Semantic Drift in Multilingual Representations
This work addresses the issue of understanding semantic drift and language relations in AI models for linguists and NLP researchers, though it is incremental in proposing a new analysis method rather than a breakthrough.
The authors tackled the problem of analyzing semantic relations between languages in computational multilingual representations, and found that these representations, trained only on monolingual text and bilingual dictionaries, can reconstruct phylogenetic trees resembling expert linguistic assumptions without etymological information.
Multilingual representations have mostly been evaluated based on their performance on specific tasks. In this article, we look beyond engineering goals and analyze the relations between languages in computational representations. We introduce a methodology for comparing languages based on their organization of semantic concepts. We propose to conduct an adapted version of representational similarity analysis of a selected set of concepts in computational multilingual representations. Using this analysis method, we can reconstruct a phylogenetic tree that closely resembles those assumed by linguistic experts. These results indicate that multilingual distributional representations which are only trained on monolingual text and bilingual dictionaries preserve relations between languages without the need for any etymological information. In addition, we propose a measure to identify semantic drift between language families. We perform experiments on word-based and sentence-based multilingual models and provide both quantitative results and qualitative examples. Analyses of semantic drift in multilingual representations can serve two purposes: they can indicate unwanted characteristics of the computational models and they provide a quantitative means to study linguistic phenomena across languages. The code is available at https://github.com/beinborn/SemanticDrift.