Home Latest Articles
Latest Articles
  • Zefan Zeng, Yuehang Si, Qing Cheng, Zhong Liu
    Journal of Computer Research and Development. 2026, 63(10): 2697-2714.

    Event causality identification (ECI) aims to detect causality between events in textual contexts. Compared with sentence-level ECI which focuses on single sentences, document-level ECI (DECI) needs to process more complex contexts such as long texts and cross-sentence events. Current DECI approaches face two key challenges: learning high-quality context-aware event representation to capture event-specific causal semantic and effectively integrating structural and semantic information for causal reasoning. To address these challenges, we propose MAIGO (multi-granularity context-aware and iterative graph optimization), a novel DECI framework that achieves context-aware event representation through multi-granularity attention mechanisms and dynamically optimizes the event causality graph structure via iterative embedding refinement. MAIGO enhances event embeddings through sentence-level, document-level, and graph-level attention, capturing event-event, event-sentence, and sentence-sentence interactions to produce better representation for events. The framework further constructs an event causality graph (ECG) and jointly optimizes both the ECG structure and event embeddings via masked attention in an iterative optimization process. Extensive experiments and analysis on two benchmark datasets and a multilingual dataset demonstrate MAIGO’s effectiveness, achieving performance gains of 4.1% (causality existence identification) and 7.3% (causality direction identification) in inter-sentence ECI over state-of-the-art models. Further analysis confirms the advantages of the multi-granularity context-aware enhanced representation strategy and iterative graph optimization mechanism.

  • Li Lin, Zongsheng Li, Shuai Wang
    Journal of Computer Research and Development. 2026, 63(10): 2715-2732.

    Inventory scheduling plays a critical role in modern supply chains, with its primary objective being to balance production and demand through appropriate inventory allocation. As market demand becomes increasingly uncertain and product attributes more diverse, traditional experience-based inventory scheduling methods often struggle to effectively integrate product characteristics with demand fluctuations. Moreover, their decision outcomes lack interpretability, making it difficult to trace the rationale behind inventory schedules and reducing stakeholders’ trust. The semantic understanding and knowledge-reasoning capabilities of pre-trained large language models, when combined with abundant real-world multi-source external data and multi-category demand data, offer opportunities to enhance existing LLM-based inventory scheduling methods. To capitalize on these opportunities, LLM-based inventory scheduling decision tasks have to face the following challenges: 1) how to effectively embed the inventory scheduling workflow into a multi-step reasoning process, yielding accurate and coherent reasoning trajectories; 2) how to balance the interpretability of allocation decisions with decision accuracy, avoiding a decline in interpretability caused by excessive focus on accuracy. To this end, we propose a novel multi-step reasoning framework for inventory scheduling based on large language models, termed InvLLM, with following primary contributions: the design of a self-calibrating multi-step reasoning method based on the inventory scheduling workflow and a multi-objective balanced optimization method for inventory scheduling decisions. Additionally, to address distortion in the reasoning fine-tuning phase, a “planning-token” supervised fine-tuning approach is proposed, combined with a generative reward model, achieving a balance between high accuracy and logical consistency. Experimental evaluations on a large-scale real-world dataset provided by JD.com demonstrate that, compared with existing inventory scheduling methods, the proposed InvLLM framework increases service rate by 3.24% and improves overall inventory evaluation scores by 1.55%, while reducing inventory rate by 1.33%, thereby validating its effectiveness and practical applicability.

  • Bowen Dong, Xueli Liu, Xiaoyang Feng, Zhenzhen Mai, Wenjun Wang
    Journal of Computer Research and Development. 2026, 63(10): 2733-2751.

