Abstract
Multi-view document clustering (MvDC) is a critical task for organizing cross-lingual information. However, existing deep learning models often fail to capture intermediate semantic relationships and treat all data instances with equal importance, leading to suboptimal performance on difficult cases. This paper presents the Ambiguous Instance-Aware Contrastive Network with Multi-Level Matching (AICN-MLM). By integrating multi-level alignment—from raw feature distributions to high-level cluster prototypes—and introducing a weight modulation mechanism for ambiguous instances, AICN-MLM achieves superior clustering accuracy and robustness across diverse multilingual datasets.
Introduction
The goal of multi-view document clustering is to group semantically related documents that appear in different "views," such as different languages (e.g., English vs. Thai). Despite recent progress, two significant challenges persist:
- Single-Perspective Alignment: Most existing models align views either at the feature level or the label level, neglecting the rich information available at intermediate stages.
- The "Equal Treatment" Fallacy: Standard contrastive learning assigns equal weight to all document pairs. This approach often over-optimizes for "easy" pairs while failing to learn the discriminative features necessary to resolve "ambiguous" instances that are difficult to classify.
Proposed Methodology: AICN-MLM
The AICN-MLM framework addresses these challenges through three integrated modules:
1. Multi-view Data Reconstruction
The model utilizes autoencoders to compress high-dimensional text data into low-dimensional latent feature vectors. This process filters out noise and extracts the most salient features for downstream clustering.
2. Multi-Level Matching Module (MLM)
To ensure comprehensive alignment across views, the MLM module operates on two levels:
- Similarity Distribution Matching (SDM): This sub-module uses Maximum Mean Discrepancy (MMD) to align the statistical distributions of document similarities within each view.
- Cross-View Prototype Matching (CVPM): To resolve "prototype shift," where cluster centers in different languages do not naturally align, CVPM uses Jensen-Shannon (JS) divergence to match cluster prototypes across views.
3. Ambiguous Instance-Aware Contrastive Learning (AIACL)
The AIACL module introduces a dynamic weight modulation function. By identifying "ambiguous" instances—pairs where the model's confidence is low—the network increases their importance during training. Conversely, "clear" instances are down-weighted. This forces the model to focus on the most informative and difficult data points, sharpening its discriminative boundaries.
Experimental Evaluation
The authors evaluated AICN-MLM on two major benchmarks:
- KUST: A multilingual dataset comprising seven language pairs (e.g., English-Thai, Chinese-Vietnamese), each with 10,000 instances across 10 themes.
- Reuters: A standard public dataset containing English and French documents across six classes.
Results and Discussion
AICN-MLM consistently outperformed over 10 state-of-the-art methods, including DEMVC and CVCL. On the KUST-ETD (English-Thai) dataset, the model achieved an Accuracy (ACC) of 89.08%, representing a significant improvement over the previous state-of-the-art (~79.32%).
Ablation studies confirmed the importance of each component. Specifically, removing the Ambiguous Instance (AI) component led to a drop in accuracy from 89% to 83% on KUST-ETD, highlighting that focusing on hard cases is a primary driver of success.
Conclusion
The AICN-MLM framework sets a new benchmark for multi-view document clustering by treating ambiguity as an asset rather than a hindrance. By aligning views across multiple levels and prioritizing difficult instances, the model achieves high accuracy and robustness. Future work will explore the scalability of this approach to "many-view" scenarios involving dozens of languages simultaneously.
Summary updated