Fig. 6: The Top-5 words the model relies on for label prediction. | Communications Medicine

Fig. 6: The Top-5 words the model relies on for label prediction.

From: A BERT model generates diagnostically relevant semantic embeddings from pathology synopses with active learning

Fig. 6

The color of each cell represents the L2-normalized importance score of the word. The top-1 words for the majority of labels are their acronyms or their name. For example, the top-1 influence word for "chronic myeloid leukemia'' is “CML” and the top-1 influence word for “metastatic” is “metastatic”. Furthermore, "normal'' has no words with high influence, which corresponds with clinical practice, as when specific no abnormal findings are identified by a hematopathologist, the case is semantically interpreted as normal. In this way, our knockout method provided some insight into the complex and opaque prediction process of the model.

Back to article page