next up previous contents
Next: Metarules Up: Appendices Previous: Thematic Roles

   
Tree Families in XTAG

In the XTAG English grammar, tree families provide us with a compact means to specify the set of trees which a particular lexical anchor takes. This set represents all syntactic environments in which this anchor appears. One way to represent tree families is to have a universal set of trees, with each individual anchor taking some subset of that larger set. However, since syntactic transformations are not usually sensitive to the anchor of the tree family, in practice we assign a tree family to entire classes of words. Chapter 4 defines the notion of tree family used in the XTAG grammar. In this chapter we discuss in more detail what it means for a particular anchor to select several tree families keeping in mind that entire classes of words which share the same subcategorization are assigned tree families in the syntactic database. Also, there are some syntactic transformations that are sensitive to the properties of a particular word within the same subcategorization frame. In general, the trees within a particular tree family can be thought of as being those which are syntactically or transformationally related. The tree in Figure B.1 allows us to parse sentences such as ((561)). This tree represents the subcategorization of the verb and forms the base for the generation of an entire tree family. Each family has one such base tree. One method for generating all the trees in the family that are syntactically related to the base tree is to use metarules (see Appendix C). One such member of the transitive tree family is the tree in Figure B.2 which is the tree where the object of the nx0Vnx1 has been extracted. This new tree W1nx0Vnx1 now allows us to parse sentences such as ((562)).
(560)0(560
(561)
Susie likes Hobbes.  (561)0(561
(562)
Who does Susie like? 

ps/verb-class-files/alphanx0Vnx1.ps.gif

{Declarative Transitive Tree: $\alpha$nx0Vnx1

 

ps/extraction-files/alphaW1nx0Vnx1.ps.gif

{Transitive tree with object extraction: $\alpha$W1nx0Vnx1

 

The semantic interpretation of arguments across different trees in a family remains constant. That is, if is the recipient argument of the predicate in the base indicative tree, then it is the recipient when it is fronted as well. There are some syntactic transformations, however, that are sensitive to the properties of a particular word within the same subcategorization frame. The ergative (or transitive-inchoative alternation) for transitive verbs is one such transformation. Only a subset of the transitive verbs can undergo this transformation. Let us call such transformations lexical rules to distinguish them from more general syntactic transformations (like wh-extraction) which are not lexically idiosyncratic. To explain how these two different kinds of transformations interact let us take the example of the ergative subset of the transitive verbs: Ergative verbs such as melt undergo the ergative transformation (see Chapter 7). Such verbs are a subset of the set of transitive verbs, which do not all take this transformation; for example verbs like borrow cannot take this transformation. To handle the transitive/ergative distinction, we could create a tree family which exclusively handles the regular transitives and another one which handles the ergatives. In doing so, however, we would be duplicating structure since the ergative family would contain all of the trees found in the transitive family in addition to the purely ergative trees. An alternative is to consider the ergative distinction as arising from a lexical rule which takes the base tree of the transitive family (nx0Vnx1) and produces the base tree for a strictly ergative family (Enx1V). Then all of the syntactic transformations which are available in the grammar can be applied to that new base tree to form the entire ergative tree family. In this particular instance, the indexation on the syntactic object NP of nx0Vnx1 ( ) is retained on the syntactic subject of the ergative tree, which has the same semantic role in both constructions. The same problem appears in the ditransitive tree family, in which a subset of anchors also undergo the dative-shift transformation. We give an identical solution to this case as we do for the ergatives. Hence, tree families can be related to each other by lexical rules. Construing tree families in this way allows us to take advantage of shared structure when creating a grammar. Introducing this shallow hierarchy in the definition of family allows us to have a more compact organization of the grammar. This departs from earlier conceptions of tree families which maintained that argument positions in different tree families were in no way related; thus, before, if a lexical item anchored two different tree families, each anchoring instance was considered a different predicate. Under both the previous and the present conceptions, tree families are the exhausitive domain of a base tree and the trees generated by syntactic transformations on that base. However, the present conception of tree families provides for a new level of distinction - that of families related by a lexical rule. These related families are the exhaustive domain of a particular predicate that is affected by a lexical rule. The XTAG grammar encodes the notion of lexical rules implicitly rather than explicitly. It is implicit in the anchorings of certain lexical items. For example, ergative verbs such as melt anchor both Tnx0Vnx1 and TEnx1V, and their lexical family is the set of trees in the union of these two tree families. See the chapters on ergatives (see Chapter 7) and ditransitives (see Chapter 10) for more in-depth discussions of particular cases of lexical families. To briefly summarize, we can now give the following definitions:

tree family:
the set of trees generated from a base tree by the syntactic transformations which are available in the grammar.
lexical family:
the set of tree families generated for a given predicate from an initial base tree and the base trees which are generated by the lexical rules which are specified for that predicate.
It should be noted, however, that we will continue to refer to tree families and lexical families both by the generic name tree family, unless the distinction is crucial to the point we are trying to make.
next up previous contents
Next: Metarules Up: Appendices Previous: Thematic Roles
XTAG Project
http://www.cis.upenn.edu/~xtag