A Frame-Based Semantics of Locative
Alternation in LTAG
Yulia Zinova
Heinrich Heine University Düsseldorf
zinova@phil.uni-duesseldorf.de
Abstract. In this paper I present the analysis of locative alternation
phenomena in Russian and English. This analysis follows the approach
proposed in [13] and uses LTAG and Frame Semantics. The combination
of a syntactic theory with an extended domain of locality and frames
provides a powerful mechanism for argument linking. Metagrammar fac-
torization allows to determine not only lexical, but also constructional
meaning that is essential for locative alternation analysis.
1 Introduction
There is a number of formalisms that capture the idea that the meaning of a verb-based
construction depends both on the lexical meaning of the verb and on the construction
in which the verb is used ([8], [19]). The question is how exactly the components of the
meaning are distributed and how they combine.
In [13] a combination of Lexicalized Tree Adjoining Grammars ([9]) and Frame Seman-
tics is introduced. It is shown that the resulting framework is very flexible with respect
to the factorisation and combination of lexical and constructional units on the syntax
and semantics level and is also suitable for computational processing.
Though there already exist a number of different approaches to semantic construction
using LTAG ([10], [5], [12]) and approaches that combine other syntactic formalisms
with Frame Semantics ([3], [4]), the novel combination of an LTAG and Frame Seman-
tics benefits from both extended domain of locality and underspecification allowed by
frames.
In this paper I want to present the analysis of locative alternation that benefits of
flexibility offered by the novel framework.
2 Tree Adjoining Grammar
Tree Adjoining Grammar (TAG, [9]) is a tree-rewriting grammar formalism. It consists
of a finite set of elementary trees with labelled nodes with two operations on them:
substitution and adjunction. All elementary trees are either auxiliary trees or initial
trees. An auxiliary tree is a tree which has exactly one foot node - a leaf that is marked
with an asterisk. Leaf nodes can be labelled with terminals and other nodes are labelled
only with non-terminals. The derivation process starts from an initial tree and in the
final derived tree all the leaves must be labelled by terminals.
Substitution allows to replace a non-terminal leaf with a new tree and adjunction is
used for replacing an internal node with an auxiliary tree. Adjunction to the node
R.K. Rendsvig and S. Katrenko (Eds.): ESSLLI 2012 Student Session Proceedings, CEUR Work-
shop Proceedings vol.: see http://ceur-ws.org/, 2012, pp. 202–210.
A Frame-Based Semantics of Locative Alternation in LTAG 203
labelled X is allowed if the root and foot nodes of the adjoining auxiliary tree have the
same label X. It is also possible to indicate nodes where adjunction is obligatory or not
allowed with OA and NA subscripts respectively.
Figure 1 shows an example of a derivation: the initial tree for Mary substitutes into
the subject slot of the elementary tree for laughs, and the sometimes auxiliary tree for
the VP modifier adjoins to the VP node. Feature-structure based TAG, or FTAG, is a
derived tree:
S
S
NP VP
NP NP VP VP Mary Adv VP
Mary V Adv VP* sometimes V
laughs sometimes laughs
Fig. 1. Example of a TAG derivation
variant of TAG in which elementary trees are enriched with feature structures ([20]).
In an FTAG each node has a top feature structure and all the nodes except substi-
tution nodes have a bottom feature structure. Feature unification happens during the
derivation process when adjunction and substitution take place. Due to the extended
domain of locality, nodes within one elementary tree can share features, allowing to
express constraints among dependent nodes easily.
For natural languages a specific version of TAG called lexicalized TAG, or LTAG is used.
In an LTAG, each elementary tree must have at least one non-empty lexical item, called
lexical anchor. The second important principle for a natural language TAG is that ev-
ery elementary tree where the lexical anchor is a predicate must contain slots (leaves
with non-terminal labels) for all arguments of this predicate, including the subject, and
for nothing else (theta-criterion for TAG, [6]).
