An Expectation Maximisation Algorithm for Automated Cognate Detection

Proceedings of the 24th Conference on Computational Natural Language Learning, pages 476–485 Online, November 19-20, 2020. c⃝2020 Association for Computational Linguistics https://doi.org/10.18653/v1/P17 476 An Expectation Maximisation Algorithm for Automated Cognate Detection Roddy MacSween Homerton College & Computer Laboratory University of Cambridge rlm72@cantab.ac.uk Andrew Caines Paula Buttery Computer Laboratory University of Cambridge {apc38|pjb48}@cam.ac.uk Abstract In historical linguistics, cognate detection is the task of determining whether sets of words have common etymological roots. Inspired by the comparative method used by human lin- guists, we develop a system for automated cog- nate detection that frames the task as an infer- ence problem for a general statistical model consisting of observed data (potentially cog- nate pairs of words), latent variables (the cog- nacy status of pairs) and unknown global pa- rameters (which sounds correspond between languages). We then give a specific instance of such a model along with an expectation- maximisation algorithm to infer its parameters. We evaluate our system on a dataset of 8140 cognate sets, finding its performance of our method to be comparable to the state of the art. We additionally carry out qualitative anal- ysis demonstrating various advantages it has over existing systems. We also suggest several ways our work could be extended within the general theoretical framework we propose. 1 Introduction In historical linguistics, two words are deemed cog- nate if they share a root in a parent language, for ex- ample German ‘Nacht’ and English ‘night’ which both originate from Proto-Indo-European *n´okwts. The task of cognate detection is interesting in its own right, but is also an important part of the larger task of proto-language reconstruction. As described in Campbell (2013), the compar- ative method used by historical linguists to recon- struct an ancestral language from a set of poten- tially cognate words in daughter languages con- sists of three steps: assembling cognates, estab- lishing sound correspondences and reconstructing proto-sounds. Automation of this method is made challenging by the fact that “the comparative lin- guist typically jumps back and forth among these steps” because of a mutual dependence between cognacy judgements and hypothesised sound corre- spondences. Whether two words should be deemed cognate depends on whether the sounds in them correspond according to known rules for the lan- guages. For instance, the t and ts sounds in English and German are known to correspond, which is ev- idence for “tooth” and “Zahn” being cognate. But sound correspondence rules are themselves theo- rised based on which sounds coincide in known cognate pairs. Therefore a linguist must iteratively adapt their hypotheses about which words are cog- nate and which sounds correspond until they can reach a definite conclusion. Existing approaches to automated cognate detec- tion (ACD) fail to fully capture this idea of dealing with mutual dependence using an iterative method. Some early approaches are not iterative at all, while several more recent methods are iterative to some extent but either only carry out a small fixed num- ber of iterations or use incomplete and ad hoc meth- ods to …

Content truncated to 500 words