    Knowledge graph reasoning has attracted extensive attention because of its ability to alleviate the inherent incompleteness of knowledge graphs. However, existing approaches mainly rely on embedding models or chain-based rules, but they still suffer from limitations in interpretability and expressiveness. Existing rule learning methods are often limited to linear dependency modeling and struggle to capture multi-path interactions and hierarchical semantic relations, while traditional rule-mining methods usually follow the “enumeration–verification” paradigm, resulting in high computational costs and insufficient modeling of deep semantic correlations among rules. To address these challenges, we propose CGRL (contrastive graph rule learning), a novel method for learning graph-structured rules through contrastive learning. CGRL enhances rule learning through a threefold mechanism. First, it applies an adaptive clustering strategy to generate soft type labels for entities, introducing semantic priors that improve rule generalization. Second, it designs a multi-hop neighborhood sampling method to construct semantic subgraphs and to generate high-quality positive and negative rule instances. Third, it employs a Transformer-based encoder and a contrastive learning objective to map rule pairs into a unified embedding space, thereby jointly capturing structural dependencies and semantic consistency. Extensive experiments on six public knowledge graph completion benchmarks demonstrate that CGRL achieves substantial improvements over state-of-the-art baselines. On datasets with relatively complex relational structures, such as FB15k-237 and WN18RR, CGRL yields average gains of 6.04% in MRR and 16.26% in Hits@1, confirming its effectiveness in discovering high-quality graph-structured rules and significantly enhancing reasoning performance.

  • Tingting Li, Shouqiang Sun, Qingqing Li, Ling Wan
    Journal of Computer Research and Development. 2026, 63(10): 2752-2766.

    The recognition of negative sentiment intensity in financial public opinion on social media is of great significance for financial risk prevention and market stability. Given the prevalence of multimodal information combining text and images in financial sentiment expression, and the need for further exploration of structural correlations between modalities, we propose a multimodal negative sentiment intensity recognition model based on the collaborative enhancement of graph convolutional networks and Stacking ensemble learning. The model first employs the CLIP model to extract semantic features from text and images. Subsequently, a multimodal graph structure is constructed and graph convolutional networks are introduced to achieve structural association embedding of features, thereby enhancing the ability to capture potential inter-modal correlations. Finally, a two-layer Stacking ensemble learning framework is built to optimize the decision performance of sentiment intensity recognition by integrating the prediction results of multiple diverse base classifiers. Experimental results show that under the optimal configuration, the model achieves an F1-score of 83.57% and an AUC of 95.05%, outperforming the optimal single-text model, the optimal single-image model, and the original CLIP multimodal model by 2.48, 5.27, and 1.87 percentage points in F1-score, respectively. These results verify the effectiveness of the proposed method in mining modal correlations and improving recognition performance, providing technical support for intelligent financial public opinion analysis and financial risk monitoring.

  • Boxiang Ma, Ru Li, Shaoru Guo, Yinghao Li, Gutiérrez-Basulto Víctor
    Journal of Computer Research and Development. 2026, 63(9): 2187-2201.

    Large language models (LLMs) exhibit systematic generalization limitations in scenario cognition tasks, revealing potential functional deficiencies in their knowledge memorization processes. Inspired by the pattern separation mechanism of the dentate gyrus in the human hippocampus, we hypothesize that failures in scenario cognition are associated with limitations in the internal pattern separation capability of LLMs. Based on this hypothesis, we first design a cross-entity representation diagnosis (CERD) method, which quantitatively measures entity representation confusion across different layers without additional training. Our analysis reveals that standard autoregressive training fails to induce sufficiently effective pattern separation capabilities, leaving substantial cross-entity correlation confusion within LLM representations. Building upon this observation, we further propose latent-space pattern separation (LPS), which explicitly incorporates pattern separation constraints into the training process by introducing a triplet-based contrastive regularization at target layers with severe entity confusion, thereby reducing correlations among easily confused knowledge representations. Extensive experiments across LLMs ranging from 0.8B to 9B parameters demonstrate that LPS consistently improves scenario cognition capabilities. Compared with standard autoregressive training, our method achieves up to 101.5% relative improvement in exact match (EM), providing a new perspective for understanding and enhancing scenario cognition and knowledge memorization mechanisms in LLMs.

  • Wei Chen, Fuzhen Zhuang
    Journal of Computer Research and Development. 2026, 63(9): 2202-2214.

