Jiujiang Guo, Zhengliang Guo, Kai Wang, Meiyang Wang, Dehua Peng, Shaozu Yuan, Chengyin Hu, Shuan Ai, Yiwei Wei. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

The widespread use of online geoinformation platforms, such as Google Earth Engine (GEE), has produced numerous scripts. Extracting domain knowledge from these crowdsourced scripts supports understanding of geoprocessing workflows. Small Language Models (SLMs) are effective for semantic embedding but struggle with complex code; Large Language Models (LLMs) can summarize scripts, yet lack consistent geoscience terminology to express knowledge. In this paper, we propose Geo-CLASS, a knowledge extraction framework for geospatial analysis scripts that coordinates large and small language models. Specifically, we designed domain-specific schemas and a schema-aware prompt strategy to guide LLMs to generate and associate entity descriptions, and employed SLMs to standardize the outputs by mapping these descriptions to a constructed geoscience knowledge base. Experiments on 237 GEE scripts, selected from 295,943 scripts in total, demonstrated that our framework outperformed LLM baselines, including Llama-3, GPT-3.5 and GPT-4o. In comparison, the proposed framework improved accuracy in recognizing entities and relations by up to 31.9% and 12.0%, respectively. Ablation studies and performance analysis further confirmed the effectiveness of key components and the robustness of the framework. Geo-CLASS has the potential to enable the construction of geoprocessing modeling knowledge graphs, facilitate domain-specific reasoning and advance script generation via Retrieval-Augmented Generation (RAG).

Contrastive graph clustering is an advanced technology in the field of cluster analysis. By leveraging graph neural networks and contrastive learning paradigm, it enables the coupling of topological structure and node semantic information for attributed graph networks. Graph augmentation and positive sample selection are two essentials of contrastive graph clustering. However, existing graph augmentation methods tend to disrupt the cluster structures, and most positive sample selectors suffer from the false negative sample problem. In this paper, we propose a Topological and Semantic Contrastive Graph Clustering (TSCGC) model consisting of three learning components. The representation learning component augments original graph using Ricci curvature to preserve the cluster structure, and introduces hypergraph view to capture high-order relationships. Graph and hypergraph convolutional networks are used to encode the triple-view embeddings. Meanwhile, we develop a dual contrastive learning component to extract the topological and semantic information. To reduce the number of false negatives, it utilizes K-means to generate pseudo cluster labels to guide the selection of positive samples. The self-supervised learning component is leveraged to align the three graph views. The final clustering results are obtained by performing K-means on the aligned embeddings. We demonstrated the effectiveness by comparing the performance of TSCGC with 13 clustering baselines on six real-world networks. Ablations verified the validity of key components and the impact of parameter settings were also analyzed. We further applied TSCGC to identify the function types of 10,370 buildings in ShenZhen City, China based on multi-source geospatial data. It achieved the highest accuracy and exhibit significant potential in handling complex network structures and high-dimensional node features. The code is available at: https://github.com/ZPGuiGroupWhu/TSCGC .

As the most typical graph clustering method, spectral clustering is popular and attractive due to its remarkable performance, easy implementation, and strong adaptability. Classical spectral clustering measures the edge weights using pairwise Euclidean similarity and resolves the optimal graph partitioning by relaxing the constraints of indicator matrix and decomposing the Laplacian matrix. However, Euclidean similarity might cause skew graph cuts when handling non-spherical clusters, and the relaxation strategy introduces information loss. Meanwhile, spectral clustering requires specifying the number of clusters, which is difficult to determine without enough prior knowledge. In this work, we propose a greedy-optimized scheme for resolving the indicator matrix using path-based similarity and degree descent criterion. It yields an indicator matrix with strictly binary entries without destructive relaxation and discretization steps. Path-based similarity can enhance the intra-cluster associations of arbitrary-shaped clusters, while degree descending is theoretically proven to be the best order to minimize our proposed objective function MeanCut. Moreover, we define a density gradient factor to separate clusters with fuzzy boundaries, and develop a fast maximum spanning tree algorithm to improve the scalability of similarity calculation. The effectiveness of MeanCut has been demonstrated on synthetic datasets and real-world benchmarks. By fusing multi-view image features, MeanCut outperforms cutting-edge subspace clustering methods in face recognition. The code is available at:https://github.com/ZPGuiGroupWhu/MeanCut.

Boundary point detection aims to outline the external contour structure of clusters and enhance the inter-cluster discrimination, thus bolstering the performance of the downstream classification and clustering tasks. However, existing boundary point detectors are sensitive to density heterogeneity or cannot identify boundary points in concave structures and high-dimensional manifolds. In this work, we propose a robust and efficient boundary point detection method based on Local Direction Dispersion (LoDD). The core of boundary point detection lies in measuring the difference between boundary points and internal points. It is a common observation that an internal point is surrounded by its neighbors in all directions, while the neighbors of a boundary point tend to be distributed only in a certain directional range. By considering this observation, we adopt density-independent K-Nearest Neighbors (KNN) method to determine neighboring points and design a centrality metric LoDD using the eigenvalues of the covariance matrix to depict the distribution uniformity of KNN. We also develop a grid-structure assumption of data distribution to determine the parameters adaptively. The effectiveness of LoDD is demonstrated on synthetic datasets, real-world benchmarks, and application of training set split for deep learning model and hole detection on point cloud data. The datasets and toolkit are available at:https://github.com/ZPGuiGroupWhu/lodd.

