Siamese Convolutional Networks for Cognate Identification

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 1018–1027, Osaka, Japan, December 11-17 2016. Siamese Convolutional Networks for Cognate Identification Taraka Rama Department of Linguistics University of Tübingen, Germany taraka-rama.kasicheyanula@uni-tuebingen.de Abstract In this paper, we present phoneme level Siamese convolutional networks for the task of pair-wise cognate identification. We represent a word as a two-dimensional matrix and employ a siamese convolutional network for learning deep representations. We present siamese architectures that jointly learn phoneme level feature representations and language relatedness from raw words for cognate identification. Compared to previous works, we train and test on larger and realistic datasets; and, show that siamese architectures consistently perform better than traditional linear classifier approach. 1 Introduction Cognates are words that are known to have descended from a common ancestral language. In historical linguistics, identification of cognates is an important step for positing relationships between languages. An example of cognates are German Fuß and English foot whereas, Hindi chakra and English wheel are cognates that can be traced back to the Proto-Indo-European ∗kwekwlo− and do not exhibit similarity on surface. In NLP, automatic identification of cognates is associated with the task of determining if two words are descended from a common ancestor or not. In NLP, word similarity measures based on number of shared bi-grams, minimum-edit-distance, and length of longest common subsequence are supplied as features for a linear classifier or a sequence labeler on a set of labeled positive and negative examples; and then employ the trained classifier to classify new word pairs. The features for a classifier consist of string similarity scores (Hauer and Kondrak, 2011; Inkpen et al., 2005). It has to be noted that the Indo-European dating studies (Bouckaert et al., 2012; Chang et al., 2015; Rama, 2016) employ human expert cognacy judgments for inferring phylogeny and internal dates of a well-studied language family. Therefore, there is a need for developing automated cognate identification methods that can be applied to those families of the world that are not as well-studied as Indo-European language family. The supervised approaches (Kondrak, 2009; Bergsma and Kondrak, 2007) employ orthographic sim- ilarities and character alignments as features for training classifiers. In this work, we show how convo- lutional networks can be employed to extract phonetic features for the purpose of cognate identification. We also include a neural network approach to integrate language features for jointly training the neural networks. To the best of our knowledge, this work is the first to apply convolutional networks (CNN) for the purpose of cognate identification. The work is organized as follows. In section 2, we define the task of cognate identification. In section 3, we motivate and describe convolutional network architectures for cognate identification. In section 4, we describe the related work for cognate identification. We present the experimental setup in section 5 and results in section 6. Finally, we present our conclusions in section 7. This work is licenced under a Creative Commons Attribution 4.0 International License. License details: http:// creativecommons.org/licenses/by/4.0/ …

Content truncated to 500 words