PROCESSING WORD ORDER VARIATION WITHIN A MODIFIED ID/LP FRAMEWORK 
Pradip Dey 
University of Alabama at Birmingham 
Birmingham~ AL 35294~ U.S.A. 
ABSTRACT 
From a well represented sample of world languages 
Steele (1978) shows that about 70% of languages exhibit 
significant word order variation. Only recently has this 
wlde-spread phenomenon been drawing appropriate attention. 
Perhaps ID/LP (Immediate Dominance and Linear 
Precedence) framework is the most debated theories in this 
area. We point out some difficulties in processing standard 
ID/LP grammar and present a modified version of the 
grammar. In the modified version, the right hand side of 
phrase structure rules is treated as a set or partially ordered 
set. An instance of the framework is implemented. 
1. Introduction 
From a well represented sample of world languages 
Steele (1978) ~,~hows that about 76~ of the languages exhibit 
significant word order variation | . Until recently thls 
widespread phenomenon was not given proper attention in 
natural language processing. The primary goal of tbis study 
is to develop eomputatlonally efficient and linguistically 
adequate strategles for parsing word order variation. The 
strategies are implemented in a network based parser. At 
first we characterize the basic problem at an abstract level 
without going into details of the problem in any specific 
language (in Section-2). Then~ in Sectlon-3, the details of the 
problems in a specific language~ namely~ Iilndi~ are presented. 
The Immediate dominance and linear precedence (ID/LP) 
framework~ developed by Gazdar and Pullum~ is one of the 
most debated theories in the study of word order variation 
(Pullum 1982~ Uszkorelt 1082, Shieber 1983, Barton 1985). 
The basic idea behind ID/LP framework is to separate 
immediate dominance from linear precedence in rewrite rules. 
Pullum (1982) expresses this via a metagrammar. The 
modified version presented in this paper expresses this directly 
in the object grammar eliminating the need for a 
metagrammar. It treats the right hand side of a PS (Phrase 
Structure) rule as a set or partially ordered set. Parsing with 
this type of rule can proceed by checking set membership. 
2. The Word Order Problem in General 
The word order problem is the problem of processing the 
whole range of word order variation occurring in natural 
languages. Some Australian languages such as Warlpiri show 
extreme word order variation (Hale 1983). Hindi, Japanese 
and German also allow considerable word order variation. In 
this section we develop descriptive formalisms and parsing 
mechanisms that are adequate for the whole range of word 
order variation. 
Consider a grammar thdt allows a node labeled S to have 
daughters labeled $, O, and ¥ in any linear order~ and 
nothing else. Such a grammar can be presented with a set of 
rules such as that given in (2.1). 
(2.1) S --> $ OV, S --> $VO, S --> 0 SV, 
S--> OV$, S--> V$O, S--> VO$ 
The problem with a grammar such as that given in 
(2.1) is that i~ needs too many rules to capture word order 
variation (in this case free word order). For 5 'words' such a 
grammar will need 5! = 120 rules. With the increase in the 
number of words, such a gralnmar will grow factorially. That 
is, for a number of words it will need n/ rules. 
There is a convYnient way of 'collapsing' rules in GPSG 
(Generalized Phrase Structure Grammar) of Gazder (1981). It 
uses metarules that operate on basic rules to generate derived 
rules which then function as basic rules in derivations. Thus, 
(2.1) can be abbreviated as (2.2). 
(2.2) Basic rule: \[_~ $ O V\] 
Metarule: \[5....X...Y...\] =:> \[S....y...X...\] 
where X and Y range over $, O, V. 
Within GPSG Pullmn (1982) suggests another solution 
which also involves a metagrammar, tie suggests that a 
grammar such as (2.1) can be expressed via a metagrammar 
that treats immediate dominance and linear precedence 
separately. Pullum's theory is known as ID/LP analysis 2 
According to this theory grammar (2.1) "would be specified by 
means of the metagrammar" given in (2.3). Similarlyj the 
metagrammar given in (2.4) "determines" the grammar shown 
in (2.5). In (2.3) and (2.4) i ..... diate doml .... e star .... ts 
are given under a~ and linear precedence statements are given~ 
under b. In the case of (2.3) however the set of linear 
precedence statements is empty. In the case of (2.4) $ ( O 
means 'it" any rule introduces $ and O, $ linearly precedes O'. 
(2.3)a. {S--> $,O,V} b. {$} 
(2.4)a. {S--> $,O,V} h. (*<O} 
(2.5) (s->.*ov, s->$vo, s->v$o} 
An important advantage of ID/LP analysis is that it 
can account for word order variation in a general way, 
capturing "analytical intultion~ often hinted at in the 
literature, that fixing constituent order "costs" in the same 
