net: General image tampering localization network based on edge distribution guidance and contrastive learning

General Image Tampering Localization (GITL) is a challenging task that aims to locate tampered regions in manipulated images accurately. Despite notable achievements in recent years, existing methods still struggle with generalization and pixel-level segmentation, particularly in the face of diverse tampering forms and invisible traces. To address these issues, we propose EC-Net, a GITL network based on edge distribution guidance and contrastive learning. EC-Net decomposes the task into two phases: the coarse locating phase and the fine locating phase. In the coarse locating phase, a contrastive learning strategy is incorporated to improve the model's generalization capabilities. Additionally, an edge distribution prediction method is introduced to anticipate the global contour information of tampered regions. In the fine locating phase, we introduced a cross-scale fusion mechanism to synthesize feature information across different scales, enhancing the model's ability to discern subtle changes in tampered regions. Additionally, an edge reconstruction mechanism is introduced to restore details, thereby improving the accuracy of the final localization results. Evaluation on seven datasets shows that EC-Net significantly outperforms seven state-of-the-art methods, improving localization performance by 0.095 (F1), 0.088 (IoU), and 0.043 (AUC) compared to the second-best method on average. Additionally, we present a tampered image dataset (RLS26K) for model auxiliary training, featuring diverse real-life scenes with various tampering types, diverse styles, and rich content.

METEO-DLNet: Quantitative Precipitation Nowcasting Net Based on Meteorological Features and Deep Learning

Precipitation prediction plays a crucial role in people’s daily lives, work, and social development. Especially in the context of global climate variability, where extreme precipitation causes significant losses to the property of people worldwide, it is urgently necessary to use deep learning algorithms based on radar echo extrapolation for short-term precipitation forecasting. However, there are inadequately addressed issues with radar echo extrapolation methods based on deep learning, particularly when considering the inherent meteorological characteristics of precipitation on spatial and temporal scales. Additionally, traditional forecasting methods face challenges in handling local images that deviate from the overall trend. To address these problems, we propose the METEO-DLNet short-term precipitation prediction network based on meteorological features and deep learning. Experimental results demonstrate that the Meteo-LSTM of METEO-DLNet, utilizing spatial attention and differential attention, adequately learns the influence of meteorological features on spatial and temporal scales. The fusion mechanism, combining self-attention and gating mechanisms, resolves the divergence between local images and the overall trend. Quantitative and qualitative experiments show that METEO-DLNet outperforms current mainstream deep learning precipitation prediction models in natural spatiotemporal sequence problems.

Learnable fusion mechanisms for multimodal object detection in autonomous vehicles

AbstractPerception systems in autonomous vehicles need to accurately detect and classify objects within their surrounding environments. Numerous types of sensors are deployed on these vehicles, and the combination of such multimodal data streams can significantly boost performance. The authors introduce a novel sensor fusion framework using deep convolutional neural networks. The framework employs both camera and LiDAR sensors in a multimodal, multiview configuration. The authors leverage both data types by introducing two new innovative fusion mechanisms: element‐wise multiplication and multimodal factorised bilinear pooling. The methods improve the bird's eye view moderate average precision score by +4.97% and +8.35% on the KITTI dataset when compared to traditional fusion operators like element‐wise addition and feature map concatenation. An in‐depth analysis of key design choices impacting performance, such as data augmentation, multi‐task learning, and convolutional architecture design is offered. The study aims to pave the way for the development of more robust multimodal machine vision systems. The authors conclude the paper with qualitative results, discussing both successful and problematic cases, along with potential ways to mitigate the latter.

Molecular Mechanism of Autophagosome-Lysosome Fusion in Mammalian Cells.

In eukaryotes, targeting intracellular components for lysosomal degradation by autophagy represents a catabolic process that evolutionarily regulates cellular homeostasis. The successful completion of autophagy initiates the engulfment of cytoplasmic materials within double-membrane autophagosomes and subsequent delivery to autolysosomes for degradation by acidic proteases. The formation of autolysosomes relies on the precise fusion of autophagosomes with lysosomes. In recent decades, numerous studies have provided insights into the molecular regulation of autophagosome-lysosome fusion. In this review, an overview of the molecules that function in the fusion of autophagosomes with lysosomes is provided. Moreover, the molecular mechanism underlying how these functional molecules regulate autophagosome-lysosome fusion is summarized.

