Cross-lingual laws of semantic change

Semantic divergence in related languages is a key concern of historical linguistics. In order to complement existing research on semantic change, which in most previous approaches relies on the comparative analysis of diachronic corpora, we propose a novel approach for measuring semantic shifts synchronically. In this chapter, we investigate semantic change across languages by measuring the semantic distance of cognate sets using cross-lingual word embeddings. We define a measure of cognate divergence and show how it can be used as a measure of language semantic divergence. We propose an algorithm to detect and correct false friends as an application for natural language processing. We hypothesize that false friends fall on a spectrum, and define a corresponding notion of “falseness” and a methodology for its quantification. We evaluate the algorithm based on WordNet and on manually curated lists of true cognates and false friends, showing accuracy values exceeding 80%, and we show how choosing a falseness level as a threshold for detecting false friends can affect the performance of the detection algorithm. We further study the properties of the semantic divergence of cognates, and verify whether hypothesized laws of semantic change (namely the law of conformity and the law of innovation) hold in the multilingual setting, thus formulating the first laws of cross-lingual semantic change. We further study the mathematical relation between polysemy and frequency on the one hand, and falseness on the other hand, and identify polynomials that optimally model their relationship, leading to equations describing cross-lingual semantic change in relation to word properties