way that having special NP case-marking rules or verb 
agreement rules does" (Pnllum 1982: 211). The main 
disadvantage of the standard ID/LP framework is th~:t it is 
difficult to process (Shieber 1983, Barton 1985). 
The alternative solution proposed in this study treats the 
rigbt hand side of a rule as a set3 . Thus~ the grammar in 
(2.1) can be p ..... ted in this format, eith .... (2.6a) or as 
(2.6b). The latter rule is to be understood under the node 
admissibility condition. 
(2.a)a. S->($,o,v} b. (5 $'°'v} 
Since the right hand side of the rule is a se G the order of 
$~ O and V does not matter. In parsing, this solution has 
definite advantages. Firstly~ the factorial growth of rules is 
eliminated. Secondly~ parsing can proceed by checking set 
membership or set difference. That is, instead of 'ordered 
match' the parser has to do 'unordered match'. The precise 
way of doing it will vary from parser to parser. We describe 
one way of implementing it in the ATN (Augmented 
Transition Network (Woods 1970, Finln and Hadden 1977)) 
formalism. 
Consider the ATN fragment presented below in (2.7) for 
the grammar given in (2.6). Conditions on are are given in . 
LISP like structures within parenthesis. Thus (null $) means 'if 
65 
S-reglster is empty'. By the arc WRD $ (null $), the 'word' $ 
will be accepted if no $ has previously been found. (In natural 
language, one can assume $ ~ Subject, O ~ Object, V = 
Verb, and use PUSH arc in place of WRD in the following 
diagram.) 
(2.7) 
WRD, (N~LL 
~ :23 WRDO(NULL_ O) / / 
WRD V (NULL V) J 
~ 4 POP (NOTNULL $)(NOTNULL O)(NOTNULL V)~ 
(2.7) parses any strings generated by (2.1) by scanning 
the input from left to right and checking the set membership. 
Thus, in recognition and parsing, (2.7) correctly reflects (2.0). 
Suppose, the input string is VO$. One way to see how a 
sentence is parsed is to trace through the analysis of the 
sentence as are sequence. This string is accepted by the arc 
sequence (3, 2, 1, 4). The ATN given in (2.7) can be said to 
have conditioned multiple loops. For convenience of 
reference, we shall refer to ATN structures such as (2.7) as 
'set-loops'. Further restrictions on set-loops (such as (2.7)) can 
be imposed and all constituent order variations can be parsed 
simply by imposing additional conditions on arcs. Thus, an 
ATN parser such as (2.8b) can parse the language generated 
by the grammar given in (2.5). Formally, (2.5) is presented 
with a partially ordered set such as (2.8a) in the proposed 
framework. The partial ordering is specified as a constrain 
after "/"~ as in a context-sensitlve rule contexts are specified 
after "/''. 
(2.8)a. ( *,O,V}/$<O 
(2.8)b. ( 
WRD $ _(NUL____~L 
~D O (NULL O)(NOTNULL $) 
WRD V (NULL V) 
POP (NOTNULL $)(NOTNULL O)(NOTNULL V~ 
Suppose that $, O, and V are nontermlnals which are 
further expanded by appropriate rewrite rules. Right hand 
side of such expansions can also show word order variations as 
shown in (2.9). 
(2.9) $ --:> {a,b,c}, O --> (d,e}, V --~> {f,g,h} 
So far, we have described parsing strategies for 
constituent order variations. However, in natural language we 
often find a discontinuous constituent. That is, an element 
can be moved out of its constituent (toplcalization in English 
would be a good example if VP is a constituent) which can be 
described by categories with holes (eg. VP/NP). In cases such 
as this, VIR arcs in combination with hold lists arc used in 
ATN (Bates 1978). Alternatively, temporary registers can be 
used to parse discontinuous constituents. Temporary 
66 
registers are particularly suitable to handle large number of 
'misplaced' words that cannot be handled by usual HOLD lists 
in combination with VIR arcs. We would llke to apply the 
general strategies described above to the case of Hindi which 
shows considerable word order variation. 
3. Word Order in Hindi 
In Hindl, the order of the major constituents such as $ 
(Subject), O (direct Object), I (Indirect object), and V (Verb 
(+aux)) is free. For example, out of the four constituents 
present in (3.11.1), we can make twenty four variants of the 
same senteneep all of which are perfectly good in Hindi as is 
obvious from (3.11.1-24). 
(3.11/1. mohan ne raam ko sev dlaa thaa. ($IOV / 
Mohan ag Ram to apple gave was 
"Mohan gave the apple to Ram." 
2. mohan ne raam ko dlaa thaa sev. ($IVO) 
3. mohan ne diaa than raam ko sev. ($VIO) 
4. mohan ne diaa than sev raam ko. ($VOI) 
24. diaa thaa sev mohan ne raam ko. (VO$I) 
To capture the freedom of order of $, I, O~ and V in 
sentences such as (3.11) We can have a rule such as (3.12) in 
the grammar of Hindl. The V alone can stand as a sentence in 
Hindi since it is highly inflected (see Kachru 1980). Hence 
