Yeqing Teng, Jiasheng Si, Shuxia Lin, Linhai Zhang, Weiyu Zhang, Wenpeng Lu, Deyu Zhou, Xiaoming Wu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Toxicity detection mitigates the dissemination of toxic content (e.g., hateful comments, posts, and messages within online social actions) to safeguard a healthy online social environment. However, malicious users persistently develop evasive perturbations to disguise toxic content and evade detectors. Traditional detectors or methods are static over time and are inadequate in addressing these evolving evasion tactics. Thus, continual learning emerges as a logical approach to dynamically update detection ability against evolving perturbations. Nevertheless, disparities across perturbations hinder the detector's continual learning on perturbed text. More importantly, perturbation-induced noises distort semantics to degrade comprehension and also impair critical feature learning to render detection sensitive to perturbations. These amplify the challenge of continual learning against evolving perturbations.

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.

Argumentative essay generation (AEG) is a complex task that requires advanced semantic understanding, logical reasoning, and organized integration of perspectives.Despite showing a promising performance, current efforts often overlook the dynamical and hierarchical nature of structural argumentative planning, and struggle with flexible rhetorical expression, leading to limited argument divergence and rhetorical optimization.Inspired by human debate behavior and Bitzer's rhetorical situation theory, we propose a debate-driven rhetorical framework for argumentative writing.The uniqueness lies in three aspects: (1) it dynamically assesses the divergence of viewpoints and progressively reveals the hierarchical outline of arguments based on a depththen-breadth paradigm, improving the perspective divergence within argumentation; (2) simulates human debate through iterative defenderattacker interactions, improving the logical coherence of arguments; (3) incorporates Bitzer's rhetorical situation theory to flexibly select appropriate rhetorical techniques, enabling the rhetorical expression.Experiments on four benchmarks validate that our approach significantly improves logical depth, argumentative diversity, and rhetorical persuasiveness over existing state-of-the-art models 1 .* Corresponding authors. 1 Code and data are available at https://github.com/ zxg-x/DARE Social media affects attention and distracts people.Social media affects attention and distracts people.Social media affects attention and distracts people.Social media is a waste of time.Social media is a waste of time.Social media is a waste of time.

Medication recommendation aims to suggest personalized drug combinations to patients based on their longitudinal medical histories stored in electronic health record (EHR) datasets. Patients' Dynamic Treatment Regimes (DTRs) determine how patients' drug combinations change along with the evolution of disease treatment. DTRs are effective for comprehending disease-treatment dynamics and for recommending a timely and personalized combination of medications for patients. However, existing medication recommender systems (MRSs) overlook the multiple treatment pathways generated by the intervention of DTRs and can only recommend a single treatment paradigm, ignoring the fact that patients may be at different treatment stages and thus require different treatment regime. Such disregard leads to a significant limitation in recommending personalized medication combinations tailored to different treatment stages, yielding greatly compromised accuracy and applicability of MRSs. Moreover, existing methods often overlook the time interval information over patients' successive visits, which is critical to indicate patients' treatment evolution. To address these significant gaps, we propose a Time-aware Medication Recommendation Framework via Intervention of Dynamic Treatment Regimes, called MR-DTR. To explicitly illustrate the intervention processes of DTRs on similar patients, we employ a co-guided graph to connect various patient sequences. In addition, to fully utilize the time interval information, we design a time-aware guidance mechanism dedicated to the co-guided graph to efficiently learn medication representation using the patient's guidance information. We also introduce relative time intervals in the encoder to act as positional information. Extensive experiments on two real-world datasets demonstrate that MR-DTR surpasses state-of-the-art models in terms of recommendation performance. Our code is available at: https://github.com/liyifo/MR-DTR.