The core of image-text retrieval is to accurately measure the similarity between different modalities in a unified representation space. However, compared to textual descriptions of a certain perspective, the visual modality has more semantic variations. Therefore, images are usually associated with multiple textual captions in databases. Although popular symmetric embedding methods have explored numerous modal interaction approaches, they often learn toward outputting the average representation of multiple semantic variations within image embeddings. Consequently, information entropy in embeddings is increased, resulting in redundancy and decreased accuracy. In this work, we propose a Dynamic Visual Semantic Sub-Embeddings framework (DVSE) to reduce the information entropy. Specifically, we obtain a set of heterogeneous visual sub-embeddings through dynamic orthogonal constraint loss. To encourage the generated candidate image embeddings to capture various semantic variations, we construct a mixed distribution and employ a variance-aware weighting loss to assign different weights to the optimization process. In addition, we develop a Fast Re-ranking strategy (FR) to efficiently evaluate the retrieval results and enhance the performance. We compare the performance with existing set-based method using five image feature encoders and three text feature encoders on three benchmark datasets: MSCOCO, Flickr30K and CUB Captions. We also show the role of different components by ablation studies and perform a sensitivity analysis of the hyperparameters. The qualitative analysis of visualized bidirectional retrieval and attention maps further demonstrates the ability of our method to encode semantic variations.

Individual driving final destination prediction supports location-based services such as personalized service recommendations, traffic navigation, and public transport dispatching. However, real-time destination prediction is challenging due to the complexity of temporal dependencies, and the strong influence of travel spatiotemporal semantics and spatial correlations. Besides temporal context, the nearby urban functionalities of traveling zones and departure regions, and the crucial positions on the road network where trajectory points located would reflect the travel intentions of drivers. However, these spatial factors are rarely considered in existing studies. To fill this gap, we propose a real-time individual driving destination prediction model LSI-LSTM based on an attention-aware Long Short-Term Memory (LSTM) by taking Location Semantics and Location Importance of trajectory points into account. More specifically, a trajectory location semantics extraction method (t-LSE) enriches feature description with prior knowledge for implicit travel intentions learning. t-LSE represents urban functionality through Points of Interest (POIs) using Term Frequency-Inverse Document Frequency (TF-IDF). Meanwhile, a novel trajectory spatial attention mechanism (t-SAM) captures the trajectory points that strongly correlate to candidate destinations based on the location importance inferred from the driving status, i.e., turning angle, driving speed, and traveled distance. Comparative experiments with three baseline methods, i.e., Hidden Markov Model, Random Forest, and LSTM, demonstrate significant prediction accuracy improvements of LSI-LSTM on four individual trajectory datasets. Further analyses validate the effectiveness of the proposed semantic extraction method and attention mechanism, and also discuss the factors that may affect the prediction results.

Spatial clustering is a widely used data mining method for discovery of spatial aggregation pattern. However, existing methods often neglect scale dependence, impeding the full recognition of point patterns and the detection of hierarchical spatial structures. Spatial clustering is scale dependent and linked to the size of analysis unit as well as the hierarchy of visual cognition. Therefore, this paper proposes a novel multi-scale grid clustering (MSGC) algorithm, which fuses dual scale factors, i.e., analytical scale and visual scale that sequentially integrates multi-analytical-scale clustering (MASC) and multi-visual-scale clustering (MVSC). MASC generates multi-granularity grids to transform the analytical scales, and MVSC extracts multi-level clusters to express the hierarchy of visual cognition. Comparative experiments validated the proposed algorithm against the classical Density-based Spatial Clustering of Applications with Noise (DBSCAN) and WaveCluster algorithms on both synthetic and real-world geographic datasets. The results demonstrate that MSGC can generate multi-scale clusters for increased understanding of the spatial aggregation patterns and hierarchical structures of geographic entities. Moreover, it can eliminate noise adaptively and effectively identify clusters with arbitrary shapes. Due to the nature of grid clustering, the low computational complexity enables near real-time visual analytics and efficient point pattern mining on large spatial datasets.

With increasing point of interest (POI) datasets available with fine-grained spatial and temporal attributes, space–time Ripley’s K function has been regarded as a powerful approach to analyze spatiotemporal point process. However, space–time Ripley’s K function is computationally intensive for point-wise distance comparisons, edge correction and simulations for significance testing. Parallel computing technologies like OpenMP, MPI and CUDA have been leveraged to accelerate the K function, and related experiments have demonstrated the substantial acceleration. Nevertheless, previous works have not extended optimization of Ripley’s K function from space dimension to space–time dimension. Without sophisticated spatiotemporal query and partitioning mechanisms, extra computational overhead can be problematic. Meanwhile, these researches were limited by the restricted scalability and relative expensive programming cost of parallel frameworks and impeded their applications for large POI dataset and Ripley’s K function variations. This paper presents a distributed computing method to accelerate space–time Ripley’s K function upon state-of-the-art distributed computing framework Apache Spark, and four strategies are adopted to simplify calculation procedures and accelerate distributed computing respectively: (1) spatiotemporal index based on R-tree is utilized to retrieve potential spatiotemporally neighboring points with less distance comparison; (2) spatiotemporal edge correction weights are reused by 2-tier cache to reduce repetitive computation in L value estimation and simulations; (3) spatiotemporal partitioning using KDB-tree is adopted to decrease ghost buffer redundancy in partitions and support near-balanced distributed processing; (4) customized serialization with compact representations of spatiotemporal objects and indexes is developed to lower the cost of data transmission. Based on the optimized method, a web-based visual analytics framework prototype has been developed. Experiments prove the feasibility and time efficiency of the proposed method, and also demonstrate its value on promoting applications of space–time Ripley’s K function in ecology, geography, sociology, economics, urban transportation and other fields.