    In recent years, multimodal knowledge graphs have introduced auxiliary information such as entity textual descriptions and images, providing richer semantic sources for entity representation learning in knowledge graph completion. However, most existing multimodal knowledge graph completion methods focus on the alignment or fusion of features from different modalities, typically modeling such information as static external attributes of entities, while insufficiently exploiting the structural context formed by local topology. Since entity semantics are derived not only from their own multimodal attributes but also influenced by neighborhood relations, the lack of relation-constrained structural propagation may cause multimodal features to remain at the level of static semantics, making it difficult to form structured representations with reasoning capability. Particularly in scenarios where modal information contains strong noise, directly relying on initial multimodal features or prematurely fusing heterogeneous modalities may weaken the stability and discriminability of entity representations, thereby limiting the ability to predict complex relation patterns and potential facts. To address these issues, we propose a multimodal knowledge graph completion method based on a multi-view enhanced graph attention network, namely MEGAT. The proposed method models structural, visual, and textual information as three independent views under a shared graph topology, and constructs a relation-aware graph attention encoder for each view. In this way, relation-constrained neighborhood propagation and structural semantic modeling can be performed separately in different modal spaces. Subsequently, a gated late-fusion mechanism is employed to dynamically integrate multi-view representations, effectively alleviating the semantic interference caused by the direct fusion of heterogeneous modalities. Experimental results show that the proposed method enhances the ability of entity representations to capture graph structural context and complex relation patterns, achieving competitive completion performance on multiple public datasets.

  • Di Wang, Junping Du, Zhe Xue, Guanhua Ye, Tong Zhao, Xiaolong Meng, Haisheng Li
    Journal of Computer Research and Development. 2026, 63(9): 2215-2230.

    To meet the demand for dynamic knowledge updating and long-term knowledge retention in knowledge-enhanced artificial intelligence, dynamic multimodal knowledge graph continual completion infers missing facts incrementally as graph snapshots evolve while keeping historical knowledge stable. It thus provides fundamental support for dynamic knowledge updates. Existing methods predominantly focus on single-modal structural information, failing to fully exploit textual and visual semantics; directly incorporating multimodal information may introduce unexpected semantic shifts across snapshots, thereby exacerbating catastrophic forgetting. To this end, we propose SAREC, a semantic alignment and adaptive replay enhanced method for dynamic multimodal knowledge graph continual completion. SAREC leverages large language models to generate entity semantic descriptions and combines them with vision-language pretrained representations to construct a multimodal discrepancy-aware fusion mechanism, enhancing the completeness and robustness of entity representations. A multimodal temporal semantic alignment module is further introduced to constrain cross-snapshot representation evolution and maintain semantic consistency. Finally, a temporal-aware adaptive replay mechanism is proposed, which dynamically allocates the replay budget according to snapshot recency and selectively replays historical knowledge to alleviate catastrophic forgetting. Experiments demonstrate that SAREC achieves superior performance under various knowledge evolution modes, providing effective support for continual knowledge injection and adaptive knowledge enhancement in open dynamic environments.

  • Jiejing Shao, Ming Yang, Xiaowen Yang, Yufeng Li
    Journal of Computer Research and Development. 2026, 63(9): 2231-2247.

    Label scarcity remains a major obstacle to deploying machine learning in real applications, because many tasks require costly annotation by domain experts. Active learning reduces this cost by selecting valuable unlabeled instances for expert annotation. Existing methods usually design query criteria from a learner-centric view, such as uncertainty, diversity, committee disagreement, and core-set coverage. However, their gains are not stable across datasets, models, and annotation budgets, and an unsuitable strategy may even underperform random annotation. Meanwhile, conventional active learning often restricts experts to the role of label providers, lacking a collaborative medium that explains why samples are selected and what decision logic the model follows. As a result, expert feedback is difficult to elevate from individual labels into reusable knowledge rules, and the model cannot be effectively aligned with expert domain knowledge. To address these issues, we propose KALE, a knowledge-guided active learning method for explainable human-AI collaboration. KALE uses first-order logic rules as the collaborative medium between the model and experts. It first induces compact logic rules from the current model behavior, and then asks experts to judge their correctness according to domain knowledge. Correct rules are used for deductive weak-label expansion, while incorrect rules are treated as symbolic observations of knowledge-model disagreement and used to drive abductive disagreement querying. KALE further combines class-level rules, local rules, and sparse rule selection to keep expert interaction manageable. Experiments show that KALE achieves more robust improvement than existing active learning baselines, especially under small annotation budgets.

