Yu Yan, Chunhong Zhang, Haiyu Zhao, Ziyang Zeng, Zihao Liu, Yongkang Wu, Jianzhou Diao, YiJie Chen, Shujie Wang, Zheng Hu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

In recent years, knowledge graphs have been integrated into recommender systems as item-side auxiliary information, enhancing recommendation accuracy. However, constructing and integrating structural user-side knowledge remains a significant challenge due to the improper granularity and inherent scarcity of user-side features. Recent advancements in Large Language Models (LLMs) offer the potential to bridge this gap by leveraging their human behavior understanding and extensive real-world knowledge. Nevertheless, integrating LLM-generated information into recommender systems presents challenges, including the risk of noisy information and the need for additional knowledge transfer. In this paper, we propose an LLM-based user-side knowledge inference method alongside a carefully designed recommendation framework to address these challenges. Our approach employs LLMs to infer user interests based on historical behaviors, integrating this user-side information with item-side and collaborative data to construct a hybrid structure: the Collaborative Interest Knowledge Graph (CIKG). Furthermore, we propose a CIKG-based recommendation framework that includes a user interest reconstruction module and a cross-domain contrastive learning module to mitigate potential noise and facilitate knowledge transfer. We conduct extensive experiments on three real-world datasets to validate the effectiveness of our method. Our approach achieves state-of-the-art performance compared to competitive baselines, particularly for users with sparse interactions.

Traffic flow prediction remains a critical issue in intelligent transport systems. Despite significant efforts in traffic flow modeling, existing approaches exhibit several notable limitations: (i) Most models fail to capture traffic flow similarities over long distances and extended periods; (ii) They struggle to account for spatio-temporal heterogeneity induced by varying traffic flow patterns; (iii) Due to their static modeling approach, they struggle to effectively capture the intricate spatio-temporal entanglement. To address these challenges, we propose a traffic flow prediction framework based on self-supervised learning spatio-temporal entanglement transformer(SSL-STMFormer). This framework adopts a self-supervised learning paradigm, leveraging a transformer architecture that captures richer spatio-temporal information to better represent traffic flow patterns. Specifically, a temporal attention module and a spatial attention module are employed to capture the spatio-temporal dependencies of traffic dynamics, respectively, and spatio-temporal entanglement-aware methods are introduced to allow the model to perceive spatio-temporal entanglement and thus better modelling of real traffic environments. Furthermore, to achieve adaptive spatio-temporal self-supervised learning, adaptive data augmentation is applied to the input traffic flow data, and the traffic flow prediction task is enhanced with temporal heterogeneity module and spatial heterogeneity module. Extensive experimental evaluations conducted on six publicly available real-world transportation datasets demonstrate that our method achieves substantial improvements across these datasets.

The rapid advancement of autonomous web navigation has significantly benefited from grounding pretrained Large Language Models (LLMs) as agents. However, current research has yet to fully leverage the redundancy of HTML elements for contrastive training. This paper introduces a novel approach to LLM-based web navigation tasks, called Web Element Preference Optimization (WEPO). WEPO utilizes unsupervised preference learning by sampling distance-based non-salient web elements as negative samples, optimizing maximum likelihood objective within Direct Preference Optimization (DPO). We evaluate WEPO on the Mind2Web benchmark and empirically demonstrate that WEPO aligns user high-level intent with output actions more effectively. The results show that our method achieved the state-of-the-art, with an improvement of 13.8% over WebAgent and 5.3% over the visual language model CogAgent baseline. Our findings underscore the potential of preference optimization to enhance web navigation and other web page based tasks, suggesting a promising direction for future research.

