Enming Wang, Jianlei Wang, Xueping Peng, Hongjiao Guan, Yinglong Wang, Sibo Wei, Jianbin Guo, Ruifeng Xu, Wenpeng Lu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

The International Classification of Diseases (ICD) coding is a crucial component for standardizing medical information, and automated coding represents an important direction for improving the efficiency and accuracy of this process. However, current automated ICD coding research still faces significant challenges. First, existing studies are predominantly based on English contexts, making their models difficult to directly apply to China's localized ICD versions, while the immense pressure on its healthcare system makes this research particularly urgent in China. In addition, mainstream approaches oversimplify coding as multi-label classification, thus ignoring the coder workflow and failing to generalize to unseen codes or uncover implicit diagnoses. To address these challenges, we propose AiCoder, a novel multi-agent framework that simulates real-world coding workflows through a two-stage design. The first stage focuses on accuracy by parsing discharge diagnoses, correlating medical record evidence, and utilizing knowledge graphs for standardization; the second stage emphasizes completeness by leveraging ICD hierarchical structures and integrating clinical reasoning patterns to identify potentially missed codes. Furthermore, we construct a publicly available high-quality Chinese ICD annotation dataset, providing a valuable resource for the research community. Comprehensive experiments including ablation study and case study validate our framework's effectiveness. We also systematically analyze common coding errors of LLMs in Chinese ICD coding tasks, hoping to provide reference for future research.

The International Classification of Diseases (ICD) coding assigns standardized codes to diseases. Automating this process enhances the efficiency and accuracy of clinical records processing. However, current methods struggle with noisy and lengthy clinical texts, making it difficult to ensure the reliability of feature extraction. Furthermore, they typically make separate binary predictions for each code, overlooking the dependencies between them. To address these issues, we propose a novel model called Hierarchical Feature Optimization and Rescorer (HiRes). We employ a cascaded convolution architecture to mitigate noise and enhance feature representation. Additionally, a masked autoencoder-based rescorer is introduced to capture ICD code interdependencies, refining initial predictions for improved accuracy. The model also considers the inconsistencies in code representations. Experiments on the MIMIC datasets demonstrate the effectiveness of our model.

Multimodal Sentiment Analysis (MSA) aims to leverage heterogeneous data—typically language, vision, and acoustic modalities—to accurately interpret human emotional states. Despite recent advances, challenges persist due to the feature distribution difference caused by intrinsic modality heterogeneity. Prior works either neglect the contribution disparity among modalities, especially the dominant role of language in sentiment reasoning, or emphasize language dominance in fusion-space alignment, ignoring coordination in the early feature space. To address these limitations, we propose a novel Two-Stage Primary Modality-Guided (TSPMG) framework, which introduces primary-modality supervision into both feature-space distribution alignment and fusion-space attention modulation. This dual-level cooperative mechanism progressively amplifies the dominant modality’s influence throughout the entire representation learning pipeline. Extensive experiments on two benchmark datasets demonstrate that TSPMG achieves superior or comparable results to state-of-the-art baselines, with ablation studies further validating the effectiveness of primary-modality-guided strategies for robust and interpretable multimodal sentiment analysis. The code is available at https://github.com/Kaisa777/TSPMG.

Medical entity disambiguation (MED) plays a crucial role in natural language processing and biomedical domains, which is the task of mapping ambiguous medical mentions to structured candidate medical entities from knowledge bases (KBs). However, existing methods for MED often fail to fully utilize the knowledge within medical KBs and overlook essential interactions between medical mentions and candidate entities, resulting in knowledge- and interaction-inefficient modeling and suboptimal disambiguation performance. To address these limitations, this paper proposes a novel approach, MED with Medical Mention Relation and Fine-grained Entity Knowledge (MMR-FEK). Specifically, MMR-FEK incorporates a mention relation fusion module and an entity knowledge fusion module, followed by an interaction module. The former employs a relation graph convolutional network to fuse mention relation information between medical mentions to enhance mention representations, while the latter leverages an attention mechanism to fuse synonym and type information of candidate entities to enhance entity representations. Afterwards, an interaction module is designed to employ a bidirectional attention mechanism to capture interactions between mentions and entities to generate the matching representation. Extensive experiments on two publicly available real-world datasets demonstrate MMR-FEK’s superiority over state-of-the-art(SOTA) MED baselines across all metrics. Our source code is publicly available.

Most Seq2Seq neural model-based medical question summarization (MQS) systems have a severe mismatch between training and inference, i.e., exposure bias. However, this problem remains unexplored in the MQS task. To bridge this research gap and alleviate the problem of exposure bias, we propose a novel re-ranking training framework for MQS called Multi-view Contrastive Learning (MvCL). MvCL simultaneously considers the similarity scores between medical questions and candidate summaries as well as the average similarity scores between candidate summaries and other candidates within the same group, and utilizes contrastive learning to optimize the model’s ranking ability. Additionally, we propose a new multilevel inference approach to adapt to this training strategy. The approach first filters out candidate summaries that are dissimilar to the original medical question, and then selects the summary with the highest average similarity to other candidate summaries from the remaining candidates as the final output. We conducted extensive experiments, and the results demonstrate that our proposed MvCL framework achieves state-of-the-art results on the majority of evaluation metrics across four datasets.1