The facts that LTAGs have extended domains of locality and that elementary trees
are lexicalized and contain slots for all the predicate’s arguments, make them good
candidates for combination with frame-based compositional semantics ([13]). In the
approach proposed in [13], a single semantic representation (a semantic frame in this
case) is linked to the entire elementary tree. When coupling an elementary tree with
a semantic frame, syntactic arguments can be directly linked to their counterpart in
the semantics. Described approach is similar to ones in [7] and [14], but uses different
kind of semantic representation. Semantic composition is then modeled by unification
which is a result of performing adjunction and substitution. Figure 2 provides a sim-
ple illustration of syntactic and semantic composition. In this example, substitutions
trigger unifications between 1 and 3 and 2 and 4 which leads to correct insertion of
argument frames into the frame of loves.
Linguistic generalizations in TAGs are captured by a metagrammar. There are two
steps of factorization, which are important for this paper:
• unanchored elementary trees are specified separately from lexical anchors;
204 Yulia Zinova
S
NP NP[I= 1 ] VP NP
John V NP[I= 2 ] Mary
" # loves " #
person love person
3 4
NAME John EXPERIENCER 1
NAME Mary
THEME 2
Fig. 2. Syntactic and semantic composition for John loves Mary
• trees are organized into tree families which represent different realizations of one
subcategorization frame.
This allows to define a meaning for sets of unanchored elementary trees, i.e., a meaning
of constructions.
3 The Data
3.1 Previous approaches
(1) and (2) show basic examples of locative alternation in English and Russian. Despite
the fact that in English both constructions have a PP and it can be omitted without
losing the specific construction meaning, let us call the first variant ((1a), (2a)) prepo-
sitional phrase construction, or PPC and the second variant ((1b), (2b)) - instrumental
case construction, or ICC for convenience of referring to them.
(1) a. John 1 loaded the hay 2 into the truck 3 . (PPC)
b. John 1 loaded the truck 3 with hay 2 . (ICC)
(2) a. Ivan 1 zagruzil seno 2 v vagon 3 .
Ivan loaded hayacc,def in wagongen,indef /def .
Ivan loaded the hay into a/the wagon.
b. Ivan 1 zagruzil vagon 3 senom 2 .
Ivan loaded wagonacc,def hayinstr,indef .
Ivan loaded the wagon with hay.
PPCs are traditionally analyzed as having a change of location meaning and ICCs - as
having a change of state meaning ([11], [16], [8]). An analysis for (1) following [11] is
provided in (3). It demonstrates that there is a difference between the two constructions,
but only the difference in the perspective is shown.
(3) a. X CAUSE [BECOME [hay BE ON truck]]
b. X CAUSE [BECOME [truck z BE [WITH [hay BE ON z]]]]
The analysis proposed in [16], which can be found under (4), provides more detailed
information about the difference between PPCs and ICCs. (4a) tells us that the hay
changes its location as a result of the loading event, while (4b) describes that the result
is a change in the state of the wagon. One can notice that in (3) there is no explicit
A Frame-Based Semantics of Locative Alternation in LTAG 205
reference to the verb itself and the only component that is taken from the verb meaning
is that the result of the loading is that the THEME is on the LOCATION in the end.
(4) a. [[x ACT] CAUSE [y BECOME Ploc z] [LOAD]M AN N ER ]
b. [[x ACT] CAUSE [z BECOME []ST AT E WITH-RESPECT-TO y] [LOAD]M AN N ER ]
Using frame semantics, one can assign two frames in 3 to the two different construc-
tions. For the PPC, one has to remove the concrete verb load and replace it with
change of location effect. So the first frame tells us that the activity of the Actor (X)
causes the Theme (Y) to change its location to the Goal (Z). For the ICC’s frame in
order to introduce the Manner one can simply embed the caused change of location
frame under the manner atribute. The second frame in this case would mean that
the activity of the Actor (X) causes the Theme (Z) to change its state by means of
changing the location of the third argument (Y) to Z.
causation
" #
activity
4
cause
actor 1
change of state
theme 3
causation
" # 0 causation
activity
" #
activity
cause 4
actor 1 effect 5
cause 7
0
actor 1
manner 6
change-of-loc
change of loc
5 theme 2
effect
8 theme 2
effect
goal 3 goal 3
Fig. 3. Frame semantics for PPC and ICC
3.2 Detailed Russian and English Data
The question that arises if one looks carefully at what the sentences in (1) and (2)
mean is whether it is really the case that there is no change of state in PPC examples?