M-DETR: Multi-scale DETR for Optical Music Recognition

Optical Music Recognition (OMR) is an important way to digitize score images and has broad application prospects in fields such as the storage of music documents, music education and digital creation. As a new paradigm for object detection, DETR (detection transformer) has the ability to associate contextual information, which can be exploited to resolve the OMR task. However, the original DETR does not fit OMR well due to its high computational complexity and numerous parameters. To address the DETR defects and improve the recognition accuracy of OMR, we propose a novel multi-scale DETR (M-DETR) with a multi-scale feature fusion mechanism and improved attention mechanisms. First, a new multi-scale feature fusion mechanism is designed to let the backbone network of M-DETR get rich multi-scale information. Then, a key-region attention mechanism is incorporated based on the character that the key information is concentrated on a score image. Finally, the pre-context attention mechanism is introduced to make better use of the contextual association between recognition notes in music scores. Experiment results show that M-DETR achieves recognition accuracy of 90.6% for 7 typical small-sized notes, which is better than Faster R-CNN and YOLO v5, and the improvement rate is 10.02% compared to the original DETR algorithm. The results indicate that M-DETR is an effective way for the OMR task, which also provides a new solution for the detection of small-sized objects with contextual association.

Regulation of myogenic differentiation of myoblasts by groove structure for the design of cell culture meat scaffolds

As a meat substitute, cell cultured meat aims to simulate the composition of real meat to provide nutrient substance. The myotube formation efficiency of myoblasts in vitro is the key for simulating the structure and composition of muscle tissue. Therefore, this study is aimed to explore the most effective groove widths for myoblasts fusion and related molecular mechanism. The results of immunofluorescence showed that the grooves with a width range of 50 μm–1000 μm were beneficial to cell arrangement compared to flat structure. At the same time, the expression of myogenin (Myog) and myosin heavy chain 2 (Myh2) up-regulated. However, the cell fusion index of myoblasts was higher in grooves with a width of 100 μm, 200 μm and 300 μm. Furthermore, Cytochalasin D disrupted intracellular stress fibers and myotube formation. However, there were still many myotubes formed in the 200 μm groove with the addition of Cytochalasin D. The expression of fusion related genes was higher than the flat group. Moreover, incorporating groove structure into scaffolds of soybean protein isolate and pea protein isolate could enhance cell adhesion, activity and myogenic differentiation of myoblasts. This could provide ideas for scaffolds design of cultured meat products using groove structure.

AN OPTIMIZED RESIDUAL NETWORK FOR EMOTION RECOGNITION BASED ON A MULTI-FEATURES FUSION TECHNOLOGY AND ELECTROENCEPHALOGRAPHY SIGNALS

Presently, a significant number of students enrolled in colleges and universities encounter mental health challenges, including academic stress and difficulties in interpersonal relationships. These factors contribute to a decline in mental well-being and necessitate prompt intervention and assistance. The automation of mental health identification in students greatly benefits from the implementation of intelligent emotion recognition systems. Through the analysis of students’ emotional states, educational institutions can gain a deeper understanding of students’ mental well-being, enabling them to promptly identify and address any issues and offer more complete care and support to students. This paper proposes a method for automated emotion recognition that is driven by data. The method is developed based on the provided background information. The proposed approach incorporates a feature fusion mechanism and an attention mechanism into the ResNet-34 model. This integration enhances the model’s capability to analyze intricate details and subsequently improves its classification performance when applied to electroencephalography (EEG) signals. This paper introduces several key innovations. First, an optimization technique is applied to the input component of the model, enabling multifeature fusion. Second, an attention mechanism is incorporated after the residual network module, enabling the model to prioritize parts that contribute to classification and enhance feature extraction. Finally, the network parameters are optimized using both softmax loss and center loss functions. The findings from the analysis of the sentiment EEG public data SJTU Emotion EEG Dataset (SEED) indicate that the proposed sentiment recognition approach not only enhances the classification performance of the model on the sentiment EEG data but also improves the stability of the results. This paper presents a novel approach that enables automatic and efficient recognition of students’ emotions on commonly used platforms. The findings of this study hold significant implications for mental health assessment and detection in real-life production settings, offering substantial reference value in this domain.

