Information-based Case Gramlnar 
Keh-jiann CHEN 
Institute of h2Jbrmation Scietzce 
AccMemia Sinica 
Taipei, Taiwwz 
R.O.C 
Clm-Ren HUANG 
b~stitute of Itistory and Philology 
Academia Sinica 
Taipei, Taiwwt 
R.O.C. 
KCHEN%IS @TWNCTUO1.BITNET HSCHUREN@TWNAS886.BITNET 
Abstract: In this paper we propose a framework of Information- 
based Case Grammar (ICG). This grammatical formalism entails 
that the lexical entry for each word contain both semantic and 
syntactic feature structures. In the feature structure of a phrasal 
head, we encode syntactic and semantic constraints on 
grammatical phrasal patterns in terms of thematic structures, and 
encode the precedence relations in terms of adjunct structures. 
Such feature structures denote partial information which defines 
the set of legal phrases. They also provide sufficient information 
to identify thematic roles. With this formalism, parsing and 
thematic analysis can be achieved simultaneously. Due to the 
simplicity and flexibility of Information-based Case Grammar, 
context dependent and discontinuous relations such as 
agreements, coordinations, long-distance dependencies, and 
control and binding, can be easily expressed. ICG is a kind of 
unification-based formalism. 'H:erefore it inherits the advantages 
of unification-bmscd formalisms and more. 
In the feature structure of a phrasal head, we encode syntactic 
and semantic constraints on grammatical phrasal patterns ha 
terms of thematic structures, and encode the precedence relations 
in terms of adjunct structures. The feature structure of a 
potential phr~al head denotes partial information for defining the 
set of legal/grammatical phrases. It also provides enough 
information to identify the thematic roles for arguments and 
adjuncts \[Chert 89\]. In other words, with ICG, parsing and 
thematic analysis are achieved simultaneously without additional 
operation; and generation with thematic structure can be done 
with the identical formalism. 
We take Mandarin Chinese as our representational target. 
Thus, the features were selected to account for Chinese only. 
Itowever, the abstract design of this formalism is not limited to 
only the representation of Chinese. Since the Chinese lexicon is 
impoverished in inflection, it is necessary to fully stipulate both 
semantic and syntactic information for the purpose of both 
parsing and generation. Furthermore the precedence relationship 
of constituents is defined over thematic roles. This seems to be 
more appropriate for Chinese. By coincidence, Bresnen and 
Kanerva's \[Bresnan 89\] lexical mapping theory represents a shift 
towards the possibility of senmntics major approaches. 
l. lntroductiozz 
In this paper, we plopose a lexicon-based gramrrmtical 
formalism called hfformatiou-based Case Grammar (ICG). This 
forma\[isnl entails that the lexical entry ff~r each word contain 
both sen:antic and syntactic information. It wiI1 bc argtmd that our 
lexicon-based representation approach better focuses information 
for parsit:g a::d generatiotL hi colltrast, the phrase-structure rule 
approaches lack ,ule focusing capability. Even with the LR 
parsing strategy \[Tomita 8@ ulmecessary branchiltg and 
backtracking cannot be avoided when adopting these approaches. 
Therefore, modern linguistic theories share the tendency to be 
lexicon-based and to reduce I'S rules. For instance, LFG and 
ftPSG stipulate the argumetU structure as one of the attributes 
for each verb \[Bresnan 82, Pollard 87\] and Karttunen \[Karttunen 
86\] proposes a radical lexicatism's approach to do without PS rules 
by representing syntactic itfformation in terms of categorial 
grail\]mar. Categorial granunar, however, offers no 
straightfop.vard and intuitive mechanisms to handle context 
dependent or discontinuous relations such as control and long 
distance dependency \[Uszkoreit 86\]. llence, we adopt an 
alterlaative approach equivalent to the ID/LP (immediate 
dominance and linear prcccdcace) format of GPSG \[Gazdar 87\]. 
2. Feature Structures for Mandarin Chinese 
Chinese is a weakly marked language with no inflection. 
Nevertheless, the linear order of arguments and adjuncts are 
relatively free. Hence syntactic-only representations would cause 
tremendous ambiguities. For representational precision and for 
parsing adequacy, semantic information is indispensibte. The most 
important semantic information includes 1. argument structures 
and their semantic restrictions, and 2. the semantic features for 
each word which ;ire necessary to identify thematic roles, l tcnce 
the following feature structure (1) was selected so that each 
texical entry can be uniformly represented by the same structure 
with lexicalty or syntactically defined value (including null). 
