MERGING - THE ART OF REPRESENTING DIFFERENT LEVELS OY 
SENTENCE STRUCTURE IN A SINGLE ANALYSIS TREE. 
Frank VanEynde 
Departement Linguistlek, Blljde Inkomststraat, 21 
B - 3000 Leuven, Belgium 
!~ An unproblematic example 
The first example illustrates the merging of a part of 
speech analysis with a functional analysis, 
ltl t Part of speech analysis and the on line principle 
A part of speech analysis consists in the assignment 
of category labels to lexicai units. By adding parsing rules 
to the categorial analysis one could also carry out the next 
step of@rouping those lexical units into larger synta~atic 
units, 
The principle governing the part of speech analysis is 
the on llne principle:, syntagmatic units are analysed from 
the left to the right without changing the word order~ 
• Example: (I) Harry promised me a new car 
(' ") Is ~ H~,y~ Cv Pr°misedJ/~o =eli ~T a\] 
\[A new\] C. car? \] 
lt2t Functional analEsie and the dependency principle 
functional analysis consists in the assignment of 
function labels to synta~natlc units. 
The principle governing the functional analysis is the 
dependency principle: every syntaEmatic unit (~ sentence) 
- 95 - 
contains one and only one lexical unit functioning as its 
he ad • 
Example: ( I" ) 
SUJ G0V I0 DO 
GOVI promised J | I 
Harry me a GOV car 
I 
new 
The sister nodes of each GOV-node are 
either terminal nodes, in which case there is no function 
label specified• cf. the node of the indefinite artic- 
le. 
or non-terminal nodes, in which case the function label 
takes one of the following values: SUJ, DO, IO, ~F 
(= modifier), ... 
The function labels specify the kind of relation holding bet- 
ween the head and its dependents• Por instance, "Harry" is the 
subject of "promised", "new" is a modifier of "oar", etc. 
1,3, A synthetic representation 
Assumption: the part of speech analysis and the functional 
analysis can be represented in the same tree, sin- 
ce the on llne principle and the dependency prin- 
ciple are compatible. . 
Example: (I) 
I I 1 i  ,.ovJ I 
C a.l~" Harry me a 
l 
new 
- 96 - 
2 t A problematic example: the result clause 
I'll be co~ erned with the external structure of the 
result clause only, not with its internal structure; the 
result clause will, consequently, be treated as an unanalyeed 
synt~gmatic unit. 
2,1, A part of speech anal2sis of the result clause 
Example: (2) He left so early on Tuesday that we missed him 
(2°) \[4RO heJ IV left'lAdy so~ \[Adv early~ \[p on~ 
4  uesdaY3 rs that we  ssed 
We could add some further structure to this bracketing 
by subsuming the adjacent adverbs under one node: 
~\[Adv so~ lAdy early33 
We can, however, not incorporate the "that"-clause into 
this eyntagmatic unit, since the prepositional phrase "on 
Tuesday" intervenes between both parts. 
2,2~ A functional anal2sis of the result clause 
The ~rucial question concerning the functional analysis 
of a sentence llke (2) is: which constituent governs the 
"that"-clause? In other words, where do ~e have to attach 
the S-node of the result clause? 
Pot reasons (to be given in the full version of this 
paper) we propose the following dependence structure for (2): 
(2") 
SUJ GOV MP 
GOV left ~IP GOV on Tuesday 
l /'~- L he GOV 
J N~ early 
so that we missed him 
- 97- 
2.3. A synthetic representation 
2.3.1. Unfortunately, the part of speech analysis and 
the functional analysis do not match (cf. the disturbance of 
the word order in (2"). 
A marging of the labeled bracketing with the dependency 
structure would give the following result: 
(II) r~,-~\] 
@Ro~,~ EA,GOVJ on Tuesday tiat.., him 
he ~A,GOg \[-,MFJ early 
l so 
The lowest MF-node dominates no lexlcal material, and has, 
consequently, no category label. The righmost S-node, on the 
other hand, has no functional label, since it does not bear 
any dependency relation to the verb. 
Important to note is the fact that both deficient nodes 
are complementary, and that they in fact refer to each other. 
In order to make thls relation explicit, I propose to add an 
index (an arbitrary integer) to both nodes: 
(II) ,.. ~-,MPJ1 ...~S,-~i 
Thanks to this coindexlng device we are able to merge two le- 
vels of sentence structure, although they do not seem to be 
compatible at first sight. 
2.3.2. The computation of (II) can be performed in a 
straightforward way: 
... so ...  A,®V so-J "JJ'''J 
2. If there is a that-clause in S', then give it a ~S,-Ji-no- 
de, and attach it immediately under the S"-node. 
If there is no such clause, then delete the L-,MPJi-node. 
- 98 - 
Comparison with a transformational treatment. 
2.3.3. Similar analyses can. be given for all kinds of 
comparison clauses. 
~t A third example: deep and surface sub teots 
~tl~ On the notion "surface sub,~ect" 
3.2. On the notion "deep subject" 
3.3. A synthetic representation 
In a sentence like 
(3) it seems that Steve likes her 
it could be argued that the surface subject and the deep sub- 
ject do not coincide. Making use of some new notations (to be 
explained in the full version of this paper) and of the device 
already known from section 2.3.1. I propose the following ana- 
lysis tree for (3): 
(III) \[S,-,-3 
1 
! i .-...'-"'- . 
that likes her ~PRO, GOV\] seems Steve 
Merging is a technique of representing different levels 
of syntactic (and semantic) structure in one analysis tree. 
In order to make merging work one has to make sure that: 
1. each level of analysis is properly defined, i.e. 
- that there exists a list of possible values for the labels 
- that there is an algorithm for assigning those values in 
each particular case 
- that there is a unifying principle at each level (for 
instance, the dependency principle at the level of funct- 
ional analysis). 
2. The relations between the different levels are properly 
defined. In order to ~-,~rant the latter I have pleaded for 
- 99 - 
adding referenoe indices to the nodes, thus introducing a new 
formal device in the ~ra~mar 
a third dimension in the an~ysis tree. 
Some merits of the merging technique in a progrsm .for auto- 
matic translation. 
- 100 - 
