Returns ------- A tuple of the following metrics based on the accumulated count statistics: precision : float

The input is a dict whose items are batched tensors. The output is a list of dictionaries - one per entry in the batch - with the slices of t

Converts the list of spans and predicted antecedent indices into clusters of spans for each element in the batch. Parameters

For scoring the argument, we don't need the trigger spans to match exactly. We just need the trigger label corresponding to the predicted tri

Get cluster metadata in form to feed into evaluation scripts. For each entry in minibatch, return a dict with a metadata field, which

Performs division and handles divide-by-zero. On zero-division, sets the corresponding result elements to zero.

Randomly permute the documents for each batch

Computes an embedding representation of pairs of spans for the pairwise scoring function to consider. This includes both the original

Compute trigger scores for all tokens.

Given span start and end (inclusive), get the context on either side.

Compute cross-entropy loss on relation labels. Ignore diagonal entries and entries giving relations between masked out spans.

Loop over each slice and get the labels for the spans from that slice. All labels are offset by 1 so that the "null" label gets class

Run the forward pass. Since we can only have coreferences between spans in the same document, we loop over the documents in the batch

Given a document and a character in the document, get the token that the char lives in.

Create a pruner that either takes outputs of other scorers (i.e. entity beam), or uses its own scorer (the `default_scorer`).