Disentangling dialects: a neural approach to Indo-Aryan historical phonology and subgrouping

Proceedings of the 24th Conference on Computational Natural Language Learning, pages 620–630 Online, November 19-20, 2020. c⃝2020 Association for Computational Linguistics https://doi.org/10.18653/v1/P17 620 Disentangling dialects: a neural approach to Indo-Aryan historical phonology and subgrouping Chundra A. Cathcart1,2 and Taraka Rama3 1Department of Comparative Language Science, University of Zurich 2Center for the Interdisciplinary Study of Language Evolution, University of Zurich 3Department of Linguistics, University of North Texas chundra.cathcart@uzh.ch, taraka.kasicheyanula@unt.edu Abstract This paper seeks to uncover patterns of sound change across Indo-Aryan languages using an LSTM encoder-decoder architecture. We aug- ment our models with embeddings represent- ing language ID, part of speech, and other fea- tures such as word embeddings. We find that a highly augmented model shows highest accu- racy in predicting held-out forms, and inves- tigate other properties of interest learned by our models’ representations. We outline exten- sions to this architecture that can better capture variation in Indo-Aryan sound change. 1 Introduction The Indo-Aryan languages, comprising Sanskrit (otherwise known as Old Indo-Aryan, or OIA) and its descendant languages, including medieval languages like P¯al.i and modern languages such as Hindi/Urdu, Panjabi, and Bangla, form a well- studied subgroup of the Indo-European language family. At the same time, many aspects of the Indo-Aryan languages’ history remain poorly un- derstood. One reason is that there are large histori- cal gaps in the attestation of Indo-Aryan languages, making it challenging to document when certain shared innovations took place. Additionally, while the operation of sound changes are a diagnostic for subgrouping that historical linguistic often em- ploy, Indo-Aryan languages have remained in close contact for millennia, borrowing words from each other and making it difficult to establish subgroup- defining sound laws using the traditional compara- tive method of historical linguistics. While a number of large digitized multilingual resources pertaining to the Indo-Aryan languages exist, these data sets have not been widely used in studies, and our understanding of Indo-Aryan dialectology stands to benefit greatly from the ap- plication of deep learning techniques. This paper seeks to move further towards closing this gap. We use an LSTM-based encoder-decoder architecture to analyze a large data set of OIA etyma (ancestral forms) and medieval/modern Indo-Aryan reflexes (descendant forms) extracted from a digitized et- ymological dictionary, with the goal of inferring patterns of sound change from input/output string pairs. We use language embeddings with the goal of capturing individual languages’ historical phono- logical behavior. We augment this basic model with additional embeddings that may help in capturing irregular patterns of sound change not captured by language embeddings; additionally, we compare the performance of these models against a baseline model that is embedding-free. We evaluate the performance of models with dif- ferent embeddings by assessing the accuracy with which held-out forms in medieval/modern Indo- Aryan languages are predicted on the basis of the OIA etyma from which they descend, and carry out a linguistically informed error analysis. We provide a quantitative evaluation of the degree of agree- ment between the genetic signal of each model’s embeddings and a reference …

Content truncated to 500 words