Neural Entity Alignment Analysis

Neural methods have become the de-facto choice for the vast majority of data analysis tasks, and entity alignment is no exception. Not surprisingly, more than 40 different neural entity alignment methods have been published in reputed computer science venues since 2017. However, surprisingly, an in-depth empirical comparison and an analysis of the differences between neural and non-neural entity alignment methods have been lacking. We bridge this gap by performing an in-depth comparison between methods from the pre-neural and neural era. Specifically, we build upon recent benchmarking studies to select one (Paris) and two (RDGCN andBootEA) representative state-of-the-art methods from the pre-neural and neural era, respectively. We unravel and consequently, mitigate the inherent deficiencies in the experimental setup utilized for evaluating neural entity alignment methods. To ensure fairness in evaluation across the two paradigms, we also homogenize the entity matching modules of neural and non-neural methods. Our results indicate that Paris, the state-of-the-art non-neural method, statistically significantly outperforms both RDGCN and BootEA, the state-of-the-art neural methods, in terms of both efficacy and efficiency across a wide variety of dataset types and scenarios. Moreover, our findings shed light on the potential problems resulting from an impulsive application of neural methods as a panacea for all data analytics tasks. Overall, our work results in two overarching conclusions: (1) Paris should be used as a baseline in every follow-up work on entity alignment, and (2) neural methods need to be positioned better to showcase their true potential, for which we provide multiple recommendations.

This dataset contains the full versions of dbpedia, wikidata and yago (v3.1), which have been used to reproduce the results of the paper "A Critical Re-evaluation of Neural Methods for Entity Alignment".

神经网络方法已成为绝大多数数据分析任务的事实上的标准选择，实体对齐（entity alignment）也不例外。不出所料，自2017年以来，已有超过40种不同的神经网络实体对齐方法在知名计算机科学学术发表平台上发表。然而令人意外的是，当前仍缺乏针对神经网络与非神经网络实体对齐方法之间差异的深度实证对比与分析。
本文通过对前神经网络时代与神经网络时代的实体对齐方法开展深度对比，填补了这一研究空白。具体而言，本文依托近期基准研究，分别从前神经网络时代与神经网络时代中选取了1款（Paris）与2款（RDGCN、BootEA）具有代表性的最先进（state-of-the-art）实体对齐方法。本文梳理并纠正了现有神经网络实体对齐方法评估实验中存在的固有缺陷，为确保两种范式下的评估公平性，还对神经网络与非神经网络方法的实体匹配模块进行了同质化处理。
实验结果表明，作为当前最先进的非神经网络实体对齐方法，Paris在各类数据集类型与应用场景下的效能与效率上，均在统计学意义上显著优于RDGCN与BootEA这两款当前最先进的神经网络实体对齐方法。此外，本文的研究结果揭示了将神经网络方法盲目作为万能解药应用于所有数据分析任务所可能引发的潜在问题。
综上，本文得出两项核心结论：其一，在所有后续实体对齐相关研究中，均应将Paris作为基准方法；其二，神经网络方法需要更合理的定位以展现其真实潜力，本文为此提供了多项针对性建议。

本数据集包含DBpedia、Wikidata与YAGO（v3.1）的完整版本，可用于复现论文《实体对齐神经网络方法的批判性重新评估》的实验结果。