Rethinking Uncertainly Missing and Ambiguous Visual Modality in Multi-Modal Entity Alignment

TL;DR

This paper analyzes the challenges of missing and ambiguous visual data in multi-modal entity alignment, introduces a robust method called UMAEA, and demonstrates its superior performance on a new benchmark dataset.

Contribution

The paper presents UMAEA, a novel approach that effectively handles visual modality incompleteness in multi-modal entity alignment, outperforming existing methods on a comprehensive benchmark.

Findings

UMAEA achieves state-of-the-art results on 97 benchmark splits.

Models tend to overfit noise and perform poorly with high missing modality rates.

Additional multi-modal data can sometimes harm entity alignment performance.

Abstract

As a crucial extension of entity alignment (EA), multi-modal entity alignment (MMEA) aims to identify identical entities across disparate knowledge graphs (KGs) by exploiting associated visual information. However, existing MMEA approaches primarily concentrate on the fusion paradigm of multi-modal entity features, while neglecting the challenges presented by the pervasive phenomenon of missing and intrinsic ambiguity of visual images. In this paper, we present a further analysis of visual modality incompleteness, benchmarking latest MMEA models on our proposed dataset MMEA-UMVM, where the types of alignment KGs covering bilingual and monolingual, with standard (non-iterative) and iterative training paradigms to evaluate the model performance. Our research indicates that, in the face of modality incompleteness, models succumb to overfitting the modality noise, and exhibit performance oscillations or declines at high rates of missing modality. This proves that the inclusion of additional multi-modal data can sometimes adversely affect EA. To address these challenges, we introduce UMAEA , a robust multi-modal entity alignment approach designed to tackle uncertainly missing and ambiguous visual modalities. It consistently achieves SOTA performance across all 97 benchmark splits, significantly surpassing existing baselines with limited parameters and time consumption, while effectively alleviating the identified limitations of other models. Our code and benchmark data are available at https://github.com/zjukg/UMAEA.