How to interpret Allen NLP Coreference resolution model output?

I am working on extracting people and tasks from texts (multiple sentences) and need a way to resolve coreferencing. I found this model, and it seems very promising, but once I installed the required libraries allennlp and allennlp_models and testing the model out for myself I got:

Script:

predictor = Predictor.from_path("https://storage.googleapis.com/allennlp-public-models/coref-spanbert-large-2021.03.10.tar.gz")
prediction = predictor.predict(
    document="Paul Allen was born on January 21, 1953, in Seattle, Washington, to Kenneth Sam Allen and Edna Faye Allen. Allen attended Lakeside School, a private school in Seattle, where he befriended Bill Gates, two years younger, with whom he shared an enthusiasm for computers.")
print(prediction)

Output:

{'top_spans': [[0, 1], [3, 3], [5, 8], [5, 14], [8, 8], [11, 13], [11, 14], [13, 13], [16, 18], [16, 22], [20, 22], [24, 24], [26, 52], [33, 33], [36, 36], [37, 37], [38, 52], [41, 42], [47, 47], [48, 48], [49, 52]], 
 'antecedent_indices': [[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20], [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20], [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20], [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20], [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20], [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20], [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20], [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20], [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20], [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20], [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20], [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20], [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20], [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20], [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20], [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20], [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20], [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20], [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20], [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20], [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20]], 
 'predicted_antecedents': [-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, 0, -1, 5, 11, -1, -1, -1, 11, -1, -1], 
 'document': ['Paul', 'Allen', 'was', 'born', 'on', 'January', '21', ',', '1953', ',', 'in', 'Seattle', ',', 'Washington', ',', 'to', 'Kenneth', 'Sam', 'Allen', 'and', 'Edna', 'Faye', 'Allen', '.', 'Allen', 'attended', 'Lakeside', 'School', ',', 'a', 'private', 'school', 'in', 'Seattle', ',', 'where', 'he', 'befriended', 'Bill', 'Gates', ',', 'two', 'years', 'younger', ',', 'with', 'whom', 'he', 'shared', 'an', 'enthusiasm', 'for', 'computers', '.'], 
 'clusters': [[[0, 1], [24, 24], [36, 36], [47, 47]], [[11, 13], [33, 33]]]}

I'm having trouble interpreting the format of this output. I was expecting something like

{entity_0_spans: [LIST_OF_INDEX_TUPLES],  # Paul Allen in this example
 entity_1_spans: [LIST_OF_INDEX_TUPLES],  # Seattle in this example
 ...}

or something that more closely resembles the visualisation available on the demo page:

I've looked through https://demo.allennlp.org/coreference-resolution but couldn't find a breakdown of how to use the model output yet - can anyone suggest some resources that will help me? Any pointers are much appreciated!

Solution

The information you are looking for is in 'clusters', where each list corresponds to an entity. Within each entity list, you will find the mentions referring to the same entity. The number are indices that mark the beginning and ending of each coreferential mention. E.g. Paul Allen [0,1] and Allen [24, 24].