Return-conditioned supervised learning (RCSL) in offline reinforcement learning (RL) leverages Transformers to extract behavioral patterns from offline datasets for decision-making. However, it suffers from inherent limitations in comprehensively capturing multi-scale temporal relationships in historical trajectories. Moreover, its return-conditioning mechanism offers limited guidance in exploiting high-quality behavioral patterns, often resulting in suboptimal action generation during inference. To address these challenges, we propose the Advantage Decision ConvMamba (ADCM), a method that integrates multi-scale sequence modeling (MSSM) with advantage policy guidance (APG). ADCM reconstructs historical sequences through patch partitioning and employs Mamba architecture together with causal convolutions to model sparse global dependencies and dense local Markovian dependencies for behavioral pattern discovery. By incorporating relative advantage action sampling based on the Mixture-of-Experts (MoE) framework, ADCM prioritizes high-quality actions during inference, thereby reducing reliance on low-quality behavioral patterns in the dataset. We evaluate ADCM on multiple offline RL benchmarks from D4RL. Experimental results show that ADCM achieves significant improvements over baseline models, with particularly strong performance on suboptimal datasets. The code for ADCM is available at https://github.com/iTom233/ADCM.git.

Open Information Extraction (OpenIE) aims to extract structured information in the form of triples from unstructured text, serving as a foundation for various downstream NLP tasks. Despite the success of neural OpenIE models, their dependence on large-scale annotated datasets poses a challenge, particularly in low-resource settings. In this paper, we introduce a novel approach to address the low-resource OpenIE task through two key innovations: (1) we improve the quality of training data by curating small-scale, high-quality datasets annotated by a large language model (GPT-3.5), leveraging both OpenIE principles and few-shot examples to form LSOIE-g principles and LSOIE-g examples; (2) we propose CycleOIE, a training framework that maximizes data efficiency through a cycle-consistency mechanism, enabling the model to learn effectively from minimal data. Experimental results show that CycleOIE, when trained on only 2k+ instances, achieves comparable results to models trained on over 90k instances. Our contributions are further validated through extensive experiments, demonstrating the superior performance of CycleOIE and our curated LSOIE-g datasets in low-resource OpenIE as well as revealing the internal mechanisms of CycleOIE.

Liping Liu, Chunhong Zhang, Likang Wu, Chuang Zhao, Zheng Hu, Ming He, Jianping Fan. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Knowledge Tracing (KT) aims to model students' knowledge states based on their historical learning sequence, playing a critical role in online education platforms.As the performance of sequence-based KT methods continues to improve, their increasing model complexity and lack of transparency have become significant limitations.In contrast, educational theory-driven KT methods incorporate educationally meaningful features (such as question difficulty or time spent on questions) to enhance interpretability and performance.However, these models typically adopt simpler structures to reduce complexity and avoid overfitting, which limits their ability to effectively capture the sequential characteristics of learning compared to sequence-based methods.To address these limitations, this paper aims to integrate the strengths of both types of methods by proposing an Interpretable KT approach with Difficulty-Aware Attention and Selective State Space Model (ASIKT).Specifically, leveraging educational context, we design a difficulty-enhanced attention mechanism to model students' knowledge retrieval process

Self-supervised contrastive learning has demonstrated effectiveness in time series representation learning. However, existing methods still exhibit three major limitations: limited robustness to noise, missing values, and distribution shifts; reliance on a binary contrastive objective that overlooks similarity information between time series instances; and coarse-grained contrast on raw observations that fails to capture true underlying relationships between time series instances. To address these limitations, we propose a novel framework that achieves robust representation learning for time series through decomposition and fine-grained similarity-guided contrast. Specifically, we apply decomposition at the input stage to smooth noise and missing values, and the resulting disentangled trend-seasonal representations provide adaptability to distribution shifts. Furthermore, we introduce a similarity-guided contrastive loss that incorporates similarity information between instances. Additionally, our method enables fine-grained contrast through separate trend contrasting and seasonal contrasting. Extensive experiments on forecasting, anomaly detection, and classification tasks demonstrate that our framework achieves state-of-the-art performance. Further analyses validate its robustness and the effectiveness of its design.