Neural network based on convolution and self-attention fusion mechanism for plant leaves disease recognition

In agriculture, early and timely detection and identification of plant disease categories can help growers take timely countermeasures. The use of deep learning techniques for plant disease category detection prevents further spread of the disease and helps to prevent crop production losses. In this paper, Based on the Next-Vit neural network model, we proposed a lightweight neural network CAST-Net based on the combination of convolution and self-attention, and we adopted self-distillation based on this model to achieve increased accuracy in classifying plant leaf diseases while reducing the number of model parameters and flops. Our model and method achieved 98.4% accuracy on the tomato subset of the data-enhanced PlantVillage dataset, a 4.9% improvement over the Next-Vit model, and 99.0% accuracy on the full PlantVillage set, a 6.9% improvement over the Next-Vit model. We also propose a new dynamic learning rate function that is applied to the training phase to prevent the loss from reaching the optimal value. The results show that our model and method have higher accuracy, fewer parameters, shorter training time and lower computational complexity than existing models.

Multi-scale surface defect detection method for bottled products based on variable receptive fields and Gather–Distribute feature fusion mechanism

The inspection of visual quality represents a crucial step in the production process of bottled products. Numerous machine vision methodologies have demonstrated proficient identification of significant defects on bottle surfaces in well-controlled imaging environments. However, the actual production scenario introduces a myriad of surface defect types on bottled products, exhibiting diverse shapes, with the majority of defect instances characterized by relatively small individual areas. Faced with such diverse and small-sized defects, traditional convolutional neural networks (CNNs) encounter limitations due to the utilization of fixed-size convolutional kernels. This choice results in a constrained receptive field, hindering the capture of sufficiently effective contextual information. Additionally, pooling operations within traditional CNNs lead to a substantial reduction in feature map dimensions, causing excessive blurring or outright neglect of small-sized defects. Consequently, this detrimentally impacts the accuracy of surface defect detection and recognition. This study proposes a novel multi-scale defect detection model incorporating a variable receptive field and Gather–Distribute feature fusion mechanism to overcome limitations of traditional CNNs. Nine different scale defect images of bottled products, including worn, wrinkles, joint markings and etc, were collected and enhanced to construct a surface defect detection dataset for the actual production of bottled products. Enhancements to convolutional layers and the C2f module improve feature extraction for different defect shapes and sizes. Integration of the Gather–Distribute feature fusion module (GDFFM) reduces feature information loss and enhances shallow-layer feature utilization. An efficient detection head (E-Detect) with ”parameter sharing” is proposed to reduce computational complexity and improve detection speed without experiencing significant accuracy loss. Experimental results demonstrate that the model’s superiority over advanced defect detection algorithms across various categories. Notably, it achieves accuracies of 89.9%, 55.6%, 89.4%, and 75% for challenging defects like small worn, subtle wrinkles, inclined external cover, and misaligned labels, with improvements of 3.2%, 7%, 2%, and 11.8% over the baseline model, respectively. The model’s mean Average Precision (mAP) also increases by 2.7%.

Object Detection in Remote Sensing Images Based on Adaptive Multi-Scale Feature Fusion Method

Multi-scale object detection is critical for analyzing remote sensing images. Traditional feature pyramid networks, which are aimed at accommodating objects of varying sizes through multi-level feature extraction, face significant challenges due to the diverse scale variations present in remote sensing images. This situation often forces single-level features to span a broad spectrum of object sizes, complicating accurate localization and classification. To tackle these challenges, this paper proposes an innovative algorithm that incorporates an adaptive multi-scale feature enhancement and fusion module (ASEM), which enhances remote sensing image object detection through sophisticated multi-scale feature fusion. Our method begins by employing a feature pyramid to gather coarse multi-scale features. Subsequently, it integrates a fine-grained feature extraction module at each level, utilizing atrous convolutions with varied dilation rates to refine multi-scale features, which markedly improves the information capture from widely varied object scales. Furthermore, an adaptive enhancement module is applied to the features of each level by employing an attention mechanism for feature fusion. This strategy concentrates on the features of critical scale, which significantly enhance the effectiveness of capturing essential feature information. Compared with the baseline method, namely, Rotated FasterRCNN, our method achieved an mAP of 74.21% ( 0.81%) on the DOTA-v1.0 dataset and an mAP of 84.90% (+9.2%) on the HRSC2016 dataset. These results validated the effectiveness and practicality of our method and demonstrated its significant application value in multi-scale remote sensing object detection tasks.

