Multilingual LAMA: Investigating Knowledge in Multilingual Pretrained Language Models

Recently, it has been found that monolingual English language models can be\nused as knowledge bases. Instead of structural knowledge base queries, masked\nsentences such as "Paris is the capital of [MASK]" are used as probes. We\ntranslate the established benchmarks TREx and GoogleRE into 53 languages.\nWorking with mBERT, we investigate three questions. (i) Can mBERT be used as a\nmultilingual knowledge base? Most prior work only considers English. Extending\nresearch to multiple languages is important for diversity and accessibility.\n(ii) Is mBERT's performance as knowledge base language-independent or does it\nvary from language to language? (iii) A multilingual model is trained on more\ntext, e.g., mBERT is trained on 104 Wikipedias. Can mBERT leverage this for\nbetter performance? We find that using mBERT as a knowledge base yields varying\nperformance across languages and pooling predictions across languages improves\nperformance. Conversely, mBERT exhibits a language bias; e.g., when queried in\nItalian, it tends to predict Italy as the country of origin.\n