Post-hoc analysis of Arabic transformer models.
<p>Analyze the internal representations of Arabic transformer models.</p>
<p>While there’ve been an extrinsic evaluation of Arabic transformer (AT) models, <em>no work has been carried out to analyze their internal representations.</em><br />
This<a href="https://arxiv.org/abs/2210.09990" rel="noopener ugc nofollow" target="_blank"> paper</a> probe how <strong>Arabic linguistic information</strong> is encoded in AT models. The authors performed a <strong>layer & neuron analysis</strong> on the models using <strong>morphological tagging</strong> <strong>tasks</strong> for different <strong>dialects </strong>and a <strong>dialectal identification task</strong>.</p>
<p>The overall idea is to extract feature vectors from the learned representations and train probing classifiers towards understudied auxiliary tasks (of predicting morphology or identifying dialect).<br />
Additionally, they used <strong>the Linguistic Correlation Analysis method </strong>to identify salient neurons with respect to a downstream task.<br />
The analysis enlightens interesting findings such as:</p>
<p><a href="https://medium.com/@Mustafa77/post-hoc-analysis-of-arabic-transformer-models-95790745c712"><strong>Website</strong></a></p>