Sequential recommendation has been developed to predict the next item in which users are most interested by capturing user behavior patterns embedded in their historical interaction sequences. However, most existing methods appear to exhibit limitations in modeling fine-grained dependencies embedded in users’ various periodic behavior patterns and heterogeneous dependencies across multi-behaviors. Towards this end, we propose a Filter-enhanced Hypergraph Transformer framework for Multi-Behavior Sequential Recommendation (FHT-MB) to address the above challenges. Specifically, a multi-scale filter layer equipped with multi-learnable filters is devised to encode behavior-aware sequential patterns emerging from different periodic trends (e.g., daily or weekly routines), and then a hypergraph structure is devised to extract heterogeneous dependencies across users’ multiple types of behaviors. Extensive experiments on two real-world e-commerce datasets show the superiority of our proposed FHT-MB over various state-of-the-art methods.1

Sequential diagnosis prediction (SDP) is a challenging task, aiming to predict patients’ future diagnoses based on their historical medical records. While methods based on graph neural networks (GNNs) have proven successful for this task, they typically focus on modeling pairwise diseases using a global disease combination graph. However, these approaches neglect the fine-grained higher-order relations among persistent and emerging diseases within a single visit, which may contain crucial clues to predict the next diagnosis. Additionally, they fail to fully leverage patient-related clinical information present in electronic health records (EHRs). To address these challenges, this paper proposes a novel approach called the fusion of Dynamic Hypergraph and Clinical Event (DHCE) for sequential diagnosis prediction. The proposed method aims to exploit the fine-grained higher-order relations among diagnoses within a visit and leverage clinical event information from EHRs to improve the accuracy of predicting the next diagnosis. Specifically, DHCE categorizes diagnoses within a single visit in a fine-grained granularity into persistent and emerging categories based on a patient’s historical diagnoses. It then constructs dynamic hypergraphs to capture higher-order disease relations within each visit. Next, we design a transition function to extract the transitional context from previous visits in order to generate the visit representation. Furthermore, to fully leverage patient-related clinical events in a visit, we utilize Bio-Clinical BERT to encode them and generate the clinical event representation for each visit. Finally, we combine the visit representation and event representation to generate a comprehensive patient representation, which is then used to predict the patient’s next diagnosis. Experimental results on two benchmark datasets consistently demonstrate that DHCE outperforms state-of-the-art methods1.

Imbalanced classification has always been a challenging issue. The minority class usually has degraded recognition rate. The key factors are sample scarcity of the minority class and intrinsic complex distribution characteristics in imbalanced data. SMOTE and its extensions are commonly used data-level methods. These SMOTE-related methods oversample all or unsafe minority class examples, without considering sample difference or only considering sample difference in spatial distribution. In this paper, we propose a dual self-paced SMOTE (DSP-SMOTE) method, which considers temporal-spatial distribution of samples. DSP-SMOTE adopts a two-way self-paced mechanism at different stages of sampling. The majority class is undersampled according to easy example first principle, whereas the minority class is oversampled following difficult sample first principle. In DSP-SMOTE, a hybrid (including static and dynamic) difficulty is introduced to estimate easy or difficult examples. Furthermore, the trade-off between static and dynamic difficulties is adjusted adaptively according to sample distribution and training status. Experimental results demonstrate that DSP-SMOTE outperforms previous SMOTE-related methods significantly in terms of AUC and sensitivity.

Many practical applications suffer from the problem of imbalanced classification. The minority class has poor classification performance; on the other hand, its misclassification cost is high. One reason for classification difficulty is the intrinsic complicated distribution characteristics (CDCs) in imbalanced data itself. Classical oversampling method SMOTE generates synthetic minority class examples between neighbors, which is parameter dependent. Furthermore, due to blindness of neighbor selection, SMOTE suffers from overgeneralization in the minority class. To solve such problems, we propose an oversampling method, called extended natural neighbors based SMOTE (ExNN-SMOTE). In ExNN-SMOTE, neighbors are determined adaptively by capturing data distribution characteristics. Extensive experiments over synthetic and real datasets demonstrate the effectiveness of ExNN-SMOTE dealing with CDCs and the superiority of ExNN-SMOTE over other SMOTE-related methods.

Imbalanced classification exists commonly in practical applications, and it has always been a challenging issue. Traditional classification methods have poor performance on imbalanced data, especially, on the minority class. However, the minority class is usually of our interest, and its misclassification cost is higher. The critical factor is the intrinsic complicated distribution characteristics in imbalanced data itself. Resampling ensemble learning achieves promising results and is a research focus recently. However, some resampling ensembles do not consider complicated distribution characteristics, thus limiting the performance improvement. In this paper, a generalized optimization embedded framework (GOEF) is proposed based on undersampling bagging. The GOEF aims to pay more attention to the learning of local regions to handle the complicated distribution characteristics. Specifically, the GOEF utilizes out-of-bag data to explore heterogeneous local areas and chooses misclassified examples to optimize base classifiers. The optimization can focus on a single class or both classes. Extensive experiments over synthetic and real datasets demonstrate that GOEF with the minority class optimization performs the best in terms of AUC, G-mean, and sensitivity, compared with five resampling ensemble methods.

This paper proposes individualized cleaning for diverse imbalanced data sets. Existing techniques for data cleaning have difficulties with rare cases and outliers in minority class, especially, in highly unbalanced data. The drawback leads incomplete and imprecise examples to removal. In order to enhance the robustness and perform thorough data cleaning, we propose a weighted edited nearest neighbor (WENN), which detects and removes noisy examples from both classes intelligently. It considers individual characteristics of each imbalanced data, involving global class imbalance and local distribution. The main idea of the proposed method is to carefully put more focus on the majority class than the minority class during data cleaning. Extensive experiments over synthetic and real data clearly validate the superiority of our approach against other data cleaning methods.