Emotion recognition in conversations (ERC) is a pivotal component of affective computing, involving a common two-stage paradigm where pre-trained language models first extract context-independent features, followed by the encoding of contextual information and the modeling of emotional dependencies. This paradigm faces two challenges: (1) Existing methods struggle to capture both the intra-dialogue emotional continuity and the inter-dialogue semantic similarity. (2) The complexity of emotional elicitation processes gives rise to entangled dependencies, termed “latent dependencies”, which are difficult for current methods to detect and analyze. To overcome these challenges, we propose a Hybrid-Context Encoder with an Automated Latent Dependency Mining model for ERC. Specifically, we examine the emotional continuity and the semantic similarity from the standpoint of context encoders. We experimentally find that context encoders with different architectures exhibit distinct benefits. Based on these findings, we design a hybrid contextual encoding module that effectively combines the strengths of various encoders. Additionally, we design a lightweight generative module for latent dependency mining that autonomously generates a context mask, enabling the effective discovery of latent dependencies. We conduct extensive experiments on three datasets in the text modality. Our model achieves the best performance, which validates the superiority of our approach.

Social recommendations leverage social networks to augment the performance of recommender systems. However, the critical task of denoising social information has not been thoroughly investigated in prior research. In this study, we introduce a hierarchical denoising robust social recommendation model to tackle noise at two levels: 1) intra-domain noise, resulting from user multi-faceted social trust relationships, and 2) inter-domain noise, stemming from the entanglement of the latent factors over heterogeneous relations (e.g., user-item interactions, user-user trust relationships). Specifically, our model advances a preference and social psychology-aware methodology for the fine-grained and multi-perspective estimation of tie strength within social networks. This serves as a precursor to an edge weight-guided edge pruning strategy that refines the model's diversity and robustness by dynamically filtering social ties. Additionally, we propose a user interest-aware cross-domain denoising gate, which not only filters noise during the knowledge transfer process but also captures the high-dimensional, nonlinear information prevalent in social domains. We conduct extensive experiments on three real-world datasets to validate the effectiveness of our proposed model against state-of-the-art baselines. We perform empirical studies on synthetic datasets to validate the strong robustness of our proposed model.

Multimodal sentiment analysis has garnered increasing attention. The bulk of existing work in multimodal sentiment analysis primarily focuses on designing various networks to align and subsequently fuse representations from individual modalities. Contrastive learning, recognized for its intrinsic alignment capabilities, has also been extensively applied in multimodal sentiment analysis. However, current contrastive learning methods are often limited to pairwise modalities and typically perform contrastive learning prior to modality fusion, neglecting the consistency of interactions across multiple modalities. Moreover, they overlook the overall consistency within samples. To address these issues, we introduce a novel Multi-Level Contrastive Learning (MLCL) framework for multimodal sentiment analysis, composed of Uni-Modal Contrastive Learning (UMCL), Bi-Modal Contrastive Learning (BMCL) and Tri-Modal Contrastive Learning (TMCL). UMCL enhances intra-modal representations by creating positive pairs using modality-specific random dropout, while BMCL leverages the asymmetry of attention mechanisms, using two directional attentions as positive samples. TMCL aligns non-overlapping uni-modal and bi-modal representations, underscoring the complementarity of tri-modal information. The effectiveness of MLCL is demonstrated through its performance on multiple datasets. Our comprehensive experiments across multiple datasets demonstrate the superiority of the MLCL framework, which achieves new state-of-the-art performance.

Event Argument Extraction (EAE) aims to identify and extract key information such as entities, times, and locations related to specific events from text and serves as a fundamental task for many NLP applications. Recent researches have utilized large language models (LLMs) for EAE, effectively addressing the resource-intensive nature of annotating training datasets for this task. However, when performing EAE on longer texts (document-level EAE), the presence of descriptions unrelated to the events within document-level EAE can lead LLMs to identify incorrect arguments. To address this issue, we propose Simple-Hard Refining: a novel prompt framework that segments EAE into straightforward and complex extraction tasks. Based on the complexity of inference, we divide EAE task into simple-argument extraction and hard-argument extraction. By utilizing a chain of prompt to perform simple and hard argument extraction sequentially, noise introduced by irrelevant description for simple-argument extraction can be effectively alleviated. Furthermore, we explore the potential of LLMs to furnish dependable explanations for their extraction outcomes. We design an explanation-based prompting method that involves a three-step explanation process: relevant sentence extraction, argument role semantic analysis, and argument role entity localization. This method further enhances the extraction accuracy at each stage of the framework. Our experiments demonstrate that our method achieves state-of-the-art performance, surpassing various baselines that utilize LLMs for the EAE task. Ablation studies further verify the effectiveness of each stage of our framework and show the ability of our proposed approach to effectively mitigate noise. Our work contributes to the structured extraction of event argument information using LLMs.

