Interpreting results about spert HOT 2 CLOSED

Hi,

regarding the output interpretation:
entity_clf is a (B x S x E) tensor where B is the batch size (in sentences), S the count of all token spans up to a specified length (10 per default) and E the number of entity types (+1 for "None"). It contains the model's (softmax) confidences that a given span belongs to a certain type of E. In case the span is assigned to the "None" type (=no entity), it is disregarded in the relation extraction step.

rel_clf is a (B x P x R) tensor where B is again the batch size, P the count of all entity pairs (= spans not assigned to the None class) and R the number of relation types. For each entity pair, it contains the (sigmoid) scores for each relation type.

rels is a (B x P x 2) tensor that contains the corresponding entity indices (in entity_clf, entity_masks, entity_sizes etc.) for each entity pair. With this, you can for example access the corresponding entity scores (by indexing entity_clf with rels).

Regarding your second question: It is (span start, span end, entity type, score). Here 'span end' is exclusive. Also, it corresponds to BPE tokens (byte-pair encoded, as in BERT), not to raw tokens.

from spert.

Thank you Markus for your help

from spert.