Yuhang Niu, Hongyuan Xu, Ciyi Liu, Bofan Wei, Jiaqi Ye, Yanlong Wen, Xiaojie Yuan. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Task-oriented semantic communication, leveraging learning-based joint source-channel coding (JSCC), has emerged as a key paradigm for low-latency, high-precision edge-assisted Internet of Things systems. However, the direct mapping of source data to continuous channel symbols in JSCC poses a great challenge in compatibility with existing digital systems. To address this, we propose a digital semantic communication scheme, i.e., an AdaptiveSemanticCompression with jointResourceAllocation andModulation (Adaptive-SCRAM) optimization scheme for multi-user image classification. This scheme, with the semantics quantized by a compressed codebook, enables the discrete semantic transmission with adaptive modulation, while achieving high accuracy and low latency with transmission resources optimized in multi-user classification task. Specifically, we first design a vector quantized-variational autoencoder-based digital JSCC framework with regional quantization, by jointly maximizing the semantic entropy and minimizing the codebook training loss with various SNRs and modulation orders considered in Rayleigh fading. Then based on the well trained end-to-end architecture, we mathematically fit the classification accuracy with respect to the effects of both compressed codebook size and received SNR under different modulation orders, providing an effective premise for the task performance optimization. Finally, we consider to maximize the overall multi-user classification accuracy under the transmission delay constraint, by optimizing the compression, modulation, power and bandwidth allocation for each user. To address the highly non-convex issue, we develop a dual-layer optimization algorithm. The outer-layer problem, which optimizes the compressed codebook size and modulation order, is solved by a cross-entropy-based learning algorithm. While for the inner-layer problem, a successive convex approximation method is used to optimize the power and bandwidth allocation. Simulation results show that our JSCC framework significantly reduces the semantic codebook size without compromising the classification accuracy, which is applicable to practical digital transmission systems. More importantly, compared to most existing comparable optimization schemes for image classification, our Adaptive-SCRAM optimization scheme with adaptive compression, modulation, and resource allocation can achieve much higher classification accuracy for multi-user tasks, while guaranteeing the transmission efficiency.

Multivariate time series forecasting has become increasingly crucial in fields such as energy and transportation. Recent research has focused on local lag relationships across variates, yielding impressive results. However, these methods typically require pre-calculating lag indicators and steps between variates based on historical data. This reliance on pre-calculated lag steps neglects the potential variability in lag steps over the historical and predicted time series. In this paper, we propose LagTS, a novel method that adaptively models lag relationships across variates in multivariate time series data and effectively leverages these relationships to improve prediction accuracy. Specifically, the proposed method extracts the lag relationships using a dedicated module. Moreover, it employs a lag relation loss to facilitate the adaptive modeling of the lag relationships. Extensive experimental results on four publicly available datasets and an industry dataset demonstrate the effectiveness of the proposed method.

Accurate forecasting of resource consumption, such as gas, is essential for efficient energy management, cost reduction, and sustainability. Time series forecasting (TSF) techniques like recurrent neural networks (RNNs), convolutional networks (TCNs), and Transformers have been employed to model complex temporal variations but face challenges in capturing long-term dependencies or inter-series relationships. In this paper, we propose a Hierarchical Contrastive Learning approach for Time Series (HCLTS) to improve gas consumption prediction for industrial and commercial users. HCLTS employs a prototype-based approach to construct positive and negative samples, leveraging industry labels and hierarchical sampling. We further adopt a hierarchical contrastive training objective to model underlying consumption patterns across industries. Extensive experiments on a real-world gas consumption dataset demonstrate that HCLTS achieves superior performance over existing methods, highlighting its potential for practical applications in diverse industrial contexts.

In multivariate time series forecasting, transformer-based methods have gained attention for their ability to capture complex dependencies and are often integrated with graph neural networks to improve forecasting performance. However, these approaches are computationally intensive. Mamba, a more computationally efficient architecture, has emerged as a potential solution. Despite its efficiency, Mamba fails to capture interactions across time and dependencies between variables. Furthermore, the graph structures used in most existing methods to model inter-series relationships primarily capture correlations between pairs of variables, neglecting the similarity of patterns among different variable series. In light of this, we propose a Tribe Graph enhanced Bidirectional Mamba for multivariate time series forecasting (TGBiMamba). This approach leverages bidirectional Mamba as the backbone to capture intra-series interactions and employs a designed tribe graph structure to capture inter-series intrinsic similarities. Extensive experiments on various public datasets and a proprietary industrial dataset highlight the superiority of our method.