Multimodal Sentiment Analysis (MSA) has witnessed remarkable progress and gained increasing attention in recent decade. However, current MSA methodologies primarily rely on global representations extracted from different modalities, such as the mean of all token representations, to construct sophisticated fusion networks. These approaches often overlook the valuable details present in local representations, which consist of fused representations of consecutive several tokens. Additionally, the integration of multiple local representations, and the fusion of local and global information present significant challenges. To address these limitations, we propose the Global-Local Modal (GLoMo) Fusion framework. It comprises two essential components: (i) modality-specific mixture of experts layers that integrate diverse local representations within each modality, and (ii) a global-guided fusion module that effectively combines global and local representations. The former component leverages specialized expert networks to automatically select and integrate crucial local representations from each modality, while the latter ensures the preservation of global information during the fusion process. We evaluate GLoMo on various datasets, encompassing tasks in multimodal sentiment analysis, multimodal humor detection, and multimodal emotion recognition. Extensive experiments demonstrate that GLoMo outperforms existing state-of-the-art models, validating the effectiveness of our proposed framework. Our code is publicly available at https://github.com/YetZzzzzz/GLoMo.

Time series anomaly detection is an extensively researched yet challenging task in modern academic studies. Previous studies still experience performance degradation caused by anomaly concentration. To address this challenge, we propose a framework combining Transformer and score-based generative models for time series anomaly detection, named AnoGrad. The model employs an innovative Trans-former model embedded with a variational AutoEncoder to capture temporal dependencies and distribution information. For anomaly concentration issues, we design a novel conditional stochastic differential equation based on score matching and use a new Variance Hybrid process to perturb the probability density space uniformly and controllably, improving adaptation to concentrated anomalies in diverse scenarios. Besides, we design a score estimation network that takes temporal and distributional features as conditional inputs to reduce uncertainty in the time series data generation process. Extensive experiments on multiple real-world datasets show that AnoGrad achieves competitive performance, often rivaling state-of-the-art benchmarks.

Social networks exhibit a distinct "celebrity effect" whereby influential individuals have a more significant impact on others compared to ordinary individuals, unlike other network structures such as citation networks and knowledge graphs. Despite its common occurrence in social networks, the celebrity effect is frequently overlooked by existing social recommendation methods when modeling social relationships, thereby hindering the full exploitation of social networks to mine similarities between users. In this paper, we fill this gap and propose a Celebrity-aware Graph Contrastive Learning Framework for Social Recommendation (CGCL), which explicitly models the celebrity effect in the social domain. Technically, we measure the different influences of celebrity and ordinary nodes by mining social network structure features, such as closeness centrality. To model the celebrity effect in social networks, we design a novel user-user impact-aware aggregation method, which incorporates the celebrity-aware influence information into the message propagation process. Additionally, we design a graph neural network-based framework which incorporates social semantics into the user-item interaction modeling with contrastive learning-enhanced data augmentation. The experimental results on three real-world datasets show the effectiveness of the proposed framework. We conduct ablation experiments to prove that the key components of our model benefit the recommendation performance improvement.

Multivariate time series anomaly detection is a challenging task due to the intricate intra- and inter-metric dependencies present in the data. Previous methods have mainly focused on capturing intra-metric dependency, while neglecting the utilization of the dependency information across metrics. More recent approaches have considered inter-metric dependency, but have not been able to capture the complex inter-metric dynamics accurately and adaptively. To address these challenges, we propose a novel dual-attentional multivari-ate time series anomaly detection framework. Our approach improves on the original Dot-Product Attention to robustly capture intra-metric dependency. Additionally, we introduce automatic graph structure learning and graph attention mechanism to adaptively capture inter-metric dependency and utilize it effectively. Experiments on five public datasets from different domains demonstrate that comprehensive and accurate modeling of both dependencies enables our proposed method to outperform baseline methods in accurately detecting anomalies.

