Zhiqing Cui, Binwu Wang, Qingxiang Liu, Yeqiang Wang, Zhengyang Zhou, Yuxuan Liang, Yang Wang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Text-based person search (TBPS) enables the retrieval of person images from large-scale databases using natural language descriptions, offering critical value in surveillance applications.However, a major challenge lies in the labor-intensive process of obtaining highquality textual annotations, which limits scalability and practical deployment.To address this, we introduce two complementary modules: Multi-Turn Text Generation (MTG) and Multi-Turn Text Interaction (MTI).MTG generates rich pseudo-labels through simulated dialogues with MLLMs, producing finegrained and diverse visual descriptions without manual supervision.MTI refines user queries at inference time through dynamic, dialogue-based reasoning, enabling the system to interpret and resolve vague, incomplete, or ambiguous descriptions-characteristics often seen in real-world search scenarios.Together, MTG and MTI form a unified and annotation-free framework that significantly improves retrieval accuracy, robustness, and usability.Extensive evaluations demonstrate that our method achieves competitive or superior results while eliminating the need for manual captions, paving the way for scalable and practical deployment of TBPS systems.

Learning from Label Proportions (LLP) is a weakly-supervised paradigm that uses bag-level label proportions to train instance-level classifiers, offering a practical alternative to costly instance-level annotation. However, the weak supervision makes effective training challenging, and existing methods often rely on pseudo-labeling, which introduces noise. To address this, we propose RLPL, a two-stage framework. In the first stage, we use unsupervised contrastive learning to pretrain the encoder and train an auxiliary classifier with bag-level supervision. In the second stage, we introduce an LLP-OTD mechanism to refine pseudo labels and split them into high- and low-confidence sets. These sets are then used in LLPMix to train the final classifier. Extensive experiments and ablation studies on multiple benchmarks demonstrate that RLPL achieves comparable state-of-the-art performance and effectively mitigates pseudo-label noise.

Anti-CRISPR proteins play a crucial role in bacterial-phage interactions by inhibiting the CRISPR/Cas system and thus enhancing phage survival. Accurately predicting these proteins is essential for understanding phage-host immune interactions and progressing CRISPR/ Cas-based technologies. Current approaches primarily analyze proteins individually, which may overlook the intrinsic similarities and potential connections among protein sequences. This study introduces PACRGNN, a graph neural network framework that creates a heterogeneous protein network by integrating sequence and structural similarities, wherein nodes represent proteins and edges signify their relationships. By combining Graph Attention (GAT) and Graph Sample and Aggregation (GraphSAGE) layers, PACRGNN captures both local and global topological dependencies, while incorporating six protein feature categories to enrich node representations. PACRGNN achieves an accuracy of 0.9577, an F1-Score of 0.9572, and a PRAUC of 0.9876 on the validation set. The model demonstrated superior performance to existing methods on the independent test set derived from NCBI database (Jan.-Oct. 2024).

It is well-established that miRNAs play a crucial role in the occurrence and development of diseases. Current miRNA-disease associations prediction research faces several challenges, including model bias due to data sparsity, information loss from overlooked complex relationships during feature fusion and insufficient capability of existing methods to capture the intricate relationships (between miRNAs, genes, lncRNAs and diseases), thereby limiting prediction accuracy. Based on hybrid gated GNN and multi-data fusion, a method (PMDGGM) for predicting miRNA-disease associations is proposed in this study. PMDGGM constructed seven similarity networks by comprehensively considering the relationships between miRNA and related genes, miRNA, lncRNA and diseases. It provides a solid foundation for feature fusion and information propagation. Subsequently, the method captures the complex relationships between heterogeneous nodes through a bilinear pooling layer and uses a gating mechanism to fuse multi-source heterogeneous features, thereby predicting miRNA-disease associations more accurately. The experimental results show that the method performs well and has significant advantages in predicting the miRNA-disease associations. Among various evaluation metrics, especially the AUCs of ROC and PR curves, the performance of method is outstanding, reaching a high level of 0.9413 and 0.9362. The study conducted case analyses on two diseases heart failure and acute myeloid leukemia. The predicted associated miRNAs can be validated by existing biomedical research efforts. The source code and data of PMDGGM can be publicly accessed on GitHub for further research and verification: https://github.com/WangYeQianger/PMDGGM