[XTAG summary] tree families

Mon Dec 14 21:32:17 1998
[This discussion is rather long, although we have tried to make it as
short as possible. We need your comments on it, so please try and read
it before the next meeting.]

During the XTAG meeting on December 10th, we discussed the definition
of tree families and whether we should modify the notion which was
held until now.  The discussion focussed on the ergatives and the
ditransitives and their roles in the organization of the XTAG grammar
and the notion of tree family.  Anoop, Tonia, and Jason met the
following day and tried to take the two opposing viewpoints and see if
either could be extended to make a consistent description of the
entire XTAG grammar. In the course of this experiment some issues came
up which appear to clarify the issues that we the group had wavered on
during the meeting on the 10th.

Formerly, tree families carried with them the implication that they were the
exhaustive instance of the kinds of transformation which a particular syntactic
frame could undergo. One method of defining a tree family is through the
following process: given a base tree selected by an anchor, the entire family
is generated by metarules which represent syntactic
transformations. Importantly, the anchor now selects a set of trees (a family)
and two trees in separate families cannot be related to each other

However, this can lead to certain redundancies if we want some classes
of words to undergo transformations which are a superset of another
class.  An example of this is the ergatives---we wish verbs like
"sink" to have all the transitive trees e.g. "John sank the boat",
plus the inchoative/ergative alternation e.g. "The boat sank". However, we
do not want verbs like "borrow" to take this alternation.

A way of handling this is to have two tree families, one which has
only transitive trees, and another which has transitive and ergative
trees.  Then "borrow" will anchor the former and "sink" will anchor
the latter.  But then we have duplicated the transitive trees, which
means more maintenance, longer loading times, etc. than if we could
have provided a more compact representation.  We will call this the
full-specification solution (FULL).

The previous solution to allow a compact representation was to put the
transitive trees and the ergative trees in one family, Tnx0Vnx1, and use a
feature (trans) to partition the lexical items so that ergative verbs could
take all the trees with no difficulty while pure transitives would have a
feature conflict when they anchored ergative trees in the family.  We will call
this the feature solution (FEAT).

Another compact solution is to create a separate tree family which has
only the ergative trees in it.  Purely transitive verbs then anchor
the transitive-only family, while ergative verbs anchor the transitive
family and the ergative family.  We will call this the union solution

Both of these compact solutions (FEAT and UNION) allow the sharing of syntactic
transformations, but they also both break the definition of tree family given
in the second paragraph.

* FEAT breaks the definition because after the metarules apply to give
  us a set of trees, the anchor no longer selects all the trees in the
  family but only a subset of them. 

* UNION breaks the definition because the ergatives are no longer
  transformationally derived from the transitive base tree by a

The case of the ergatives is not the only one where such a solution is
required. Naturally, if we want the grammar to be consistent, all such cases
should be handled in an identical manner. This is what we set out to do: to see
if other cases similar to the ergative/transitive problem can help us decide
between the two alternatives. The ditransitives are such a case. 

The handling of ditransitives until now was done in a
full-specification manner (FULL).  The ditransitives with PP shift family,
Tnx0Vnx1tonx2, was a superset of the ditransitive family, Tnx0Vnx1nx2.
Let us test out both solutions FEAT and UNION on this case.

Extending FEAT to this case: use a partitioning feature (call it
"shift"), then let verbs which exhibit the dative ditransitive
alternation, e.g. "give", be shift+ and those which do not,
e.g. "bake", to be shift-, and finally require that a verb be shift+
in order to take the PP shifted trees like alpha_nx0Vnx1tonx2 (John
gave the book to Mary).

However, this does not handle all the duplication in the grammar. The PP
shifted trees share syntactic transformations with the Tnx0Vnx1Pnx2 family
("ascribe to"). Using FEAT, they would also become part of the ditransitive
family using the shift feature to deselect all the distransitive trees except
the PP shifted ones. In order to implement this in the XTAG system, the P will
now be a substitution node rather than an anchor (since there are other
prepositions which can occur in these trees). By making P a substitution node,
different prepositions will have to be selected using features, i.e. "give"
would select for "to" using a feature. So using FEAT carries with it
implications not just for grammar organization but representation as well.

Another case will be the benefactives (John baked Mary a cake). They do in fact
undergo a transformation similar to the dative shift (John baked a cake for
Mary), however in this case the for-PP is adjoined on (so "bake" just takes an
ordinary transitive tree). In order treat this as an alternation, we could
split the benefactives off from the ditransitive-with-dative-PP-shift family
and we are once again in a situation where we need to replicate
structure. Alternatively, we need one big tree family for the transitives,
ditranstives, and PP shift trees, and the notion of tree family dissolves since
there are no tree family divisions.

Let us use UNION for the same case. We have one family for the
ditransitives such as "John asked Mary a question" and another one for
"John gave a book to Mary."  Then non-shifting ditransitives ("ask") will
anchor only the former while those which do shift ("give") will anchor both.
We then let the tree family for dative shifting ditransitives be the
union of these two families.

However, in doing this we have destroyed the relationship which is
captured in the single family approach----namely, that NP1 and NP2
remain consistent in terms of the arguments they represent throughout
the entire family.  The ditransitive family is called Tnx0Vnx1nx2;
thus, in "John gave Mary the book", `Mary' fills the NP1 position.
The family we use for the shifted version is Tnx0Vnx1Pnx2, and for the
sentence "John gave the book to Mary", `Mary' is now NP2.  Generally,
the index is maintained in transformational movements in XTAG
(e.g. passive = nx1V) to represent that a particular argument is in a
different syntactic position yet retains the same semantic role, which
is why we have the trees alpha_nx0Vnx1tonx2 and alpha_nx0Vnx2nx1 in
the original family Tnx0Vnx1tonx2. We do not appear to have recourse
to a similar strategy when we move our notion of tree families from
individual families to sets of families.  Or do we?

It actually makes more sense to swap the indices of the ditransitive
trees, giving us Tnx0Vnx2nx1 instead of Tnx0Vnx1nx2.  Then `Mary' will
fill the NP2 position for the trees which parse the sentences "John
gave Mary a book" (alpha_nx0Vnx2nx1) and "John gave a book to Mary"
(alpha_nx0Vnx1Pnx2), and we do not need to have them in the same
"physical" family to enforce this argument coherence.  This can then
also be said to hold for the ergatives, for which the subject is NP1,
representing the fact that the syntactic subject of an ergative verb
is the logical object.  This maintains coherence across the Tnx0Vnx1
and TEnx1V families, the union of which defines the "ergative family".

Taken in this spirit, UNION does not require us to change our analysis
of these phenomena as related by syntactic transformations or
metarules. In the original definition, a set of metarules were applied
on a single base tree. The UNION solution suggests that there are some
"lexical rules" that are first used on a base tree to give us a set of
base trees (each forming the basis of a tree family) and then the
metarules apply as usual. This notion of "lexical rule" is exactly
what was proposed in the hierarchical lexical organization of the
grammar (Lexorg). In terms of XTAG nomenclature, it would mean that
tree families can be unions of other tree families.

Also, UNION allows us to make an elegant claim about object
drop. "eat" would allow object drop and thus its family is the union
of Tnx0V and Tnx0Vnx1. This contrasts with the ergatives since the
syntactic subject of Tnx0V is NP0, which corresponds to the logical

This we feel is a fair comparision of the two alternatives. If, in
fact, we have omitted something relevant please let us know as soon as
possible. As the experiment stands so far, UNION gives us a better
solution than FEAT. If you have a counterexample, please speak up.