Notes on APF Format 2004

(Lance Ramshaw, BBN Technologies, summarizing the results of an
ACE planning meeting held at LDC on 2004-02-23.)

ACE's "APF" format has been modified somewhat from that of previous
years to better represent the distinctions made in this year's
annotation.  The major changes are summarized briefly here; for more
detail, refer to the LDC website.

Some changes could be predicted given that the target types have
expanded.  There are two new entity types, entities now have subtypes,
and the entire set of relation types and subtypes has been reworked.

In addition to the expansion of the official ACE task, the LDC has
also chosen this year as a matter of annotation strategy to mark
additional distinctions that they believe promote annotator
consistency, as described in their guidelines documents.  For example,
the LDC is using a more detailed catalogue of mention types, and is
marking a distinction between referential and attributive mentions.
These additional, non-target distinctions are recorded in the APF
using optional attributes and elements whose names begin with "LDC".
For example, the ACE mention type of "cats" in "cats make nice pets"
would be NOM, but the APF also records the LDCTYPE, which is BAR (for
bare nominal).  These "LDC" attributes are left in the APF for
informational purposes, but the scorer ignores them, and ACE systems
are not expected to output them.

The set of ACE mention types in this release is {NAM, NOM, PRO, PRE},
and they map to the more detailed LDC types as follows:
  NAM - NAM
  NOM - NOM, BAR, MWH, 
  PRO - PRO, WHQ, HLS, PTV
  PRE - PRE

The PRE type covers prenominal mentions without distinguishing
between names and nominals.  Note that the sponsors are still
exploring this issue, and the PRE type may go away in future releases.

Where we used to have a boolean GENERIC attribute on entities, we now
have a three-way entity class distinction, with possible values SPC
(specific), GEN (generic), and USP (underspecified), as described in
the LDC guidelines.

For metonymy, base and intended types are no longer being marked.  If
the same entity (Macdonalds, say) is used in a document both as a FAC
and as an ORG, the connection between the two is shown using a
"METONYMY" pseudo-relation.  When a mention (say, of the city Beijing)
is used to refer to a different entity (China), the mention is marked
with the atttribute METONYMY_MENTION="TRUE".  ("Beijing" then would not
be listed as a name attribute of China.)

There have also been some strictly formatting changes to make the APF
more readable.  Entity types, for example, are now attributes, rather
than elements.  The encoding of charseq elements has also changed, so
that the START and END offset values are now attributes, and the text
of the mention is no longer in an XML comment.