Overlapping event extraction aims to identify triggers, arguments, and semantic roles when multiple events share textual components within the same sentence. In agricultural text, this task is particularly challenging because shared arguments may play different roles across events, while compound terms and long-span expressions often exhibit ambiguous boundaries. Existing methods still suffer from insufficient event conditioning and weak exploitation of boundary cues, which easily lead to role confusion and noisy span candidates during decoding. To address these issues, we propose ESBR, a structured model that integrates event-conditioned encoding, structured span inference, hierarchical boundary reasoning, and boundary-guided decoding. We further enhance training and inference stability with uncertainty-aware weighting and adaptive thresholding. Experiments on the public benchmark FewFC and our constructed agricultural benchmark FewAgri show that ESBR consistently outperforms competitive baselines, with especially notable gains on argument identification and role classification.

Extracting entities and relations from unstructured text is an important task in information extraction. In recent years, great success has been achieved in helping models understand text by formalizing the relation extraction task as a machine reading comprehension problem. However, existing approaches often lack semantic understanding of entity and relation categories, leading to poor differentiation of ambiguous types and weak semantic linkage between queries and context. To address these problems, our research infuses attribute knowledge into query representations and enhances model robustness through adversarial data augmentation. Furthermore, we introduce an interactive attention mechanism to deepen the semantic analysis between queries and context. Our experiments on the ACE2004 and ACE2005 datasets demonstrate significant improvements; our model outperforms the baseline F1 scores by 11.1% and 3.5%, respectively, achieving F1 scores of 60.5% and 63.7%.

This paper proposes a Dual-Channel Heterogeneous Agent System (DCHAS) for optimizing agricultural irrigation decisions. The irrigation decision problem is formulated as a fully cooperative Decentralized Partially Observable Markov Decision Process (DEC-POMDP), and two heterogeneous agent channels are designed: one channel focuses on historical environmental information to determine irrigation timing, while the other channel uses real-time crop growth status to determine irrigation volume. By sharing key data, the two agent channels work collaboratively to ensure the consistency and coordination of irrigation strategies. To further enhance system performance, a globally shared reward function is designed to incorporate the multi-objective optimization of crop yield and water consumption within a reinforcement learning framework, thereby promoting cooperative interaction between the agent channels. Experimental results show that, compared to baseline models, DCHAS demonstrates higher flexibility and adaptability under complex environmental and resource-constrained conditions, while also reducing training time and accelerating convergence.

Many scholars have employed dictionary-based word enhancement methods and multimodal information supplementation techniques to improve Chinese Named Entity Recognition models. However, these approaches primarily rely on static weights between different modalities, leading to a failure in capturing fine-grained correlations within the text modality and across modalities. As a result, they do not fully exploit multimodal information, leading to a loss of valuable data. To overcome these limitations, this paper proposes the Cross- and Self-Modality Attention network. This network dynamically captures correlations within the text modality and across modalities at multiple levels, effectively enhancing multimodal mutual information and reducing information loss. Additionally, we introduce two CNN structures to extract glyph visual and phonetic information. We conducted extensive experiments on Weibo, Resume, Ontonotes 4.0, and MSRA, and the results demonstrate that our approach outperforms state-of-the-art (SOTA) baseline methods.

Distantly supervised relation extraction (DS-RE) leverages existing knowledge bases to generate annotated data for relation extraction (RE), addressing the issue of scarce labeled data. However, distant supervision (DS) is often limited by coarse annotations and insufficient contextual awareness, leading to relational ambiguity and introducing noise in the labeled results. Moreover, although one can optimize the classifiers in DS-RE models through weight updates, the static nature of the guiding rules for such adjustments often falls short when addressing the challenges posed by diverse non-taxonomic relations and complex noise patterns in datasets. In this paper, we propose a DS-RE framework that capitalizes on non-taxonomic relations and a self-optimizing mechanism. We define a set of consistent DS relation candidates and combine DS with a LLM to enhance the perception of entities’ contextual states during the DS process. Then, we design a Self-Optimizing Ontology-Enhanced Non-taxonomic Relation Extraction Model (SO-NRE). The model incorporates additional entity-relation knowledge to enhance the semantic depth of Non-taxonomic relation ontologies and uses an adaptive dynamic scheduling mechanism to refine the classification strategy through iterations informed by self-perception outcomes. The experimental results show that the improved DS annotation workflow has enhanced accuracy, and SO-NRE outperforms mainstream baselines in RE performance.

With the rapid development of smart farm models, multi-machine collaborative operations are an important trend. Nevertheless, in the face of sudden disasters, the traditional multi-machine collaborative operations have unclear operation area allocations, simplistic path planning, and scheduling algorithms easily falling into the local optimal solution, resulting in low efficiency and inability to meet the task needs of urgent tasks. A multiple agricultural machine cooperative scheduling method based on farmland segmentation is proposed to solve these problems. The method first segmented large-scale farmland based on multiple machine task requirements to obtain location information for the segmented sub-regions. Next, to minimize the turning time in the headland, complete coverage path planning inside the field is planned to obtain the best coverage angle and the shortest operation time. Lastly, an optimal simulated annealing algorithm is employed to obtain the task allocation of agricultural machinery to minimize the longest working time. The simulation results show that the study has achieved optimization results. Under urgent tasks, the method can effectively schedule agricultural machine resources, ensuring task completion and generating a high-yield solution.