In fact, any loading activity leads to both a change of location of the content and
some change of state of the container (if it is specified), and the difference between two
constructions is that
• different components of the effect become more salient;
• in the case of ICC initial and result states of the container are specified.
In order to understand how the meaning of verbs and constructions should be repre-
sented let us look at the whole range of the verbs allowing locative alternation that one
can find in English and Russian. [18] provides the following classification for English:
Content-oriented classes:
206 Yulia Zinova
(a) simultaneous forceful contact and motion of a mass against a surface (brush, drape,
spread, etc.);
(b) vertical arrangement on a horizontal surface (heap, pile, stack);
(c) force is imparted to a mass, causing ballistic motion in a specified spatial distribu-
tion along a trajectory (inject, splash, spray, etc.);
(d) mass is caused to move in a widespread or non directed distribution (scatter, seed,
sow, etc.).
Container-oriented classes:
(e) a mass is forced into a container against the limit of its capacity (crowd, jam, stuff,
etc.);
(f) a mass of size, shape, or type defined by the intended use of a contained is put into
the container, enabling it to accomplish its function (load, pack, stock).
From the description of verb classes that allow locative alternation in English one can
see that the result state of the container in case of ICC is such that the action cannot be
performed any longer. There is no result state common for all the cases, so it depends
on the verb, i.e. on how the change of location happens. The easiest way to solve this
would be to assume different construction meanings for different verb classes (e.g. one
with the Effect of the Theme being full and the other one with the Effect of the Theme
being covered), but let us first look at some Russian data.
In Russian a lot of verbs allow only one of the constructions, i.e. a change of construction
requires a change of verb prefix (a list can be found in [1]). However, some of the verbs
from the list remain the same in both prepositional and instrumental constructions.
Such verbs can be organised in three groups: the first one is similar to the (f) group
in English (see example (2)), the second one is similar to group (a) in English, like in
(5), and the third class is like a combination of the first and the second: a mass is put
into a container, enabling it to accomplish its function, or on a container, covering its
surface (6).
With the vebs from the third group an interesting effect can be observed: while in the
case of PPC example (6a) there is a preposition which tells us that the content goes
in the container, in the case of ICC example (6b) two different readings are possible:
the content can be put in the container or the content can cover the container. In both
cases there is a clear result state: either the container is full or the container’s surface
is fully covered with content. This means that the verb zasypat’ (’to fill/to cover’) does
not provide information about how the theme is positioned at the goal. In case of
PPC this information comes from the preposition used (both v (’in’) and na (’on’)
are possible) and in ICC the ambiguity can be resolved only using world knowledge.
So (6) demonstrates conclusively that there should be one construction accounting for
different result states of the theme and allowing to get different interpretations of one
verb due to underspecification of how the change of location process goes.
(5) a. On namazal maslo na hleb.
He distributed butteracc,def on breadacc,indef
He distributed butter over a piece of bread.
b. On namazal hleb maslom.
He covered breadacc,def butterinstr,indef
He covered a piece of bread with butter.
(6) a. On zasypal sahar v banku.
He put suggaracc,def in canacc,indef /def
A Frame-Based Semantics of Locative Alternation in LTAG 207
He put sugar in a/the tin.
b. On zasypal banku saharom.
He covered/filled tinacc,def sugarinstr,indef
He covered/filled the tin with sugar.
4 Locative Alternation: The Analysis
4.1 Syntactic representation
In the previous section we were looking only at ”full” examples, where both container
and content are present. However, the constructions that are being discussed can be
used when only the direct object of the verb is present; in this case, they will have the
same difference in semantics. Using LTAG and metagrammar decomposition one can
obtain the tree family in 4 for the PPC and tree family in 5 for the ICC (the second
NPIN ST R stands for both NP in instrumental case in Russian and PP with preposition
”with” in English).
