Cross-lingual Entity Alignment with Incidental Supervision

Abstract

Much research effort has been put to multilingual knowledge graph (KG) embedding methods to address the entity alignment task, which seeks to match entities in different languagespecific KGs that refer to the same real-world object. Such methods are often hindered by the insufficiency of seed alignment provided between KGs. Therefore, we propose an incidentally supervised model, JEANS , which jointly represents multilingual KGs and text corpora in a shared embedding scheme, and seeks to improve entity alignment with incidental supervision signals from text. JEANS first deploys an entity grounding process to combine each KG with the monolingual text corpus. Then, two learning processes are conducted: (i) an embedding learning process to encode the KG and text of each language in one embedding space, and (ii) a selflearning based alignment learning process to iteratively induce the matching of entities and that of lexemes between embeddings. Experiments on benchmark datasets show that JEANS leads to promising improvement on entity alignment with incidental supervision, and significantly outperforms state-of-the-art methods that solely rely on internal information of KGs. 1

XLM-K: Improving Cross-Lingual Language Model Pre-training with Multilingual Knowledge

“…For cross-lingual word embeddings, most of works rely on aligned words or sentences (Ruder, Vulić, and Søgaard 2019). Cao et al 2018b;Pan et al 2019;Chen et al 2021…”

“…In fact, the knowledge conveys similar semantic concepts and similar meanings across languages (Vulić and Moens 2013;Chen et al 2021), which is essential to achieve crosslingual transferability. Therefore, how to equip pre-trained models with knowledge has become an underexplored but critical challenge for multilingual language models.…”

ClusterEA: Scalable Entity Alignment with Stochastic Training and Normalized Mini-batch Similarities

“…First, side information may not be available due to privacy concerns, especially for industrial applications [23,25,26]. Second, models that incorporating machine translation or prealigned word embeddings may be overestimated due to the name bias issue [5,21,22,28]. Thus, compared with the models employing side information, the structure-only methods are more general and not affected by bias of benchmarks.…”

From Alignment to Assignment: Frustratingly Simple Unsupervised Entity Alignment

“…We compare our method against the following three groups of advanced EA methods: (1) Structure: These methods only use the structure information (i.e., triples): GCN-Align (Wang et al, 2018), MuGNN (Cao et al, 2019a), BootEA (Sun et al, 2018), MRAEA (Mao et al, 2020), JEANS (Chen et al, 2021). ( 2) Word-level: These methods average the pre-trained entity name vectors to construct the initial features: GM-Align , RDGCN (Wu et al, 2019a), HGCN (Wu et al, 2019b), DAT (Zeng et al, 2020b), DGMC (Fey et al, 2020).…”