Distant supervision for relation extraction aims to automatically obtain a large number of relational facts as training data, but it often leads to noisy label problem. In this paper, we propose a self-directed confidence learning based latent-label denoising method for distantly supervised relatio n extraction. Concretely, a self-directed algorithm that combines the semantic information of model prediction and distant supervision is designed to predict the confidence score of latent labels. Since this mechanism utilizes the obtained latent labels of easy examples to produce the latent labels of hard examples step by step, it is a robust and reliable learning process. Besides, it facilitates dynamic exploration of the confidence space to achieve better denoising performance. Moreover, to cope with the common imbalance problem in large corpus where the negative instances account for a much larger percentage, we introduce a discriminative loss function to solve the misclassification between non-relational and relational instances. Empirically, in order to verify the generality of the proposed denoising method, we use different neural models – CNN, PCNN and BiLSTM for representation learning. Experimental results show that our method can correct the noisy labels with high accuracy and outperform the state-of-the-art relation extraction systems.

User preference profile is generally significant to marketing strategy decisions as well as user experience improvement for mobile telecom operators. To establish preference profile, perators create a hierarchical taxonomy of preference and classify records of user browsing history on mobile internet by the taxonomy to measure user preference. However, the incompleteness of recorded browsing history makes it nontrivial to observe all the users' preferences. To complete missing preferences, recommendation based methodology is commonly exploited. Although taxonomy contains the semantic relationships between preferences, there are merely a few works that explored them for recommendation. We extend these works by clearly defining the relation types and learning relation strengths among preferences in the taxonomy, on which we propose a Taxonomy-induced Matrix Factorization (TMF) model. We perform experiments on a large dataset of user browsing data from a Chinese telecom operator. The results show that our proposed model outperforms the standard matrix factorization model. In addition, the relations learned by TMF are detailed analyzed to show their inherent effects for the inference improvement.

Churn prediction, which is to identify who are prone to abandon the subscription, is of high significance for the operators to retain the potential churners. It should be noted that, in practice, the earlier the churners are identified, the more effective strategies the operators can develop to retain them. While the earlier prediction of customer churn in telecommunication industry has not been well investigated and the predicting accuracy of previous work degrades unacceptably when the interval between observed attributes and predicted labels is prolonged. In this paper, from a different perspective, we study the effectiveness to find the churners as early as possible with the accuracy being high enough, which we define as Early Churn Prediction. The predictive performance of the proposed model, which takes time series attributes and influence of churning contacts in social network into consideration, is investigated. We evaluate the method using a 12-month-long dataset collected by one of the largest operators in China. The results show that our model significantly outperforms the previous work especially when the prediction interval is larger than 3 months.

In this paper we propose the use of a two-agent learning scheme for the management of radio resources on cellular access networks. The management is materialized by the implementation of a self-configuration system governing the setup of several parameters on each base station. The two agents have independent goals; one is trying to maximize the quality of service and the other the economic benefit. Thanks to the combined use of the fuzzy logic technique and reinforcement learning, both agents will work in a complementary mode, achieving both goals simultaneously.

Since wireless network virtualization (WNV) enables physical resources abstraction and sharing, the overall resources inefficiency can be reduced dramatically. This paper investigates a pricing-based energy efficient (EE) optimization problem for orthogonal frequency-division multiple Access (OFDMA) WNV. A typical WNV environment consists of an infrastructure provider (InP), virtual network operators (VNOs) and end users. The objective of this paper is to maximize VNOs' EE in bits per joule unit. This is achieved by allocating each VNO certain amount of power. The problem is formulated as a commercial market competition based on a pricing function. A non-cooperative game is applied and a power allocation algorithm is developed to search the Nash equilibrium which is the solution of this game. The Nash equilibrium indicates the best strategy that each VNO can employ. The performances of the proposed algorithm are obtained in a frequency selective fading environment. Evaluation results reveal the VNO adaptation of power sharing strategies and also shows the inefficiency of the Nash equilibrium.