S
S
NP[I= 1 ] VP
NP[I= 1 ] VP
V⋄[S= 0 ] NP[I= 2 ] PP[I= 3 ]
V⋄[S= 0 ] NP[I= 3 ] NPIN ST R [I= 2 ]
S S
NP[I= 1 ] VP NP[I= 1 ] VP
V⋄[S= 0 ] NP[I= 2 ] V⋄[S= 0 ] NP[I= 3 ]
Fig. 4. Unanchored trees for the PPC Fig. 5. Unanchored trees for the ICC
4.2 Scales and Proposed Frame Semantics
In the case of the PPC, the semantics of the whole phrase can be compositionaly de-
rived from the semantics of the verb and its arguments, while in the case of the ICC
there is a part of the meaning, that comes from the construction itself. The goal now
is to provide such semantics for the ICC and verbs allowing locative alternation such
that in combination they form the desired frame representation of the semantics of a
sentence.
Following ideas in [17] where one can find a discussion of the representation of at-
tributes, events, and results while implementing Fillmore’s Frame Semantics ([2]) I
introduce attributes of initial and result state and a scale which is determined by its
type, start and end points. The change of state is either a decrease or an increase of the
value on an ordered scale (discussion of analysis of scalar change can be found in [15]).
The direction is given by the values of atributes ENDP and STARTP (end and start
208 Yulia Zinova
points), which replaces the LESSER atribute of ordering proposed in [17]. Some of
the verbs specify a concrete initial or result state (INIT and RESULT respectively),
but load does not have any initial or result state specified within its semantics, so it
just determines the scale and two values on it. Summarizing the ideas, one obtains the
following scheme:
• in the verb both change of location and change of state effects are specified;
• MANNER attribute is not needed because it is already described in the change of
location subframe;
• change of state is described by scale, initial state, and result state;
• SCALE atribute can have a type such as ”degree of fullness” that is a subtype of
the type ”scale” and thus replaces it during the unification;
• initial and result states are values on the scale;
• the ICC specifies that initial and result states are equal to the start and end points
of the scale, respectively.
causation
scalar change of state
theme 3
causation
actor 1
scale
theme 2
goal 3
7 startp 5
scale
" #
activity
cause
endp 6
actor 1
scalar change of state
state
0
theme 3
effect
init entity 3
degree of fullness
7
scale 0
scale 7 startp empty
endp full
value 5
change of location state
state
theme 2
effect , init
entity 3
3
goal
scale
7
entity 3
8
value
result
7
scale
state
3
value 6
entity
result
scale 7
value 9
Fig. 6. Frame for the ICC Fig. 7. Frame representation of the verb load
Figure 7 shows a lexical frame for the verb load. As one can see, when all the
arguments are filled, the right meaning for the whole PPC follows automatically. The
semantics of the ICC is a caused change of state meaning that gets further constrained
when a specific lexical anchor is inserted. Figure 6 shows how the unanchored tree
for the ICC is linked to its semantic frame. The correct argument linking happens
because I features in the syntactic tree and the thematic roles in the semantic frame
are identical. This is done in a local way (within the domain of an elementary tree)
because of LTAG’s extanded domain of locality. The S feature of the V node serves
for unification of the lexical frame for the verb and the constructional frame. When
a lexical anchor is inserted, this feature unifies with the S feature of the lexical item.
The result of this unification for the ICC with a lexical anchor load is shown in Fig. 8.
A Frame-Based Semantics of Locative Alternation in LTAG 209
causation
actor 1
theme 2
goal 3
" #
cause activity
1
actor
scalar change of state
3
theme
degree of fullness
0
scale 7 startp empty
endp full
change of location
state
effect theme 2 , entity 3
init
3
goal
scale
7
value empty
state
entity 3
result
scale 7
value full
Fig. 8. Resulting frame for the ICC and the verb load
5 Conclusion
This is a case study for combining an LTAG with Frame Semantics, in which I have
described a model for locative alternation in English and Russian. This analysis uses
LTAG’s mechanism of separation between unanchored elementary trees and lexical
anchors to separate the contribution of the lexical meaning from the contribution of
construction and follows the ideas expressed in [13]. Another advantage of combination
of an LTAG with Frame Semantics is that LTAG’s extended domain of locality allows
direct linking of thematic roles of the arguments with corresponding syntatctic slots.
As this framework is a new one, there are a lot of open questions and a wider range of
semantic phenomena should be examined.
