Wenxiang Zheng, Guo Tang, Shixin Jiang, Liangyu Huo, Xiyuan Zhang, Jian Xie, Ming Liu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Video question answering (VideoQA) necessitates simultaneous understanding of visual and linguistic information, requiring both in-depth analysis of individual modality features and the establishment of cross-modal correlations to achieve precise reasoning. However, VideoQA models often struggle with irrelevant temporal and spatial noise due to the dense events and concepts in real-world complex video contents. Previous works reduce noise by only sampling a fixed number of visual tokens at the patch level, overlooking the variation in the required granularities of features and quantities of visual cues across different question conditions. To address these, we propose an Aggregating-then-Pruning Sampler (APSam), which diversifies feature granularities and adaptively denoises on a per-question basis. Specifically, we propose a conditional token aggregator to obtain multi-granularity visual semantics by merging similar question-relevant tokens. Then, we propose a conditional token pruner, which restricts noise tokens through a variable-capacity receptive field determined by the inputs. Experimental results show that APSam achieves significant performance on three challenging complex VideoQA datasets,i.e., AGQAv2, NExT-QA, and STAR. Further analyses reveal that the APSam also exhibits high reasoning capability and interpretability.

Jiafeng Liang, Shixin Jiang, Xuan Dong, Ning Wang, Zheng Chu, Hui Su, Jinlan Fu, Ming Liu, See-Kiong Ng, Bing Qin. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Intelligent question answering over industrial databases is a challenging task due to the multicolumn context and complex questions. The existing methods need to be improved in terms of SQL generation accuracy. In this paper, we propose a question‐aware few‐shot Text‐to‐SQL approach based on the SDCUP pretrained model. Specifically, an attention‐based filtering approach is proposed to reduce the redundant information from multiple columns in the industrial database scenario. We further propose an operator semantics enhancement method to improve the ability of identifying complex conditions in queries. Experimental results on the industrial benchmarks in the fields of electric energy and structural inspection show that the proposed model outperforms the baseline models across all few‐shot settings.

Bridge cracks can threaten transportation safety and infrastructure, leading to severe failures if not detected. Timely, accurate detection is vital for bridge safety and preventing costly repairs. Traditional methods rely on manual inspections and basic image processing, making them slow, error-prone, and unable to detect fine cracks or handle noise. Current semantic segmentation methods face challenges in accurately identifying cracks, particularly in maintaining segmentation continuity and handling localized crack features. This paper proposes an improved bridge crack segmentation algorithm based on Mask2Former, focusing on intelligent automation in bridge maintenance by preserving segmentation continuity and handling diverse crack morphologies. The Mask2Former architecture is enhanced through the introduction of a multi-scale feature fusion module, which utilizes a fusion block to process concatenated feature maps via RepBlock, strengthening feature representation, particularly for small cracks. Furthermore, a high-resolution feature processing module is integrated, employing wavelet transform convolution (WTConv) with multi-level wavelet transformations for extensive receptive field coverage. WTConv is refined with detail enhancement convolution (DEConv), which maintains parameter efficiency, along with a channel and spatial attention mechanism to better capture fine details, thus improving the segmentation of localized cracks and effectively handling background noise. Experiments on benchmark datasets demonstrate that the proposed algorithm achieves mIoU of 86.33%, F1-Score of 85.07%, and pixel accuracy of 98.70%, surpassing existing methods in fine crack detection. Ablation experiments show that both MSFM and HRFPM significantly outperform the baseline, achieving mIoU scores of 86.22% and 85.99%, and F1-scores of 84.95% and 84.65%, respectively. These results highlight the effectiveness of the proposed method in improving the robustness, accuracy, and efficiency of bridge crack segmentation, with significant potential for enhancing bridge health monitoring in transportation infrastructure.