Collaborative filtering has become one of the most used approaches to provide personalized services for users. The key of this approach is to find similar users or items using user-item rating matrix so that the system can show recommendations for users. However, most approaches related to this approach are based on similarity algorithms, such as cosine, Pearson correlation coefficient, and mean squared difference. These methods are not much effective, especially in the cold user conditions. This paper presents a new user similarity model to improve the recommendation performance when only few ratings are available to calculate the similarities for each user. The model not only considers the local context information of user ratings, but also the global preference of user behavior. Experiments on three real data sets are implemented and compared with many state-of-the-art similarity measures. The results show the superiority of the new similarity model in recommended performance.

Multi-access for multi-mode terminals is possible in heterogeneous networks environment which becomes a critical issue for transmission in parallel with meeting user quality of service (QoS) performance requirements and throughput maximization. The previous optimal resource allocation schemes, which do not consider service type, may allocate scarce radio resources inefficiently. To solve this, we propose an optimal resource allocation algorithm with distinguishing the services traffic into two classes: Delay-Constraint (DC) and Best-Effort (BE). In our work, we formulate a mathematical optimal model to support such heterogeneous service requirements in multi-radio access scenario. Then, we develop a optimal radio resource allocation algorithm that achieves the goal of maximizing the total system throughput in heterogeneous networks, while efficiently satisfying the QoS requirement for DC services traffic and fairness for BE services traffic. Simulation results show that the proposed service traffic differentiation based radio resource allocation algorithm significantly outperforms other existing schemes.

In cognitive radio (CR) networks, cross-layer design is an important issue since the behavior of one protocol could affect the performance of others. However, for the energy-constraint CR networks, the previous works mostly focus on maximizing the throughput in physical or transport layer, rather than the energy efficiency of the end-to-end transmission control protocol (TCP). In this paper, we propose a novel cross-layer scheme which takes the lower layers' parameters into consideration, e.g., signal-to-noise ratio (SNR), modulation and frame size, to improve the energy efficiency of TCP. Specifically, we use a finite state Markov channel (FSMC) model to characterize the fading channel, and solve the optimization problem by a restless bandit approach. Simulation results show that the physical and data link layer parameters affect the energy efficiency of the TCP traffic significantly and the performance can be improved by the adjustment of lower layers' parameters compared with the existing method.

In the group vertical handover (GVHO) scenario, many mobile terminals (MTs) send handover requests almost at the same time. The traditional vertical handover (VHO) schemes assume that the VHO user is coming one by one, so the current user knows the decision results of previous users, then the optimal result can be obtained. In GVHO scenario, multiple VHO decisions need to be made simultaneously, if the traditional VHO scheme was applied in this scenario, it may lead to system performance degradation or network congestion, because the decision-making MT does not know the results of other concurrent VHO users, so it may selfishly select the best networks just like in common VHO scenario. Therefore, three decision-making models for GVHO are proposed in this paper, and there performance comparisons are analyzed through numerical simulations.

This paper proposes a novel distributed QoSguaranteed multi-cell radio resource allocation (RRA) scheme for OFDMA systems, which includes two separate steps operated on two time levels: the Hungarian algorithm based intra-cell fast scheduling on the frame level and the X2 interface based inter-cell interference coordination (ICIC) algorithm performed on the super-frame level. The scheme is distributed in the sense that both algorithms are carried out by the enhanced Node B (eNB). Through intense analysis, we proved that the ICIC algorithm not only provides fine robustness against ICI but also achieves good load balancing from a new perspective. Simulation results indicate that the performance of the proposed scheme, when compared to the standard "universal reuse" approach without inter-cell coordination, is significantly better in terms of increasing both the network and cell-edge throughputs.

In this article we present a novel architecture of mobile networks referred to as mobile ubiquitous service environment (MUSE), which aims to bring users always best experiences (ABE) employing heterogeneous networks and personal area networks. We aim to provide a uniform ubiquitous service environment, which consists of detectable, representable, evaluable, expansible, and adaptable abstract features. Thus, it appears as homogeneity to the services. We aim for the joint utilization of the capabilities provided by novel terminals and heterogeneous networks. As a result of this view, we introduce two interactive environments, terminal service environment (TSE) and network service environment (NSE); which provide the MUSE capabilities to bring experiences to users. In this paper, the design principles of MUSE and the ABE concept are presented, and the QoS framework and the reconfiguration, which play key roles in MUSE, are presented in more detail.