Multi-modal hashing achieves low storage costs and high retrieval speeds by using compact hash codes to represent complex and heterogeneous multi-modal data, effectively addressing the inefficiency and resource intensiveness challenges faced by the traditional multi-modal retrieval methods. However, balancing intraclass compactness and interclass separability remains a struggle in existing works due to coarse-grained feature limitations, simplified fusion strategies that overlook semantic complementarity, and neglect of the structural information within the multi-modal data. To address these limitations comprehensively, we propose a Proto-centric Multi-modal Hashing with Pronounced Category Differences (PMH-PCD) model. Specifically, PMH-PCD first learns modality-specific prototypes by deeply exploring within-modality class information, ensuring effective fusion of each modality's unique characteristics. Furthermore, it learns multi-modal integrated class prototypes that seamlessly incorporate semantic information across modalities to effectively capture and represent the intricate relationships and complementary semantic content embedded within the multi-modal data. Additionally, to generate more discriminative and representative binary hash codes, PMH-PCD integrates multifaceted semantic information, encompassing both low-level pairwise relations and high-level structural patterns, holistically capturing intricate data details and leveraging underlying structures. The experimental results demonstrate that, compared with existing advanced methods, PMH-PCD achieves superior and consistent performances in multi-modal retrieval tasks. To promote further research and reproducibility, we have publicly released the source code of PMH-PCD at https://github.com/vindahi/PMH-PCD.

Drug-target binding affinity prediction is critical to drug design. Some existing computational methods rely on single-scale data and cannot fully integrate the rich information of multi-scale data, such as molecular structures and network information. In this paper, we propose the MBC-DTA model, which effectively integrates multi-scale data, and models the complex interactions between atoms and amino acids. We introduce GCN and GAT, which effectively extract the feature representation of molecular structure scale and network scale. Afterwards, the features obtained from the molecular structure scale are fed into a bilinear attention network, which captures the intricate interaction information between drug-target pairs. Additionally, a cross-scale graph contrastive learning strategy is applied, optimizing the feature representations across different scales. Extensive experimental results on two benchmark datasets demonstrate that the MBC-DTA model outperforms existing state-of-the-art methods.

Electronic health records (EHRs) play a crucial role in the development of personalized treatment plans for patients. However, EHRs are often highly incomplete, posing significant challenges for predictive modeling. While existing deep learning models employ various imputation techniques to reconstruct missing values, they often fail to represent missing data in a personalized manner and do not learn from the missingness patterns in EHRs data. This limitation reduces their effectiveness in practical personalized healthcare applications. To address this issue, we propose SPIME, a self-supervised model that generates personalized imputations in EHRs data for patient health status prediction. We introduce a personalized missing mask (PMM) based on the frequency of feature measurements. Additionally, we incorporate a masked imputation task (MIT) loss that minimizes the loss of artificially introduced missing values, thereby enhancing the model’s capability to handle missing data. SPIME adopts self-supervised pretraining to learn representations from personalized missing patterns and reconstructs missing data in the latent space. To further enhance representation learning, two independent attention mechanisms are applied separately across the feature and temporal dimensions. Experimental results on two real-world EHRs datasets show that SPIME outperforms existing state-of-the-art methods in predicting in-hospital mortality and decompensation, demonstrating its effectiveness in reconstructing missing data and predicting patient health status. The code will be published at https://github.com/cling6666/SPIME.

Drug-target interaction (DTI) prediction is a critical step in drug discovery. Deep learning has shown great potential in DTI prediction. In recent years, many sequence-based DTI prediction methods have been proposed. However, the current methods do not fully represent drugs and proteins and ignore the local and global features of these molecules. In response to these challenges, we propose a sequence-based CNN and Transformer collaborative prediction method (SCTDTI) for drug-target interaction. We train and evaluate our proposed approach on two public drug-target datasets, and experimental results show that SCTDTI improves DTI prediction performance compared to state-of-the-art baselines.

