Datasets:
named_entity annotations are broken
There is a serious problem with this dataset as provided, to the point that I feel comfortable using a word as strong as "broken" to describe it. Consider for example the following sentence:
from document pt/nt/44/nt_4405
{
"part_id": 0,
"words": [ "There", "was", "a", "man", "named", "Ananias", "." ],
"pos_tags": [ 14, 40, 13, 24, 42, 25, 7 ],
"parse_tree": "(TOP(S(NP (EX There) )(VP (VBD was) (NP(NP (DT a) (NN man) )(VP (VBN named) (S(NP (NNP Ananias) ))))) (. .) ))",
"predicate_lemmas": [ null, "be", null, "man", "name", null, null ],
"predicate_framenet_ids": [ null, "02", null, null, "01", null, null ],
"word_senses": [ null, null, null, null, null, null, null ],
"speaker": "Acts",
"named_entities": [ 0, 0, 0, 0, 0, 0, 0 ],
"srl_frames": [ { "verb": "was", "frames": [ "O", "B-V", "B-ARG1", "I-ARG1", "I-ARG1", "I-ARG1", "O" ] }, { "verb": "named", "frames": [ "O", "O", "B-ARG1", "I-ARG1", "B-V", "B-ARG2", "O" ] } ],
"coref_spans": [ [ 8, 2, 5 ] ]
}
It should be very clear that a sentence There was a man named ___ is about to have a Person as a named entity, and yet the named_entities field for this sentence are all O. In fact, this sentence (and many others like it) simply should not have named_entities fields at all. In the original LDC2013T19 release, there is no nt_4405.name file to go with the text. Whereas many of the annotated text files have corresponding files such as data/files/data/english/annotations/nw/wsj/04/wsj_0417.name, there simply is no such labeling of nt_4405.
I looked into which files did not have corresponding .name files in the original OntoNotes dataset, and it is exactly the set of nt_ files. I propose that all of the sentences with a document_id in the nt_ set of files should have some indication that their named_entities are not labeled, such as replacing named_entities with None or simply not having that field in the dict.