CCGbank -- the machine-readable derivation files (.auto)
	--------------------------------------------------------
     		Julia Hockenmaier and Mark Steedman 

The machine-readable derivation files contain the syntactic derivations in a
format that is designed to be read in automatically. They do not indicate the
word-word dependencies in the predicate-argument structure.
Each sentence appears on one line, preceded by one line which identifies the
sentence: 


ID=wsj0001.1 PARSER=GOLD NUMPARSE=1
(<T S[dcl] 0 2> (<T S[dcl] 1 2> (<T NP 0 2> ...) )) 

The sentence ID  consists of the original Penn Treebank file name, followed by
the number of the sentence in this file.

Each node is indicated by parentheses ``(`` and ``)'' and a
description, which follows the left parenthesis.  The node description
itself is delimited by angled brackets (``<X...>''), where X is L for
leaf nodes and T for other nodes. 


Leaf nodes
----------

In leaf nodes, the description
contains six fields:

<L CCGcat mod_POS-tag orig_POS-tag word PredArgCat>


The original POS tag is the tag assigned to this word in the Penn
Treebank. The modified POS tag might differ from this tag if it was
changed during the translation to CCG. 
PredArgCat is another representation of the lexical category (CCGcat) which encodes
the underlying predicate-argument structure (described in more
detailed below).  

The node description of a leaf node is also enclosed
in parentheses:
(<L N NN NN chairman N>)


PredArgCat, the last field of the node description of leaf nodes, is
designed to encode the word-word dependencies in the underlying
predicate-argument structure. This is a simplified version of the
predicate-argument structure representation presented in the CCGbank
manual.  Complex categories are recursive structures that consist of a
result and an argument category. Each category has a head index, such
that the identity of the lexical heads of two arguments can be
indicated by giving them the same head index. This mechanism
(explained in more detail in the manual) is used to indicate non-local
dependencies that are mediated through lexical items such as relative
pronouns or control verbs.

The head of a lexical category is simply the word itself, and in
complex categories, the head of the result is the same as the head of
the entire category (with the exception of determiners NP[nb]/N, where
the lexical head of the NP is the same as the head of the N.  Since we
only annotate lexical categories with their head indices, the head
index for the entire category is omitted. Similarly, the head index of
any part of a complex category is omitted if it is identical to the
head index of the entire category.


  -- For atomic categories (N, etc.), the head index is
     not indicated, since it is simply the word itself:
	 
     <L N NNP NNP Vinken N> 

  -- In ordinary complex categories that do not mediate any non-local
     dependencies (S[dcl]\NP)/NP etc.), each argument has a
     distinct head index: 
  
     <LL (S[dcl]\NP)/NP VBD VBD took ((S[dcl]\NP_6)/NP_7))>

     If an argument is complex, the head index of each of its parts is
     given:

     <L  PP/(S[ng]\NP)  IN  IN  as  PP/(S[ng]_183\NP_181)_183>

  -- In adjuncts of the form X|X, the head indices of the result
     X are identical to that of the argument X (the head index of the
     entire category, which is different, is omitted):

     <L N/N NNP NNP Pierre N_73/N_73>

     This is also the case if X itself is a complex category, eg.:

     <L (S\NP)\(S\NP) NN NN yesterday (S_9\NP_4)_9\(S_9\NP_4)_9>


     Similarly, in adjuncts that take themselves arguments, eg. (NP\NP)/NP
     or (S\NP)\(S\NP))/NP, the head indices are adjusted accordingly:

     <L (NP\NP)/NP IN IN of (NP_5\NP_5)/NP_6>
     <L ((S\NP)\(S\NP))/NP IN IN by ((S_6\NP_1)_6\(S_6\NP_1)_6)/NP_7>

  -- The mediation of non-local dependencies is indicated by
     co-indexation. As explained in the manual,
     we distinguish between locally mediated (bounded) and long-range
     (undounded) dependencies. 
     Locally mediated dependencies are indicated by :B (bounded):  
  
     <L (S[dcl]\NP)/(S[pt]\NP) VBZ VBZ has (S[dcl]\NP_3)/(S[pt]_4\NP_3:B)_4> 
  
     True long-range depdencies are indicated by :U (unbounded):
  
    <L  (NP\NP)/(S[dcl]/NP)  IN  IN  that  (NP_8\NP_8)/(S[dcl]_9/NP_8:U)_9>


Non-leaf nodes
--------------

For non-leaf nodes, the description contains the following three
fields: the category of the node CCGcat, the index of its head
daughter head (0 = left or only daughter, 1 = right daughter), and the
number of its daughters, dtrs: 

<T CCGcat head dtrs>

With the exception of argument clusters, the head corresponds generally
to the lexical head.  The following example describes a node with
category S[dcl]\NP and two children (0 and 1), the left
of which (child 0) is the head:

<T S[dcl]\NP 0 2> 


Sentences that are not translated are not indicated in this file format.