Sequential diagnosis prediction (SDP) is a complex and challenging task, aming to predict future diagnoses of patients by analyzing their historical medical records. Although graph neural networks(GNNs) has been applied to successfully address the challenge of heterogeneous data integration in electronic health records, relatively limited work has been done on GNNs for sequential diagnosis prediction. Graph neural network-based methods, aimed at capturing structural and relational patterns of EHR data for sequential diagnosis prediction, explore code-code pairwise relationships, resulting in an inability to learn fine-grained, higher-order interaction relationships among different types medical codes. As a result, they are difficult to effectively model complex, multi-dimensional interactions among different types of medical codes necessary for accurate and nuanced diagnosis predictions. To address these challenges, this paper proposes a novel approach called Multi-Channel Hypergraph Network (MCHN) predictive framework for sequential diagnosis prediction. The proposed method aims to explore the fine-grained higher-order interactions between different types of medical codes via multi-channel hypergraphs. Specifically, MCHN learns two levels of code embeddings from multi-channel hypergraph learning module and line graph learning module, respectively: (i) multi-channel hypergraph learning module, which is to learn multi-channel hypergraph level code embeddings by modeling the higher-order relationships between medical codes in different hypergraphs; and (ii) line graph learning module, which is to learn the line graph level code embeddings by modeling code-code pairwise relationships. In MCHN, we propose a novel channel-level attention mechanism to help our model attend to the informativeness of the different channel for forecasting future patient diagnoses. We also design a code-level attention mechanism, which can to pay more attention to the medical codes that are more important to the visit representation. Moreover, MCHN aggregates the learnt code embeddings in the two levels to generate the visit representation, which is used to predict the patient’s next diagnosis. Experimental results on two benchmark datasets consistently demonstrate that MCHN outperforms state-of-the-art methods1.

Traditional Chinese Medicine (TCM) is a natural, safe, and effective therapeutic approach with widespread application worldwide. The unique diagnostic methods of TCM often require a comprehensive analysis of patient information, much of which is contained in clinical text. Numerous studies have demonstrated the effectiveness of natural language processing (NLP) techniques in TCM disease classification. Therefore, this paper focuses on the task of TCM syndrome differentiation, proposing a novel matching score calculation method and a new label attention calculation method to assist the model in focusing on the relationship between TCM syndrome and disease symptom. Specifically, we enhance the model’s attention to the uniqueness of the relationship between TCM syndrome and symptom by introducing a finer-grained token-level matching score. Simultaneously, we improve the model’s attention to the generality of the relationship between TCM syndrome and symptom through a more global label attention mechanism. Additionally, we observe a severe long-tail problem in the dataset. To alleviate this issue, we propose the use of focal loss to help the model pay more attention to challenging samples. Extensive experiments on the TCM-SD dataset indicate that our approach significantly outperforms state-of-the-art baselines1.

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

Query embedding (QE) is designed to embed entities and first-order logic (FOL) queries in low-dimensional spaces, it has demonstrated considerable effectiveness in multi-hop reasoning over knowledge graphs. Recently, the practice of embedding entities and queries using geometric shapes has emerged as a promising approach, given the natural ability of these shapes to represent query answer sets and the logical relationships between them. However, existing geometry-based models, such as those employing point and box embeddings, cannot handle logical negation operations. Additionally, these models typically focus only on entities and relations, overlooking the rich semantic information available in knowledge graphs. To address these limitations, we propose the Type-aware Enhanced Cylinder Embedding (TECE) model. This model leverages three-dimensional geometric representations, integrating entity relationship type information. TECE covers a spectrum of essential First-Order Logic (FOL) operations, such as conjunctive, disjunctive, and negative. TECE improves the representation of entities and relationships in queries by incorporating semantic information. Moreover, it strengthens the model’s ability in generalization, along with deductive and inductive reasoning. Our experiment results demonstrate that TECE achieves significantly superior performance compared with existing SOTA methods on standard datasets.

Community detection is an important task in graph analysis, and it is of great significance in reality. Recently, unsupervised learning has been widely used in community detection tasks. However, only a few community detection models combine unsupervised learning with graph neural networks (GNNs). To this end, in this paper, we combine GNNs with unsupervised learning to propose a new model, Unsupervised graph neural network with Self-expressive attention for Community detection (USCom). We first use the graph attention encoder to generate node embeddings. Then we apply the self-expressive principle to optimize the node embeddings to make them more suitable for community detection tasks. Finally, we utilize a four-layer perceptron for community detection. The experimental results show that the model proposed in this paper outperforms the comparison baselines on community detection tasks.

