Generating a concept relation network for Turkish based on conceptnet using translational methods

Abstract

ConceptNet is a large-scale network of concepts and relationships, based on var ious common sense knowledge bases and built upon more than 700 thousand sentences contributed by approximately 15 thousand authors. It was originally developed for the English language and later became a multilingual tool with the addition of other languages using many different sources. It can be seen as a database of how different concepts relate to each other, especially as a valuable resource for systems that perform text analyses, meaning or context extraction. Turkish is a language that lacks similar sources for processing texts and extracting meaning. Although ConceptNet includes examples for Turkish, not many are available where both concepts are in Turkish. This study discusses various methods to create a Turkish ConceptNet using translational techniques based on English ConceptNet and explains the results herewith obtained. Multiple models are tested, using different sources including WordNet, Wikipedia and Google Translate. Results obtained from each model and approaches to improve these results are discussed, while also explaining details, assumptions and drawbacks relevant to each relation.