References
1. Dudchuk, P.: Locativnaja alternacija i struktura glagol’noj gruppy [Locative Alter-
nation and the structure of VP]. diploma thesis, Moscow State University (2006)
2. Fillmore, C.J.: Frame semantics. In: of Korea, T.L.S. (ed.) Linguistics in the Morn-
ing Calm, pp. 111–137. Hanshin Publishing Co., Seoul (1982)
3. Frank, A.: Generalizations over corpus-induced frame assignment rules. In: Pro-
ceedings of the LREC 2004 Workshop on (2004)
4. Frank, A.: Question answering from structured knowledge sources. Journal of Ap-
plied Logic 5(1), 20 – 48 (2007)
210 Yulia Zinova
5. Frank, A., van Genabith, J.: GlueTag. Linear logic based semantics for LTAG –
and what it teaches us about LFG and LTAG. In: Butt, M., King, T.H. (eds.)
Proceedings of the LFG01 Conference. Hong Kong (2001)
6. Frank, R.: Syntactic Locality and Tree Adjoining Grammar: Grammatical, Acqui-
sition and Processing Perspectives. Ph.D. thesis, University of Pennsylvania (1992)
7. Gardent, C., Kallmeyer, L.: Semantic Construction in FTAG. In: Proceedings of
EACL 2003. pp. 123–130. Budapest (2003)
8. Goldberg, A.E.: Constructions. A Construction Grammar Approach to Argument
Structure. Cognitive Theory of Language and Culture, The University of Chicago
Press, Chicago and London (1995)
9. Joshi, A.K., Schabes, Y.: Tree-Adjoning Grammars. In: Rozenberg, G., Salomaa,
A. (eds.) Handbook of Formal Languages, pp. 69–123. Springer, Berlin (1997)
10. Joshi, A.K., Vijay-Shanker, K.: Compositional semantics with lexicalized Tree-
Adjoining Grammar (LTAG): How much underspecification is necessary? In:
Blunt, H.C., Thijsse, E.G.C. (eds.) Proceedings of the Third International Work-
shop on Computational Semantics (IWCS-3). pp. 131–145. Tilburg (1999)
11. Kageyama, T.: Denominal verbs and relative salience in lexical conceptual struc-
ture. In: Kageyama, T. (ed.) Verb semantics and syntactic structure, pp. 45–96.
Kurioso Publishers, Tokyo (1997)
12. Kallmeyer, L., Joshi, A.K.: Factoring Predicate Argument and Scope Semantics:
Underspecified Semantics with LTAG. Research on Language and Computation
1(1–2), 3–58 (2003)
13. Kallmeyer, L., Osswald, R.: A frame-based semantics of the dative alternation
in lexicalized tree adjoining grammars (2012), submitted to Empirical Issues in
Syntax and Semantics 9
14. Kallmeyer, L., Romero, M.: Scope and situation binding in LTAG using semantic
unification. Research on Language and Computation 6(1), 3–52 (2008)
15. Kennedy, C., Levin, B.: Measure of change: The adjectival core of degree achieve-
ments. In: McNally, L., Kennedy, C. (eds.) Adjectives and adverbs. syntax, seman-
tics, and discourse. Oxford University Press, Oxford (2008)
16. Levin, B., Rappaport Hovav, M.: Morphology and lexical semantics. In: Spencer,
A., Zwicky, A.M. (eds.) Handbook of morphology, pp. 248–271. Blackwell Publish-
ers, Oxford (1998)
17. Osswald, R., Van Valin, Jr., R.: Framenet, frame structure, and the
syntax-semantics interface (2012), http://www.phil-fak.uni-duesseldorf.
de/fileadmin/Redaktion/Institute/Allgemeine_Sprachwissenschaft/
Dokumente/Rainer_Osswald/FrameStructure-Osswald-VanValin-2012-02-20.
pdf, manuskript Heinrich-Heine Universität
18. Pinker, S.: Learnability and cognition: The aquisition of argument structure. MIT
Press, Cambridge, MA (1989)
19. Van Valin, R.D., LaPolla, R.J.: Syntax. Cambridge University Press, Cambridge
(1997)
20. Vijay-Shanker, K., Joshi, A.K.: Feature structures based tree adjoining grammar.
In: Proceedings of COLING. pp. 714–719. Budapest (1988)