Efficient robotic disassembly of end-of-life products is often impeded by inherent uncertainties in product condition and unknown internal structures. Conventional disassembly methods face challenges when adaptive exploration is required—particularly in cap-shaft disassembly, where connection mechanisms are frequently concealed. This paper proposes a novel robotic twist-and-pull disassembly strategy that integrates compliance control with reinforcement learning (RL). By enabling the robot to adapt to unknown connection geometries and systematic misalignments, the approach enhances the capabilities of robotic skill and reduces dependence on precisely pre-programmed trajectories. Experimental results confirm that the proposed strategy substantially improves robotic disassembly performance, improves RL training success rate, and demonstrates strong domain transferability, supporting its application across varied disassembly contexts.

Reinforcement learning (RL) has great potential for skill acquisition in robotic disassembly, but the process in real-world scenarios remains inefficient. In response, we present an efficient RL-based method for acquiring disassembly skills, incorporating relevant external knowledge to enhance learning performance. This method aims at a rapid and stable acquisition of disassembly skills which cable with transferability. The successful disassembly of various types of clearance-fit components highlights the practical applicability of the method. In this content, we design both control policies and reward functions according to the disassembly knowledge to guide and optimize the RL process. Our findings indicate that augmenting RL with external knowledge not only enhances learning efficiency, but also significantly improves skill performance. Comparative evaluations of disassembly skills obtained by different methods on the same tasks demonstrate that the proposed method yields skills with better transferability. Additionally, the method outlined is adaptable and can be implemented with various RL algorithms, making it a robust solution for enhancing robotic disassembly operations.

As a classical optimal trade execution algorithm, Volume Weighted Average Price (VWAP) strategy is favored by brokers. Since it is schedule-based and cannot perform well in a dynamic stock market, optimizing the traditional VWAP strategy via reinforcement learning is worth investigating. Most of the existing reinforcement learning-based execution strategies focus on formulating trading volumes or trading prices separately, ignoring the cooperation between trading volumes and trading prices. To address this issue, we propose a Multi-Agent Deep Q-Network (MADQN) trading framework to optimize the traditional VWAP strategy, which can dynamically adapt to the complex stock market and simultaneously formulate trading prices and volumes at each transaction period. Specifically, we design two different types of agents: 1) volume-driven agent for determining trading volumes at each transaction period and 2) price-driven agent for deciding trading prices. We model our stock market environment in which multiple agents participate as a fully cooperative stochastic game. The volume-driven agent and the price-driven agent take joint actions to interact with the stock market environment and then update their network respectively. We use 9-month level-2 data of eight stocks from different sectors on Shanghai Stock Exchange as experimental data. Experimental results demonstrate that our MADQN trading framework outperforms baselines in terms of several evaluation metrics.

Inspection of physical surface defects is a significant concern in many industrial areas. In railway systems, this process mainly includes the detection and classification of defects in rails and wheels, for which laser‐based optical inspection technologies have gradually been applied in the form of 2D profile measurement, benefiting from its high precision and robustness to surface conditions. However, defect classification and evaluation after the initial detection works still rely heavily on human inspectors to make maintenance suggestions. The linear nature of rails makes it possible to increase the dimension of rail measurement data from 2D to 3D by aligning 2D profiles along the rail, from which more comprehensive diagnosis information becomes available. In combination with appropriate artificial intelligence algorithms, this approach can potentially replace human‐dominated defect classification and evaluation work. This study presents a 3D model‐based railway track surface defect classification and evaluation method. A set of geometrical features are extracted from the 3D model of track surface defects to describe a distinguishable pattern for each category of defect. Multi‐class classifiers are then tested and have shown promising results on a group of artificial track surface defects, giving a systemic solution for 3D model‐based automatic track surface defect inspection.

Summary Correlation analysis is both popular and useful in a number of social networking research, particularly in the exploratory data analysis. In this paper, three well‐known and often‐used correlation coefficients, Pearson product–moment correlation coefficient, Spearman, and Kendall rank correlation coefficients, are compared from definition to application domain. Based on the characteristics of the pump's vibration dataset, the nonparametric and distribution‐free Spearman rank correlation coefficient is introduced to analyze the relationship between the pump's working state and each of the 207′880 variables. The percentage of variables and exact variables' tables with high Spearman's correlation coefficients for states I and II, states I and III, states II and III, and three states in different files are obtained respectively, which has important valuation for the future research of the unsupervised machine learning system. Copyright © 2015 John Wiley & Sons, Ltd.