Parses a lex attribute (might be nested) into a list of basic lex form (i.e., no nested lex)

Give each coref the index of it parent in the list of corefs.

Remove items that didn't get spans, don't have an allowed coref type, or weren't assigned a cluster

Generates a binary indicator for every pair of spans. This label is one if and only if the pair of spans belong to the same cluster.

Computes scores for every pair of spans. Additionally, a dummy label is included, representing the decision that the span is not core

Run the forward pass for a single document. Important: This function assumes that sentences are going to be passed in in sorted orde

TODO(dwadden) document me and add comments.

TODO(dwadden) document me and add comments.

Create trigger / argument pair embeddings, consisting of: - The embeddings of the trigger and argument pair. - Optionally, th

Given bioul predictions of one document, return entities in the span format

Flatten the coref labels.

Spans are input with each minibatch as a sentence. For coref, it's easier to flatten them out and consider all sentences together as

Spans are input with each minibatch as a sentence. For coref, it's easier to flatten them out and consider all sentences together as

This method generates possible antecedents per span which survived the pruning stage. This procedure is `generic across the batch`. T

Compute cross-entropy loss on argument labels.

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

Get text span of coref. We have inclusive endpoints.

Load in text and strip out tags.

Convert to nested list of cluster spans, as in scierc data.

Add a few special cases to spacy tokenizer so it works with ACe mistakes.

Randomly permute the documents for each batch

Randomly permute the documents for each batch

Given bioul predictions of a list of documents, return three tensors, spans, span labels, and mask

Make sure all entries in dict have same length.

Make sure all entries in dict have same length.

Split cluster dict into clusters in current sentence, and clusters that come later.

Create a dataset mapping the ner id's to the coref id's and vice versa.

Largely copy-pasted from what happens in dygie.

Largely copy-pasted from what happens in dygie.