Suppose I have several journal articles. I would like to train a binary classifier on whether the journal article is insightful. NLP models such as BERT certainly fit my need by scanning the whole article.
To improve the accuracy, I would like to further use the tabular data and figures in the article. However, the number of figures and tables varies across different articles. So does the content of them.
How can I utilize these data, when the number and its content are variable?