Tu-Phuong Mai, Minh-Ha H. Le, Duc-Luong Tran, Phuong-Anh Chu, Duy-Cat Can, Hoang-Quynh Le. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Multi-document summarization for low-resource languages faces a critical trade-off: large language models are computationally prohibitive for most institutions, while smaller models suffer from severe hallucination in abstractive generation. We address this through extractive sentence retrieval, which guarantees faithfulness while operating within constrained computational budgets. Our approach combines language-adaptive mixture-of-experts embeddings with graph neural networks that model discourse structure, addressing linguistic challenges across typologically diverse low-resource languages. With only 3.2M trainable parameters, our model requires 28 times less training time than comparable transformer-based approaches, making it practical for single-GPU environments typical in resource-limited settings. We demonstrate applicability to some main Southeast Asia (SEA) countries including Vietnam, Thailand, Laos, Indonesia, and Malaysia, representing three distinct language families: Austroasiatic, Kra-Dai, and Austronesian.

Case-to-Statute Retrieval—identifying applicable statutes from case facts—is essential for judicial efficiency but remains underexplored in low-resource languages like Vietnamese, where annotated legal corpora are scarce and domain-specific challenges persist. To advance research in this setting, we introduce ViCSR, a new benchmark of 10,000 Vietnamese criminal judgments and 1,122 statutory articles with citation-based relevance labels. We further propose a lightweight two-stage framework that performs efficient candidate retrieval with a fine-tuned Vietnamese bi-encoder and leverages citation structure for reranking with a compact heterogeneous GNN. Experiments show clear improvements over strong sparse and dense baselines, highlighting the importance of both domain adaptation and structure-aware modeling in low-resource legal IR. We will publicly release the benchmark, code, and checkpoints.

Counterfactual explanations (CEs) provide an intuitive way to understand recommender systems by identifying minimal modifications to user-item interactions that alter recommendation outcomes. Existing CE methods for recommender systems, however, have been evaluated under heterogeneous protocols, using different datasets, recommenders, metrics, and even explanation formats, which hampers reproducibility and fair comparison. Our paper systematically reproduces, re-implement, and re-evaluate eleven state-of-the-art CE methods for recommender systems, covering both native explainers (e.g., LIME-RS, SHAP, PRINCE, ACCENT, LXR, GREASE) and specific graph-based explainers originally proposed for GNNs. Here, a unified benchmarking framework is proposed to assess explainers along three dimensions: explanation format (implicit vs. explicit), evaluation level (item-level vs. list-level), and perturbation scope (user interaction vectors vs. user-item interaction graphs). Our evaluation protocol includes effectiveness, sparsity, and computational complexity metrics, and extends existing item-level assessments to top-K list-level explanations. Through extensive experiments on three real-world datasets and six representative recommender models, we analyze how well previously reported strengths of CE methods generalize across diverse setups. We observe that the trade-off between effectiveness and sparsity depends strongly on the specific method and evaluation setting, particularly under the explicit format; in addition, explainer performance remains largely consistent across item level and list level evaluations, and several graph-based explainers exhibit notable scalability limitations on large recommender graphs. Our results refine and challenge earlier conclusions about the robustness and practicality of CE generation methods in recommender systems: https://github.com/L2R-UET/CFExpRec.

Recommendation systems have faced significant challenges in cold-start scenarios, where new items with a limited history of interaction need to be effectively recommended to users. Though multimodal data (e.g., images, text, audio, etc.) offer rich information to address this issue, existing approaches often employ simplistic integration methods such as concatenation, average pooling, or fixed weighting schemes, which fail to capture the complex relationships between modalities. Our study proposes a novel Mixture of Experts framework for multimodal cold-start recommendation (MAMEX), which dynamically leverages latent representation from different modalities. MAMEX utilizes modality-specific expert networks and introduces a learnable gating mechanism that adaptively weights the contribution of each modality based on its content characteristics. This approach enables MAMEX to emphasize the most informative modalities for each item while maintaining robustness when certain modalities are less relevant or missing. Extensive experiments on benchmark datasets show that MAMEX outperforms state-of-the-art models with superior accuracy and adaptability.

