Fig. 3: Comparison of the top five predictions for the BERT (large-uncased version) and Pl@ntBERT (large-species version trained on folds 1–9) models for our sample text of ‘prunus padus, [MASK] and crataegus monogyna are constant species of temperate hardwood riparian forests (T13)’. | Nature Plants

Fig. 3: Comparison of the top five predictions for the BERT (large-uncased version) and Pl@ntBERT (large-species version trained on folds 1–9) models for our sample text of ‘prunus padus, [MASK] and crataegus monogyna are constant species of temperate hardwood riparian forests (T13)’.

From: Learning the syntax of plant assemblages

Fig. 3

The percentages next to each predicted species represent the probabilities assigned by the models for replacing the [MASK] token, normalized so that the top five predictions sum to 100%. On the one hand, the candidates from BERT are all trees, which shows that the model ‘understood’ we are in a forest. However, all of them are common plant names (and not scientific names of taxa) and, except for the oak, which is the last candidate, are not found within the T13 habitat type. On the other hand, the candidates from Pl@ntBERT are all scientific names of constant species from the required habitat type.

Back to article page