Assumes `xs` is a 1-d array. The usual cumsum has elements [x[1], x[1] + x[2], ...]. This one has elements [0, x[1], x[1] + x[2], ...]. U

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

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

Search through coreference clusters and return the one containing the query entity, if it's part of a cluster. If we don't find a mat

Flatten a nested list into a single list. Note this only works for lists nested one deep. For the general case we'd need recursion

TODO(dwadden) change this.

TODO(dwadden) Write documentation. The trigger embeddings are just the contextualized token embeddings, and the trigger mask is

Extracts the top-k scoring items with respect to the scorer. We additionally return the indices of the top-k in their original order,

TODO(dwadden) Write documentation.

TODO(dwadden) Write documentation.

Get the different types of coref in the GENIA data set, and save to file. Once this is done once, just hard-code the list of coref types.

Get all metrics from all modules. For the ones that shouldn't be displayed, prefix their keys with an underscore.

Get all metrics from all modules. For the ones that shouldn't be displayed, prefix their keys with an underscore.

Get the number of batches.

Make each line of a file into a list entry. Apply a conversion function to each line (e.g. convert to int). By default, the conversion does n

Finds all bibliography entries in a grobid xml.

Save a list as text, one entry per line.

To compute coreference evaluation metrics, the evaluator needs access to the list of coreference clusters, given in the same form as