  • Yuehang Si, Zefan Zeng, Boyan Wang, Jing Huo, Yang Gao, Jincai Huang
    Journal of Computer Research and Development. 2026, 63(9): 2248-2263.

    Temporal knowledge graph (TKG) reasoning aims to predict future events based on historical facts. However, existing reinforcement learning (RL)-based reasoning methods merely conduct path searches at the entity level, lacking high-level semantic and logical guidance for the reasoning process. Moreover, ontologies, as conceptual abstractions of graphs, have strong logicality, but their construction typically relies on static statistical rules and is difficult to adapt dynamically to varying data distributions. To effectively address these issues, we propose ODRT, an ontology-graph dual-agent reinforcement learning method for TKG reasoning. Specifically, it jointly initializes the TKG ontology by leveraging both large language model (LLM) semantic embeddings and pretrained TKG reasoning model structural embeddings. Furthermore, it dynamically optimizes the ontology structure through learnable frequency weights during the end-to-end training process. Within the proposed dual-agent RL framework, the upper-level ontology agent performs concept-level walks on the constructed ontology graph, providing effective semantic navigation for the lower-level graph agent through carefully designed shared historical states. Meanwhile, the lower-level graph agent conducts fine-grained entity-level searches on the temporal graph under the constraint of hierarchical soft reward functions. Extensive experiments on four publicly available reasoning datasets demonstrate that ODRT achieves state-of-the-art performance among existing RL-based baselines and reaches the current state-of-the-art level on two encyclopedic datasets. However, on event-based datasets with heterogeneous entity behavior patterns and predictions that rely on historical reuse, its performance still lags behind that of partially embedded methods. Comprehensive experimental analyses further verify the significant advantages of the proposed ontology-guided mechanism and the dual-agent reinforcement learning framework.

  • Jie Ma, Hanchi Zhang, Ning Qu, Haoquan Xue, Xinping Wang, Jun Liu
    Journal of Computer Research and Development. 2026, 63(9): 2264-2284.

    Multimodal retrieval-augmented generation introduces external textual and visual knowledge, providing important support for complex document understanding, knowledge-based question answering, and step-by-step task guidance. However, existing methods usually delegate the interleaved text-image generation process to a single multimodal large language model. Specifically, the retrieved text chunks, candidate images, and image descriptions are simultaneously fed into the same model, which is required to perform text organization, visual evidence selection, image insertion position determination, and multi-image ordering within a single generation process. In complex long-context scenarios, however, the candidate content often contains weakly relevant text, visually similar images, and redundant visual evidence, which can easily lead to missing key images, selecting irrelevant images, and misordering multiple images. To address these problems, we propose a multimodal retrieval-augmented interleaved text-image generation method based on a multi-agent debate mechanism. The proposed method decomposes the generation process into three collaborative roles: a text reasoning agent, a visual perception agent, and a global judge agent. Through dual-track draft generation, set conflict and ordering conflict detection, multi-round adversarial debate, and global judgment and fusion, the method jointly optimizes image selection, dynamic image supplementation, redundant image removal, and image reordering. Experimental results show that the proposed method achieves competitive performance across the network data, academic document, and lifestyle domains of MRAMG-Bench. In particular, in the lifestyle domain, compared with the strong baseline M2IO-R1, the proposed method improves image recall and ordering consistency by 7.02 and 7.77 percentage points, respectively. In the network data domain, the proposed method improves the average performance by 7.90 percentage points compared with the best complete baseline. These results demonstrate the effectiveness of the proposed method in complex long-document interleaved text-image generation tasks. Anonymous code and reproduction instructions are available at: https://github.com/mira-ai-lab/MRAMG-MultiAgent-Debate