Fig 1.
The main framework of BERT.
Fig 2.
Given parallel data from two languages, a student model can be trained such that the generated vectors for the two languages sentences are close to the teacher language sentence vector.
Table 1.
Examples of different levels of correlation between the sentences in STS dataset.
Table 2.
Some examples of the proposed translations along with original English sentences.
Fig 3.
A framework of models generation using the third approach.
Table 3.
Accuracy of machine translation based and interleaved MSA models tested based on Spearman rank correlation between the cosine similarity of sentence representations and the reference labels of the testing dataset in [8].
Table 4.
Accuracy of knowledge distillation-based MSA models tested based on Spearman rank correlation between the cosine similarity of sentence representations and the reference labels of the testing dataset in [8].
Table 5.
Accuracy of main Egyptian models tested based on Spearman rank correlation between the cosine similarity of sentence representations and the reference labels of the testing dataset in [8] after translation to Egyptian Arabic.
Table 6.
Accuracy of main Saudi Arabian models based on Spearman rank correlation between the cosine similarity of sentence representations and the reference labels of the testing dataset in [8] after translation to Saudi Arabic.
Table 7.
Comparisons between the proposed models and current state-of-the-art Arabic STS models based on Spearman rank correlation between the cosine similarity of sentence representations and the reference labels of the testing dataset in [8].