Bo Lv, Jingbo Sun, Jianwei Lv, Chen Tang, Shaojie Zhang, Nayu Liu, Guoxin Yu, Zihao Li, Qichao Zhang, Dongbin Zhao, Ping Luo, Yue Yu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Vehicle detection based on remote sensing images is widely used in urban traffic management and disaster rescue. RGB images, which are used more, lead to poor detection performance in low light conditions due to the imaging mechanism. At present, the main solution is to improve the detection performance in low light by fusing with infrared images. However, the current methods often overlook the impact of illumination changes on RGB images, and ignore the important role of high-frequency information for object detection, especially for low-light target detection. In this paper, we propose a Cross-modality and Cross-domain Low-light Detector (CCLDet) for low-light vehicle detection, including three improvements. First, an object illumination-aware module (OIAM) is proposed, which can adjust adaptively the weight of different modalities according to the object illumination intensity in the training phase and enables the detector to adapt to different lighting conditions. Second, we propose a visibility loss, which converts the position deviation into the illumination intensity deviation of each point in the object area. Compared with relying only on semantic information for object localization, the illumination makes the information that can be used for localization more abundant. Third, we design a cross-domain feature fusion module (CDFFM), which can enhance high-frequency features and enrich target information when low-frequency features are lost due to low light pollution. Extensive experiments on three challenging RGB-infrared objects detection datasets demonstrated the mAP and the parameter quantities of CCLDet over popular object detectors.

With the rapid advancement of satellite technology, satellite video has emerged as a key method for acquiring dynamic terrestrial information, facilitating multiple object tracking (MOT). Satellites are capable of surveying vast urban landscapes, yet the observed objects are small and dispersed among complex interference from the background, heightening the challenges in detection and association tasks for object tracking. However, current trackers often dissociate the classification task from the localization task, leading to drift in tiny object detection (TOD), and rely on prior knowledge for clue ranking, limiting model robustness. In this article, we introduce the graph-based multiclue fusion tracker (GMFTracker). Initially, we introduce a sparse sampling-based feature map correction approach to rectify the misalignment between the classification and localization feature maps. Furthermore, we developed graph neural networks (GNNs) for object relationship modeling, free from presuppositions, to tackle association challenges using relational features. GMFTracker was rigorously tested on VISO, CGSTL, and TinyPerson datasets, demonstrating its competitive performance relative to contemporary studies.

As a challenging sub-task of event argument extraction, implicit event argument extraction seeks to identify document-level arguments that play direct or implicit roles in a given event. Prior work mainly focuses on capturing direct relations between arguments and the event trigger; however, the lack of reasoning ability imposes limitations to the extraction of implicit arguments. In this work, we propose anArgument-argumentRelation-enhancedEventArgument extraction (AREA) learning framework to tackle this issue through reasoning in event frame-level scope. The proposed method leverages related arguments of the expected one as clues, and utilizes such argument-argument dependencies to guide the reasoning process. To bridge the distribution gap between oracle knowledge used in the training phase and the imperfect related arguments in the test stage, we introduce a conventional knowledge distillation strategy to drive a final model that can work without extra inputs by mimicking the behaviour of a well-informed teacher model. In addition, considering that conventional knowledge distillation methods transfer knowledge individually, we integrate it with a novel relational knowledge distillation mechanism to explicitly capture the structural mutual argument-argument relation. Moreover, since the training process is not compatible with the real situation, a curriculum learning method is further introduced to make the training process smoother. Experimental results demonstrate that the learning framework obtains state-of-the-art performance on the RAMS and Wikievents datasets. Ablation study and further discussion also show it could handle long-range dependency and implicit argument problems effectively.

Many recent sentence-level event detection efforts focus on enriching sentence semantics, e.g., via multi-task or prompt-based learning. Despite the promising performance, these methods commonly depend on label-extensive manual annotations or require domain expertise to design sophisticated templates and rules. This paper proposes a new paradigm, named dialogue-based explanation, to enhance sentence semantics for event detection. By saying dialogue-based explanation of an event, we mean explaining it through a consistent information-intensive dialogue, with the original event description as the start utterance. We propose three simple dialogue generation methods, whose outputs are then fed into a hybrid attention mechanism to characterize the complementary event semantics. Extensive experimental results on two event detection datasets verify the effectiveness of our method and suggest promising research opportunities in the dialogue-based explanation paradigm.

Event detection (ED) identifies and classifies event triggers from unstructured texts, serving as a fundamental task for information extraction.Despite the remarkable progress achieved in the past several years, most research efforts focus on detecting events from formal texts (e.g., news articles, Wikipedia documents, financial announcements).Moreover, the texts in each dataset are either from a single source or multiple yet relatively homogeneous sources.With massive amounts of user-generated text accumulating on the Web and inside enterprises, identifying meaningful events in these informal texts, usually from multiple heterogeneous sources, has become a problem of significant practical value.As a pioneering exploration that expands event detection to the scenarios involving informal and heterogeneous texts, we propose a new large-scale Chinese event detection dataset based on user reviews, text conversations, and phone conversations in a leading e-commerce platform for food service.We carefully investigate the proposed dataset's textual informality and multi-source heterogeneity characteristics by inspecting data samples quantitatively and qualitatively.Extensive experiments with state-of-the-art event detection methods verify the unique challenges posed by these characteristics, indicating that multisource informal event detection remains an open problem and requires further efforts.Our benchmark and code are released at https: //github.com/myeclipse/MUSIED.

Abstract Hazy image obtained in the low‐light environment has the characteristics of low contrast, non‐uniform illumination, color cast and much noise. In this paper, a method is put forward which can be properly applied to recover low‐light hazy images. The original image is first decomposed into glow layer and haze layer with a modified color channel transformation for glow artifacts and color balanced. A new light segmentation function is proposed next by using gamma correction of channel difference and setting threshold levels to determine if the pixel belongs to light source regions. Then the ambient illuminance map is estimated using maximum reflectance prior to computing the atmosphere light in the light and non‐light regions. Finally, a novel nonlinear image depth estimation model is established to build the relationship between the image depth map and three image features including luminance, saturation and gradient map for the light areas. The experimental results prove that the dehazing algorithm is reliable for removing haze and glow artifacts of active light sources, reducing much noise and improving the visibility.