(3.13) is more appropriate for Hindi where $~ I, and O are 
given within paratheses to show their optional occurrence. 
(3.12) (S ~'I'O'V } 
(3.13) ( ~ (,), (I), (o), v } 
We have been referring to rules such as (3.13 / as set 
rules. An ATN fragment, such as (3.14) would be appropriate 
for (3.13 / . 
IC PUSH CNULL,I _j / / 
5 POP (NOTNULL V) f 
(Assume appropriate subnets for $, I, O, V) 
Suppose we are parsing (3.11.1) mohaa nc raam ko sev diaa than 
"Mohan gave the apple to Ram". It is accepted by the arc 
seq ..... (1, 2, 3, 4, 5). The sentence given in (3.11.24) is 
accepted by the arc sequence (3, 1, 4, 2, 51 . (3.14) captures 
constituent order variation in Hindl in a general way. 
However, it is to be noted that sentences such as (3.11.1) have 
bl-transltive (or double transitive) V. We have to impose 
more conditions on arc 5, POP, to parse intransitive and 
transitive sentences. Informally, the conditions are: (1 / If the 
V is intransitive then the I and O must be empty. (2) If the V 
is transitive then the I must be empty. We have implemented 
a large parser of Hindl wlth wide coverage of construction 
types including relative clauses, interrogatives, passives, 
dative subjects, compound verbs and gapping which interact 
with word order variation (see Dey 1982, 1984). 
Word order variation in Hindl is fairly restrictive. Thus, 
in the sentences of (3.11) the main verb must precede the 
AUX. (3.15) is unacceptable because it violates this restriction. 
(3.15) * thaa dlaa sev mohan ne raam ko. 
was gave apple Mohan ag Ram to 
Similarly, (3.16) violates the restrictlon that the case elements 
must follow the noun (Verma 1970). 
(3.18) * m~ mohan ram ko sev diaa thaa. 
ag Mohan Ram to apple gave was 
As obvious frmn (3.15-16), all rules of tIindi are not 
'set.rules'. Thus, the subject"NP, $, cannot be expanded by 
(3.17a); rat\] .... w .... th ...... 1 (3.17b). 
(3.17)a. {-*e NP, K } b. \[ $ NP K \] 
(Assume \[NP mohan\], \[K ne\]) 
In the modified ID/LP fl'amework we allow strict order 
rules such as (3.17b), fr .... d ..... I ...... h as (3.14) and 
partial order rules such as (2.8a). We also allow notions llke 
subject and object. That means the grammar is an annotated 
PS grammar. The parsing strategy suggested above for this 
grammar has an important consequence. It does not recognize 
VP (thab dominates V, O, I) as a constituent. It advocates a 
'flat' structure for sentences as shown in (3.1.8). 
(3.18) S 
( 1 o v 
mohan ne raam ko sev dl aa 
"Mohan gave tbe apple to Ram." 
It should be noted that actual structural representations 
should be given with mort details. Some parse trees given by 
the parser are presented below: 
(3.19) (p ..... (inoh ....... aln k .... dlaa thaa)) 
(S (NP-subj (NP (DEW nil) (ADJ) (N mohan)) (K-ag he)) 
(NP-ind (NP (DET nil) (ADJ) (N ...... )) (K-dat ko)) 
(NP-obj (NP (DET nil) (ADJ) (N sev))) 
(VX (ADV) (V dlaa (AUX thaa))))t 
(3.20) (parse (diaa thaa sev raam ko mohan ne)) 
(S (NP-sul,j (NP (DET nil) (ADJ) (i mohan)) (K-ag ne)) 
(NP-ind (NP (DET nil) (hDJ) (i raam)) (K-dat ko)) 
(iP-obj (il' (DET nil) (ADJ) (N sev))) 
(VX (ADV) (V diaa (AUX thaa))))t 
It is t~o be noted that though case words llke ne and ko 
often help to identify subjects, objects etc. the parser must 
use semantic information in order to identify them in 
sentences mlch as the ones given in (3.21-22) (see Dey 1984). 
(3.21.) (parse (mohan anDaa khaataa hat)) 
Mohan egg eats is 
"Mohan eats an egg" 
(S (NP-subj (NP (DET nil) (ADJ) (N mohan)) (K-ag nil)) 
(NP-Ind nil (K-dat nil)) 
(NP-obj (NP (DET nil) (ADJ) (i anDaa))) 
(VX: (ADV) (V khaataa (AUX hal))))t 
(3.22) (p .... (anDaa mohan khaataa hal)) 
"Mohan eats an egg" 
(S (NP-subj (NP (DET nil) (ADJ) (N mohan)) (K-ag nil)) 
(NP-ind nll (K-dat nil)) (NP-obj (NP (DET nil) (ADJ) (N anDaa))) 
(VX (ADV) (V khaataa (AUX hai))))t 
4. Concluding Remarks 
Processing word order variation with new techniques 
within the modified ID/LP framework seems to be revealing. 
But, it is not eontext"fl'ee nnlike other ID/LP based parsers. 
Detailed comparison of II)/LP based parsers is a subject of 
further research. 
Footnotes: 
1. I am grateful to A. K. Joshi, A. Kroch, T. Finln, D. 
Itindle, S. Gambhlr, K. Reilly, D. Kaemmerer, K. Ryan, II. 
Bullock and the anonymous COLING-86 referees for thelr 
helpful suggestions and comments. 
2. See Uszkoreit (1982) for an implementation of ID/LP 
framework. 
3. The right hand side of a rule should be treated as a 
restricted set rather than as a pure set. The restriction can bc 
stated as follows: a membvr of a set can occur only once in the 
set unless specified otherwise. Thus, though formally the 
following two sets are equal, under the restrictions imposed 
they are notequah { $, O, V } =#= { $, O,V, $,V } 