Multi-view graph pooling with coarsened graph disentanglement

Multi-view graph pooling utilizes information from multiple perspectives to generate a coarsened graph, exhibiting superior performance in graph-level tasks. However, existing methods mainly focus on the types of multi-view information to improve graph pooling operations, lacking explicit control over the pooling process and theoretical analysis of the relationships between views. In this paper, we rethink the current paradigm of multi-view graph pooling from an information theory perspective, subsequently introducing GDMGP, an innovative method for multi-view graph pooling derived from the principles of graph disentanglement. This approach effectively simplifies the original graph into a more structured, disentangled coarsened graph, enhancing the clarity and utility of the graph representation. Our approach begins with the design of a novel view mapper that dynamically integrates the node and topology information of the original graph. This integration enhances its information sufficiency. Next, we introduce a view fusion mechanism based on conditional entropy to accurately regulate the task-relevant information in the views, aiming to minimize information loss in the pooling process. Finally, to further enhance the expressiveness of the coarsened graph, we disentangle the fused view into task-relevant and task-irrelevant subgraphs through mutual information minimization, retaining the task-relevant subgraph for downstream tasks. We theoretically demonstrate that the performance of the coarsened graph generated by our GDMGP is superior to that of any single input view. The effectiveness of GDMGP is further validated by experimental results on seven public datasets.

A Two-Stage Selective Fusion Framework for Joint Intent Detection and Slot Filling.

Spoken language understanding (SLU) is the core of the speech-centric human-robot interaction system, which mainly involves intent detection and slot filling. The recent SLU research focuses on the joint modeling of the two tasks due to their correlation. Furthermore, the slot information consists of slot position and slot type. Although the slot types are semantically related to the intent, the slot positions of the same intent may vary a lot in different utterances due to the diversity of spoken language. Thus, the conventional one-stage slot filling task may introduce unrelated information for slot position prediction in the slot-intent interaction of the joint modeling. Therefore, we propose a novel two-stage selective fusion framework for joint intent detection and slot filling. Unlike the previous one-stage framework, the proposed framework decomposes the slot filling into two stages, i.e., the slot proposal and slot classification. The slot proposal network consisting of BERT and bidirectional long short-term memory (Bi-LSTM)-conditional random field (CRF) predicts the slot positions. Instead of the tokenwise fusion in the existing methods, the slot-intent feature fusion is only performed in the slot classification. A selective fusion mechanism is designed to facilitate the slot-intent interaction within each slot candidate for more accurate slot-type classification. Experiments on five standard benchmarks (i.e., ATIS, SNIPS, MixATIS, MixSNIPS, and DSTC4) show that the proposed framework achieves the best performance in comparison with several state-of-the-art methods.

A Sequential-Multi-Decision Scheme for WiFi Localization Using Vision-Based Refinement

Currently, most mobile devices have WIFI and camera modules to locate their position. However, there are two main challenges in large, highly similar indoor environments (localization accuracy and localization time). Aiming to balance these problems, we propose a sequential-multi-decision integrated system that combines WIFI and vision to acquire users' locations. This system has two phases: sequential fusion localization and adaptive multi-decision fusion localization. The former employs WIFI-based localization first, then image-based localization and fusion localization are used within the constraints of WIFI-based localization. In the WIFI-based localization phase, the gaussian process regression (GPR) model is used to construct a WIFI indoor map. Subsequently, we propose to apply the hybrid whale optimization algorithm (HWOA) to WIFI-based localization to improve its accuracy and stability. The latter uses an adaptive multi-decision fusion mechanism that integrates WIFI-based localization, image-based localization, and fusion localization to obtain the users' location finally. The experiments show the effectiveness of HWOA applied to WIFI-based localization. We also experimentally evaluate the proposed fusion algorithm with other state-of-the-art fusion algorithms (e.g., accuracy and time) in a real environment (an area larger than 10,000 <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$m^{2}$</tex-math></inline-formula> ). The experimental results show that the proposed fusion system is competitive.

