Developing a Large Benchmark Corpus for Urdu Semantic Word Similarity

Abstract

The semantic word similarity task aims to quantify the degree of similarity between a pair of words. In literature, efforts have been made to create standard evaluation resources to develop, evaluate, and compare various methods for semantic word similarity. The majority of these efforts focused on English and some other languages. However, the problem of semantic word similarity has not been thoroughly explored for South Asian languages, particularly Urdu. To fill this gap, this study presents a large benchmark corpus of 518 word pairs for the Urdu semantic word similarity task, which were manually annotated by 12 annotators. To demonstrate how our proposed corpus can be used for the development and evaluation of Urdu semantic word similarity systems, we applied two state-of-the-art methods: (1) a word embedding–based method and (2) a Sentence Transformer–based method. As another major contribution, we proposed a feature fusion method based on Sentence Transformers and word embedding methods. The best results were obtained using our proposed feature fusion method (the combination of best features of both methods) with a Pearson correlation score of 0.67. To foster research in Urdu (an under-resourced language), our proposed corpus will be free and publicly available for research purposes.