Measuring Spelling Similarity for Cognate Identification

The most commonly used measures of string similarity, such as the Longest Common Subsequence Ratio (LCSR) and those based on Edit Distance, only take into account the number of matched and mismatched characters. However, we observe that cognates belonging to a pair of languages exhibit recurrent spelling differences such as “ph” and “f” in English-Portuguese cognates “phase” and “fase”. Those differences are attributable to the evolution of the spelling rules of each language over time, and thus they should not be penalized in the same way as arbitrary differences found in non-cognate words, if we are using word similarity as an indicator of cognaticity.

This paper describes SpSim, a new spelling similarity measure for cognate identification that is tolerant towards characteristic spelling differences that are automatically extracted from a set of cognates known apriori. Compared to LCSR and EdSim (Edit Distance-based similarity), SpSim yields an F-measure 10% higher when used for cognate identification on five different language pairs.