Structural health monitoring (SHM) systems have been extensively used to ensure the operational safety of long-span bridges. Large-scale bridge structural response and loading data observed from various sensors show obvious big data characteristics. However, serious “data island” problems, which exist in the conventional SHM solutions, inevitably limit the effects of sensory data analysis and information sharing. A unified bridge SHM semantic representation model is much in demand. By taking the advantages of Semantic Web technologies, this article presents a novel model, called the bridge structure and health monitoring ontology, to achieve fine-grained modeling of bridge structures, SHM systems, sensors, and sensory data from multiple perspectives. A bridge SHM big data platform is used to demonstrate the usefulness. Several representative data accessing and rule-based reasoning scenarios are employed as to illustrate the advantages of the proposed manner.

In tennis training, beginner players can fail to return the ball when the ball moves faster than what they can react to. In this paper, we propose a new training process of mediated-timescale learning (MTL) to manipulate the incoming ball’s motion. The ball first moves in slow motion, allowing more time for players to react and develop skills. The ball then moves in faster motion, challenging players with improved skills. To evaluate MTL, we implemented it in a virtual reality (VR)-oriented tennis training system. We piloted the MTL implementations (N = 12) to study players’ physical enjoyment. We then conducted an efficacy study (N = 8) to evaluate MTL’s training effects on player’s real-world performance. We found that in comparison to real-world training, five participants improved more in hitting the sweet spot after training with MTL.

In [1], the affiliation for Y. Gao, K. Wang and J. Tian should have appeared as follows:.

In[1], the affiliation for Y. Gao, K. Wang and J. Tian should have appeared as follows:.

Fluorescence Molecular Tomography (FMT) is a powerful imaging modality for the research of cancer diagnosis, disease treatment and drug discovery. Via three-dimensional (3-D) imaging reconstruction, it can quantitatively and noninvasively obtain the distribution of fluorescent probes in biological tissues. Currently, photon propagation of FMT is conventionally described by the Finite Element Method (FEM), and it can obtain acceptable image quality. However, there are still some inherent inadequacies in FEM, such as time consuming, discretization error and inflexibility in mesh generation, which partly limit its imaging accuracy. To further improve the solving accuracy of photon propagation model (PPM), we propose a novel compactly supported radial basis functions (CSRBFs)-based meshless method (MM) to implement the PPM of FMT. We introduced a series of independent nodes and continuous CSRBFs to interpolate the PPM, which can avoid complicated mesh generation. To analyze the performance of the proposed MM, we carried out numerical heterogeneous mouse simulation to validate the simulated surface fluorescent measurement. Then we performed an in vivo experiment to observe the tomographic reconstruction. The experimental results confirmed that our proposed MM could obtain more similar surface fluorescence measurement with the golden standard (Monte-Carlo method), and more accurate reconstruction result was achieved via MM in in vivo application.

Bioluminescence tomography (BLT) is a powerful non-invasive molecular imaging tool for in vivo studies of glioma in mice. However, because of the light scattering and resulted ill-posed problems, it is challenging to develop a sufficient reconstruction method, which can accurately locate the tumor and define the tumor morphology in three-dimension. In this paper, we proposed a novel Gaussian weighted Laplace prior (GWLP) regularization method. It considered the variance of the bioluminescence energy between any two voxels inside an organ had a non-linear inverse relationship with their Gaussian distance to solve the over-smoothed tumor morphology in BLT reconstruction. We compared the GWLP with conventional Tikhonov and Laplace regularization methods through various numerical simulations and in vivo orthotopic glioma mouse model experiments. The in vivo magnetic resonance imaging and ex vivo green fluorescent protein images and hematoxylin-eosin stained images of whole head cryoslicing specimens were utilized as gold standards. The results demonstrated that GWLP achieved the highest accuracy in tumor localization and tumor morphology preservation. To the best of our knowledge, this is the first study that achieved such accurate BLT morphological reconstruction of orthotopic glioma without using any segmented tumor structure from any other structural imaging modalities as the prior for reconstruction guidance. This enabled BLT more suitable and practical for in vivo imaging of orthotopic glioma mouse models.