A Hierarchical Deep Learning Architecture for Robust Intrusion Detection in Time-Evolving Security Landscapes

Abstract: This paper introduces a novel architecture for Intrusion Detection Systems (IDS), designed to enhance resilience against adversarial attacks by integrating conventional machine learning (ML) models with Deep Learning (DL) models. The proposed system, termed DLL-IDS, comprises three key components: a DL-based IDS, an adversarial example (AE) detector based on local intrinsic dimensionality (LID), and an ML-based IDS. Initially, a novel AE detector is developed using LID to identify potential adversarial examples. This detector serves as a crucial component in identifying suspicious inputs that may be crafted to deceive the IDS. Subsequently, the system leverages the low transferability of attacks between DL and ML models to establish a robust ML model capable of discerning the maliciousness of potential AEs. The DLL-IDS architecture operates as follows: if incoming traffic is flagged as an AE by the detector, the MLbased IDS is employed to evaluate its maliciousness; otherwise, the DL-based IDS handles the prediction. By integrating both DL and ML approaches, the system benefits from the high prediction accuracy of DL models while exploiting the low susceptibility of ML models to adversarial attacks. Experimental results demonstrate significant enhancements in IDS prediction performance under adversarial conditions, achieving high accuracy with minimal resource consumption. This fusion mechanism effectively combines the strengths of DL and ML models, thereby bolstering the overall robustness of the IDS against sophisticated intrusion attempts.

DMFF: Deep multimodel feature fusion for building occupancy detection

The 2022 Global Status Report for Buildings and Construction (Buildings-GSR) indicates that construction activities have returned to pre-pandemic levels in most major economies, alongside more building energy consumption. To achieve the Net Zero emissions target by 2050, particularly in the post-pandemic era, accurate occupancy information is important to enhance building energy efficiency and improve occupancy comfort. While remarkable progress has been made in existing studies, they struggle to make full use of multi-sensor data to achieve high accuracy. Furthermore, there is a high expectation of the multimodel multi-temporary fusion by Transformer. In this study, we present a Transformer-based multimodal, multi-temporal feature fusion method (DMFF) for occupancy detection. To transfer domain knowledge from artificial intelligence into the building area, DMFF includes a pretrain-finetune pipeline and leverages pre-trained visual and sound models. Multiple pre-trained Transformer encoders are employed to extract features of different modalities. Then, we propose a self-attention mechanism for modality fusion to learn relationships among various sensors. Our DMFF method demonstrates superior performance on a real dataset, outperforming machine and deep learning methods (e.g., Convolutional Neural Networks, Random Forest, and Multilayer Perceptrons). Applied to a room setting, DMFF shows promising potential for building energy savings. The code and demo are accessible at https://github.com/kailaisun/multimodel_occupancy.

Deepfake detection based on cross-domain local characteristic analysis with multi-domain transformer

Deepfake videos present a significant challenge in the current media landscape. While current deepfake detection methods demonstrate satisfactory performance, there is still room for improvement in their ability to generalize and detect unseen scenarios, particularly those involving imperceptible cues. This paper introduces a novel multi-modal deepfake detection model named SpectraVisionFusion Transformer (SVFT), which incorporates spatial and frequency domain statistical artifacts to improve generalization performance. The SVFT framework uses two different backbone encoder models to take advantage of both spatial and frequency domain cues in video sequences, along with a decoder and classifier, for common cross-attention and classification, respectively. The spatial domain branch uses a convolutional transformer-based encoder to analyze facial visual features. In contrast, the frequency domain branch employs a language transformer encoder. Additionally, we introduce a weighted feature embedding fusion mechanism that integrates spectral-based statistical feature embeddings and visual cues to achieve a more comprehensive and balanced spatial-frequency feature representation. By coordinately analyzing these modalities, our model exhibits improved detection and generalization capabilities in unseen scenarios. Our proposed SVFT model achieved 92.57% and 80.63% accuracy in extensive cross-manipulation and dataset evaluation, respectively, while surpassing the performance of traditional and single-domain-based approaches.

Multi-level multilingual semantic alignment for zero-shot cross-lingual transfer learning

Recently, cross-lingual transfer learning has attracted extensive attention from both academia and industry. Previous studies usually focus only on the single-level alignment (e.g., word-level, sentence-level), based on pre-trained language models. However, it leads to suboptimal performance in downstream tasks of the low-resource language due to the missing correlation of hierarchical semantic information (e.g., sentence-to-word, word-to-word). Therefore, in this paper, we propose a novel multi-level alignment framework, which hierarchically learns the semantic correlation between multiple levels by leveraging well-designed alignment training tasks. In addition, we devise an attention-based fusion mechanism (AFM) to infuse semantic information from high levels. Extensive experiments on mainstream cross-lingual tasks (e.g., text classification, paraphrase identification, and named entity recognition) demonstrate the effectiveness of our proposed method, and also show that our model achieves state-of-the-art performance across various benchmarks compared to other strong baselines.