(l) Semantic: 
Syntactic: 
Meaning: 
Features: 
Arguments: 
Adjuncts: 
'Class: 
Constraints:\[Form: 
\] Basic Patterns: 
\[Adjunct Precedence: 
A typical example (2) is given here i)efore we introduce 
and summarize the values for each feature path below. 
54 1 
(2) C'hiuan °persuade": 
Semantic: meaning: "persuade" 
features: 
arguments:f AGENT: feature: +Human 
|GOAL : \[\] feature: + 11uman 
\[ feature: +Event \[I'HEME: \[ argument: AGENT: \[\] 
adjuncts: time, location, manner, .... 
Syutactic: chtss: VNI,,V p 
constraints: form:( time \[{NP, DM, PP, GP, ADV}, +time\] 
\[location \[{PP, ADV}, +location\] l 
mammr \[ADV, + m.anner\] 
BP: AGENT \[NP\] < * < GOAL \[NP\] < TItEME \[VP\] 
AP:(1. {time, location} < * 
L 2. AGENT < rammer < * 
Meaning: an atomic value denoting the motoring of tile 
word. 
Features: a set of atomic values which are tile semantic 
features of the word; e.g. + animate, -~ 
physical. 
Arguments: a set of thematic arguments for the head if 
the word is a possible phra.sal head; null 
othmwise. The value for each thematic 
argument is a feature structure of the same 
type ~u~ the value for tile feature path 
'semantic'. 
E.g. 'a nice boy' is an agent, as in 'A nice 
boy drove the cattle.' 
(3) AGENT: 
• meaning: 'boy' 
features: + 1 Itmmn 
arguments: none 
adjuncts:/quantifier:,naeaning: 'a' 
\] \[fcatures: -defirfite, + singular 
\[property: \[meaning: 'nice' 
Argument structures for a verb are equivalent to 
case frames and case restrictions of this verb 
\[Fillmore 68, Winograd 83\]. C~me restrictions 
indicate semantic preferences of thenmtic roles and 
function as a guide to identifying each case role 
\[Chen 89\]. The argument structures fur other 
phrasal heads such as prepositions, post-positions 
;.tnd conjunctions serve similar purposes. 
Adjuncts: a set of pcrmissible adjuncts of the head word. 
The value of each adjunct is a feature structure 
of the same type as arguments. 
E.g. 'yestm day' 
(4) time: meaning: 'yesterday' 
features: + time 
arguments: none 
\[ac juncts: none 
Arguments, adjuncts and head form the right- 
hand side of an immediate dominance rule. 
Syntactic Class: atomic values denoting the syntactic class 
of the word. 
"Ilm syntactic class of a word serves two 
purposes. The first is to denote the syntactic type. 
The second is as an index for inheriting common 
syntactic properties belonging to the mother node 
in the syntactic hierarchy, 
Syntactic Form: a set of syntactic and semantic 
definitions for arguments and 
adjuncts. 
"Ilm syntactic form for each thematic role in fact 
constrains the syntactic structures and semantic 
features of this role. We consider semantic 
restriction as part of the syntactic constraint. For 
instance, temporal expressions are instantiated by 
at least five different syntactic categories which are 
nouns phrase (NPs), compounds with 
determinatives and measures (DM, such as Swz. 
ditto Shi-fen three-hours ten-minute, i.e. 'three- 
ten'), post-position phrases (GPs), preposition 
phrases (PPs), br adverbs (ADVs). They all share 
a cornmon semantic feature +time regardless of 
their categories. Therefore tile temporal 
cxprcssions can be expressed as time \[{NP, DM, 
PP, GP, ADV}, +time\]. For complex expressions, 
we adopt the choice system used by tlalliday in the 
systemic grammar \[Winograd 83\]. 
Basic Patterns: a set of linear precedence rules governing 
arguments and heads. 
The basic patterns for verbs denote the possible 
sententkd patterns, including optional argument 
omission. The familiar syntactic generalizations 
with regard to passivizatio% topicalization etc. can 
be captured by lexical rules \[Gazdar 85, Pollard 87, 
Sells 85\]. Adopting the lexical mapping theory, 
however, is a theoretical possibility to make basic 
patterns obsolete \[Bresnan 89, Huang 89\]. 
2 55 
Adjunct Precedence: a set of linear precedence and 
cooccurrence constraints for 
adjuncts. 
The following notations were adopted. 
* : denotes a phrasal head. 
< : e.g. "a < b" denotesa precedesb. 
< < : e.g. "a < < b" denotes a immediately 
precedes 12. 
> < : e.g. "a > < b" denotes a and b can not cooccur. 