Drug-drug interaction (DDI) has been a challenging problem in healthcare machine-learning research. DDI might cause changes in drug pharmacological activity and trigger adverse patient reactions. Therefore, it is critical for both patients and society to effectively identify potential DDI. Most recent studies have used graph-based learning methods to predict DDI, but these methods usually have the following limitations: i) modeling drug information on a single view; ii) ignoring the importance of different neighboring nodes; and iii) failing to integrate the drug-embedded information well. Therefore, this paper proposes a multi-view feature fusion strategy based on graph attention networks(MV-GAT). In MV-GAT, we use the graph representation learning method of bond-aware message passing neural network to obtain the local features of each atom in the molecular graph and the global features of the molecular graph. Meanwhile, We propose an attention mechanism based on a fusion strategy to handle the fusion of drug features and topological information under each view, which can efficiently integrate the features extracted from molecular and interaction maps. In addition, to ensure the diversity of node features, we use an unsupervised contrastive learning component in each Graph Neural Networks (GNN) layer to address the over-smoothing problem during information transfer. Comprehensive experiments on multiple real datasets show that MV-GAT has good generalization performance.

Recently, recommendation methods based on graph convolutional networks (GCNs) have received much attention and have shown competitive performance. These methods usually perform graph convolutions of user and item embeddings in Euclidean space. However, user-item interaction graphs usually present a tree-like hierarchy in large-scale recommender systems. To better fit user and item data to the embedding space, hyperbolic space becomes a good choice, which grows exponentially in volume with radius and is well suited for embedding hierarchical data. In addition, most previous works aggregate neighborhoods through projection operations on tangent spaces, which inevitably cause distortion. In this paper, we propose a Lorentzian graph convolutional network model for collaborative filtering (LCF), which makes the learned user and item representations strictly follow hyperbolic geometry, and the learned embeddings are optimized based on margin ranking loss and geometric views. Specifically, we first pass the initialized user and item information through multiple Lorentzian graph convolutional layers, and ensure that the learned node embeddings are not out of the hyperbolic space, while capturing rich high-order neighborhood information. Then, the model is effectively learned under the composite loss function to take full advantage of the hyperbolic space. We conduct extensive experiments on three datasets and compare with many baselines. Experimental results show that our method performs better than many state-of-the-art baseline methods.

COVID-19 has become a worldwide epidemic. Prediction of COVID-19 is an effective way to control its spread. Recently, some research efforts have made great progress on this task. However, these works rarely combine both the temporal and spatial domains for case number prediction. Moreover, most of them are only suitable for short-term prediction tasks, which cannot achieve good long-term predicting effects. Therefore, we use a method that combines human-mobility factors and time-series factors - the Spatio-temporal convolutional network (G-TCN) to deal with these problems. Firstly, we use data on the mobility of people between regions to generate graphs of regional relationships. Secondly, to process the spatial information at each moment, we apply multi-layer graph convolutional neural networks (GCNs) to aggregate multi-layer neighborhood information. And we input the information obtained by GCNs at different moments into temporal convolutional networks (TCNs), which are used to process the time-series information. Finally, we tested the proposed G-TCN method using datasets from four countries. The experimental results show that G-TCN has lower prediction errors than other comparison methods and can better fit the trend of COVID-19 development.

Chinese sentence matching is a critical and yet challenging task in natural language processing. Recent work on modeling sentence semantic relations with deep neural models has shown its great potential in improving the performance of sentence matching. However, existing sentence matching methods usually focus on generating word-level sentence representation, which neglects the character-level information and leads to weak semantic representations. Also, they usually capture the interactive features with an attention-based alignment, which are typically implemented on sentence level and neglect the interactions among characters, words and sentences. This paper proposes a novel Chinese sentence matching model with Multiple Alignments and Feature Augmentation (MAFA). Specifically, the model first employs the multi-level embedding layer to accept the character and word sequences of sentences, and introduces the multiple alignment layer to capture the interactions among characters, words and sentences in turn. Then, the feature augmentation layer is applied to combine the interactive features to generate the final semantic matching representations. Finally, the prediction layer is utilized to judge the matching degree of the input sentences. Substantial and extensive experiments are conducted on two real-world data sets to show that MAFA significantly outperforms the competing methods and achieve comnarable nerformance with BERT-based methods.