Skip to content

Data format

Jump to bottom Edit New page
Ben Hachey edited this page Mar 8, 2014 · 2 revisions

To evaluate your system output against the gold-standard, you will need to output in tab-separated format.

  • Each document is started with the -DOCSTART- (some_doc_id) line, where some_doc_id might be something like 1163testb SOCCER
  • Each sentence is separated by a blank line
  • Each document is separated by a blank line
  • Each token is on its own line (we re-use the gold-standard tokenisation)

The column ordering for token lines is:

  • Token
  • Mention span: B for mention begin, I for inside mention, empty column for outside
  • Mention text: this is a bit redundant, but a sanity check when reading the output (text == ' '.join(mentiontokens)
  • Entity identifier: where a mention is linked to the KB, this will be the id/title (e.g., a Wikipedia title). Where the mention is a NIL, this column should be blank
-DOCSTART- (some_doc_id)
Some
headline
about
two
Named	B	Named Entities	Named_Entity
Entities	I	Named Entities	Named_Entity
.

By
John	B	John Smith
Smith	I	John Smith
Add a custom footer
Add a custom sidebar
Clone this wiki locally