Yifei Li, Weidong Guo, Lingling Zhang, Rongman Xu, Muye Huang, Hui Liu, Lijiao Xu, Yu Xu, Jun Liu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Hang Yan, Fangzhi Xu, Rongman Xu, Yifei Li, Jian Zhang, Haoran Luo, Xiaobao Wu, Anh Tuan Luu, Haiteng Zhao, Qika Lin, Jun Liu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Traditional knowledge graph (KG) embedding methods aim to represent entities and relations in a low-dimensional space, primarily focusing on static graphs. However, real-world KGs are dynamically evolving with the constant addition of entities, relations and facts. To address such dynamic nature of KGs, several continual knowledge graph embedding (CKGE) methods have been developed to efficiently update KG embeddings to accommodate new facts while maintaining learned knowledge. As KGs grow at different rates and scales in real-world scenarios, existing CKGE methods often fail to consider the varying scales of updates and lack systematic evaluation throughout the entire update process. In this paper, we propose SAGE, a scale-aware gradual evolution framework for CKGE. Specifically, SAGE firstly determine the embedding dimensions based on the update scales and expand the embedding space accordingly. The Dynamic Distillation mechanism is further employed to balance the preservation of learned knowledge and the incorporation of new facts. We conduct extensive experiments on seven benchmarks, and the results show that SAGE consistently outperforms existing baselines, with a notable improvement of 1.38% in MRR, 1.25% in H@1 and 1.6% in H@10. Furthermore, experiments comparing with fixed dimensions methods show that SAGE achieves optimal performance on every snapshot, demonstrating the importance of adaptive embedding dimensions in CKGE.

Charts are high-density visualization carriers for complex data, serving as a crucial medium for information extraction and analysis. Automated chart understanding poses significant challenges to existing multimodal large language models (MLLMs) due to the need for precise and complex visual reasoning. Current step-by-step reasoning models primarily focus on text-based logical reasoning for chart understanding. However, they struggle to refine or correct their reasoning when errors stem from flawed visual understanding, as they lack the ability to leverage multimodal interaction for deeper comprehension. Inspired by human cognitive behavior, we propose ChartSketcher, a multimodal feedback-driven step-by-step reasoning method designed to address these limitations. ChartSketcher is a chart understanding model that employs Sketch-CoT, enabling MLLMs to annotate intermediate reasoning steps directly onto charts using a programmatic sketching library, iteratively feeding these visual annotations back into the reasoning process. This mechanism enables the model to visually ground its reasoning and refine its understanding over multiple steps. We employ a two-stage training strategy: a cold start phase to learn sketch-based reasoning patterns, followed by off-policy reinforcement learning to enhance reflection and generalization. Experiments demonstrate that ChartSketcher achieves promising performance on chart understanding benchmarks and general vision tasks, providing an interactive and interpretable approach to chart comprehension.

Fake news data, often sampled from the same communities, results in the veracity of news being highly correlated with certain textual and visual entities. This correlation leads fake news classification models to be prone to shortcut learning, quickly overfitting by capturing only shallow spurious correlations between labels and features. Consequently, neural networks trained on such data suffer from poor generalization and potential misclassification under distribution shifts. To address these critical challenges and enhance the robustness of fake news detection, in this paper, we propose a DIsentanglement-based Causality-awarE fake news detection method (DICE). DICE introduces a novel paradigm that moves beyond merely mitigating known correlations or relying on predefined bias categories. Specifically, DICE dynamically constructs multimodal news into a graph neural network, employing learnable node and edge mask disentanglers to effectively model and separate genuine causal relationships from spurious correlations between multimodal features and veracity labels. To reinforce this disentanglement process, we design a novel optimization framework that minimizes extrapolation risk and enforces representation orthogonality, leading to robust disentangled causal and biased representations. Extensive experiments demonstrate that DICE achieves superior performance on five large-scale fake news detection benchmarks. Additionally, our evaluation on a heavily biased fake news dataset demonstrates DICE’s strong generalization, suggesting its potential to inform a new paradigm in causal fake news detection. The code repo is available: https://github.com/mazihan880/DICE Code/.

Relation extraction (RE) aims to identify the relation between two entities within a sentence, which plays a crucial role in information extraction. Traditional supervised setting on RE does not fit the actual scenario, due to the continuous emergence of new relations and the unavailability of massive labeled examples. Continual few-shot relation extraction (CFS-RE) is proposed as a potential solution to the above situation, which requires the model to learn new relations sequentially from a few examples. Apparently, CFS-RE is more challenging than previous RE, as the catastrophic forgetting of old knowledge and few-shot overfitting on a handful of examples. To this end, we propose a novel flexible-prompt framework on pretrained language model named FPrompt-PLM for CFS-RE, which includes flexible-prompt embedding, pretrained-language understanding, and nearest-prototype learning modules. Note that two pools in FPrompt-PLM, i.e., prompt and prototype pools, are continual updated and applied for prediction of all seen relations at current time-step. The former pool records the distinctive prompt embedding in each time period, and the latter records all learned relation prototypes. Besides, three progressive stages are introduced to learn FPrompt-PLM's parameters and apply this model for CFS-RE testing, which includes meta-training, continual meta-finetuning, and testing stages. And we improve the CFS-RE loss by incorporating multiple distillation losses as well as a novel prototype-diversity loss in these stages to alleviate the catastrophic forgetting and few-shot overfitting problems. Comprehensive experiments on two widely-used datasets show that FPrompt-PLM achieves significant performance improvements over the SOTA baselines.