A Comparison of fastText Implementations Using Arabic Text Classification

The quality of word representation is crucial to obtain good results in many natural language processing tasks. Recently, many word representation models (word embeddings), such as fastText, have been developed. In this research, we compared the algorithms for the fastText implementation, Facebook's official implementation, and Gensim's implementation using the same pre-trained fastText model. Using multiclass classification, we evaluated these embeddings. According to the results, the Facebook implementation performed better than Gensim's implementation, with an average accuracy of 78.22% and 56.73%, respectively, for sentence embeddings and an average accuracy of 79.43% and 57.95%, respectively, for word embeddings.