References

Barton, G. \]5. Jr. 1985. "On the Complexity of ID/LP Parsing" 
Computational Linguistics, 11, 205-218 

Bates, M. 1978. "The Theory and Practice of Augmented 
Transition Network gramnlars". In L. Bole (ed.) Natural 
Language Communication with Computers. Spring Verlag, 
\]\]erlin: 191-259. 

Dey, P. \]982. "A Parser for Hindi". Presented to 4th South 
Asian Languages Round Table, Syracuse, 1982. 

...... 1984. Computatlonally Efficient and Linguistically 
Adequate Parsing of Some Natural Language Structures. 
Ph.D. diss., University of Pennsylvania. 

Finin, T. and G. IIadden \].977. "Augmenting ATNs". In the 
Proceedings of the 5th IJCAI. 

Gambhlr, V. 1980. Syntactic Restrictions and Discourse 
Functions of Word Order of Standard tIindl. Ph.D. 
diss., University of Pennsylvania. 

Gazdar, G. 1981. "Unbounded Dependencies and Coordinatc 
Stmlcture", Linguistic Inquiry 12, 155-184. 

Hale, K. 1983. "Warlpiri and the Grammar of Non- 
configurational Languages," Natural Language and 
Linguistic Theory, 1.5-48. 

Kachru, Y. 1980. Aspects of Hindl Syntax. Delhi: Monohar. 

Pullum, G. K. 1982. "Free Word Order and Phrase Structure 
Rules," NELS, 12, 209-222. 

Pullum, G. K. 1983. "Context-freeness and the Computer Processing of 
Human Languages," Prec. of the 21st ACL Conference. 

Shleber, S. 1fl83. "Direct Parsing of ID/LP Grammars," 
Lingnlstlcs and Philosophy 7:2. 

Steele, S. 1981. "Word Order Variation: A Typological Study," 
in J. Greenberg (ed.) Universals of Language, Vol. 4. 
Stanford, CA: Stanford University Press. 

Uszkorelt, H. 1982. "A Framework for Parsing Partially Free 
Word Worder," Proceedings of the 21st ACL Conference. 

Verma, M. K. 1971. The Structure of Noun Phrase in English 
and Hindi. Delhi: Motilal Banarsldas. 

Woods, W.A. 1970. "Transition Network Grammars for 
Natural Language Analysis," Comm. of ACM \]3, 591-60{5. 