Bundle recommendation aims to suggest a set of interconnected items to users. However, diverse interaction types and sparse interaction matrices often pose challenges for previous approaches in accurately predicting user-bundle adoptions. Inspired by the distant supervision strategy and generative paradigm, we propose BRIDGE, a novel framework for bundle recommendation. It consists of two main components, namely the item-sensitive instruction generation and the pseudo bundle generation modules. Inspired by the distant supervision approach, the former is to generate more auxiliary information, e.g., sampled item-sensitive instruction, for training without using external data. This information is subsequently aggregated with collaborative signals from user historical interactions to create pseudo ‘ideal’ bundles. This capability allows BRIDGE to explore all aspects of bundles, rather than being limited to existing real-world bundles. It effectively bridging the gap between user imagination and predefined bundles, hence improving the bundle recommendation performance. Experimental results and analyses validate the superiority of BRIDGE over state-of-the-art methods across four benchmark datasets. Our implementation is available at https://github.com/Rec4Fun/BRIDGE.

Existing studies on bundle construction have relied merely on user feedback via bipartite graphs or enhanced item representations using semantic information. These approaches fail to capture elaborate relations hidden in real-world bundle structures, resulting in suboptimal bundle representations. To overcome this limitation, we propose RaMen, a novel method that provides a holistic multi-strategy approach for bundle construction. RaMen utilizes both intrinsic (characteristics) and extrinsic (collaborative signals) information to model bundle structures through Explicit Strategy-aware Learning (ESL) and Implicit Strategy-aware Learning (ISL). ESL employs task-specific attention mechanisms to encode multi-modal data and direct collaborative relations between items, thereby explicitly capturing essential bundle features. Moreover, ISL computes hyperedge dependencies and hypergraph message passing to uncover shared latent intents among groups of items. Integrating diverse strategies enables RaMen to learn more comprehensive and robust bundle representations. Meanwhile, Multi-strategy Alignment & Discrimination module is employed to facilitate knowledge transfer between learning strategies and ensure discrimination between items/bundles. Extensive experiments demonstrate the effectiveness of RaMen over state-of-the-art models on various domains, justifying valuable insights into complex item set problems.

Experimental performance on the task of relation classification has generally improved using deep neural network architectures.One major drawback of reported studies is that individual models have been evaluated on a very narrow range of datasets, raising questions about the adaptability of the architectures, while making comparisons between approaches difficult.In this work, we present a systematic large-scale analysis of neural relation classification architectures on six benchmark datasets with widely varying characteristics.We propose a novel multi-channel LSTM model combined with a CNN that takes advantage of all currently popular linguistic and architectural features.Our 'Man for All Seasons' approach achieves state-of-the-art performance on two datasets.More importantly, in our view, the model allowed us to obtain direct insights into the continued challenges faced by neural language models on this task.

Motivation: Recognition of biomedical named entities in the textual literature is a highly challenging research topic with great interest, playing as the prerequisite for extracting huge amount of high-valued biomedical knowledge deposited in unstructured text and transforming them into well-structured formats. Long Short-Term Memory (LSTM) networks have recently been employed in various biomedical named entity recognition (NER) models with great success. They, however, often did not take advantages of all useful linguistic information and still have many aspects to be further improved for better performance. Results: We propose D3NER, a novel biomedical named entity recognition (NER) model using conditional random fields and bidirectional long short-term memory improved with fine-tuned embeddings of various linguistic information. D3NER is thoroughly compared with seven very recent state-of-the-art NER models, of which two are even joint models with named entity normalization (NEN), which was proven to bring performance improvements to NER. Experimental results on benchmark datasets, i.e. the BioCreative V Chemical Disease Relation (BC5 CDR), the NCBI Disease and the FSU-PRGE gene/protein corpus, demonstrate the out-performance and stability of D3NER over all compared models for chemical, gene/protein NER and over all models (without NEN jointed, as D3NER) for disease NER, in almost all cases. On the BC5 CDR corpus, D3NER achieves F1 of 93.14 and 84.68% for the chemical and disease NER, respectively; while on the NCBI Disease corpus, its F1 for the disease NER is 84.41%. Its F1 for the gene/protein NER on FSU-PRGE is 87.62%. Availability and implementation: Data and source code are available at: https://github.com/aidantee/D3NER. Supplementary information: Supplementary data are available at Bioinformatics online.