{ }: e.g. "a < {t2,_c }" denotes a precedes both b and 
_c but there are no preced-ence 
constraints between b and c. 
(5) A nice boy persuaded John to go ,o school yesterday. 
r meaning; "persuade" 
features: mst 
arguments 'AGENT: "meaning: "boy" 
feature: + }luman 
adjuncts:\[ quantifier:\[meaning: "a" 
\[ feature: -definite, +singular 
t property:(meaning: "nice" 
GOAL : L\[t me:ruing: "John" 
tfeamre: + \]hllrian 
TIIEME: meaning: "go to" 
feature: + Event 
arguments:/AGENT: El 
\[GOAL: meaning: %chool" 
adjuncts: \[ ,hne: \[ mcani ng: "yesterday" 
\[ { features: + time 
Although Basic Patterns and Adjunct Precedence both 
govern !inear precedence relations, they differ in their 
nature. Basic Patterns encode phrasal heads and their 
arguments. Linear order is but an additional piece of 
information describing the pattern. It is conceivable, i.e. in 
non-configurational languages, that linear precedence 
relations play no role in Basic Patterns. Adjunct 
Precedence Rules, on the other hand, are constraints on 
linear precedence relations among possible adjuncts. In 
other words, linear order is central to Adjunct Precedence 
statements while the presence of each adjunct is not. 
"I2ms, a legal phrase can be viewed as a sequence of 
thematic roles arranged in a proper order defined by one 
of the basic patterns and satisfying all the constraints of 
Adjuncts Precedence when qpplicable. The division of 
Basic Pattents from Adjunct Precedence is similar to the 
ID/LP format with the additional inforrnation 
differentiating adjunts from arguments. The limhed 
numbers of thematic cases, syntactic categories, and 
semantic features require only a finite number of 
notational symbols \[Gazdar 87\]. "~erefore we claim that 
ICG fails into the ch,ss of context-free grammars. 
The semantic features would be unified during parsing 
while the syntactic features are no more than constraints 
guiding appropriate unification. We \[lave a scheme to 
identify the thematic roles by' fully utilizing the semantic 
af~d syntactic information \[Chen 89\]. In \[Chen 89\], Chen 
et al. propose an information accumulation scheme 
(incremental description refinement in \[Mellish 88\]) for 
identifying thematic roles with the parametricalized 
information encoded with the ICG formalism. The four 
types of parametric infornaatitm used in Chinese are: 
a. th_2e ~utactic ~ and semantic features of 
the constituent, 
b. the case frame and sern\[n~tic rcstrictioas of the 
verb, 
c. tlae ~ntactic configuration and word order, and 
d. obliqE_e " case assi~gner, includin~ rp_L~msitions and 
postpositions. 
Following is the parsing result of 
3. 77ze Formal Definition for hzformation.based Case 
Grwnmar: The Lexicon and Principles 
rl'be ICG is composed of two major components. One is the 
lexicon which is the set of feature structures ms described in 
section 2. The other is the principles. Each feature structure can 
be viewed as a set of interpretable representations of syntactic and 
semantic information governed by a formal synt~. A parser or 
generator interpretes feature structures while parsing or 
generating sentences. The interpretation processes are guided by 
the principles of the grammar, qhe principles define well- 
formedness conditions and the rules for infommtion management 
for sentences and phrases. The parser or generator takes lexical 
information and unify it in such a way that changes of lexical 
information woukl not affect the phr,'ksing or generating process. 
Therefore we claim that ICG is declarative, qlm major principles 
of ICG are summarized below: 
1) I Iead Driven Principle 
The feature structure of a head contains the partial 
information that defines the permissible set of phrases 
with it ,'ks a phrasal head. The possible head types and 
respective phrase types for Mandarin Chinese are as 
follows. 
(6) 
I Icad Pbrasc 
V S, VI' 
N NP 
Prcp pp 
Post P GI' 
Determinant DM (determinat ire 
atld nleasurecouq~mtnd) 
Conjunction CP (conjunctive phrase) 
The syntactic classes of the heads determine the 
syntactic types of their projections. VP is defined 
as an S without subject \[Gazdar 87\]. 
56 3 
2) Well-formedness Conditions 
Like LFG \[Bresnan 82\], we have tim following 
well-formedness conditions: 
a. Completeness and functional biuniqueness 
conditions, 
b. Coherence conditions, 
and c. Linear precedence and syntactic form constraints. 
In fact, the completeness condition is enforced with 
respect to Basic Patterns (BP) only. The cases of 
argument omission are idiosyncratically determined by 
verb classes, and are nut governed by the completeness 
condition, On the other hand, adjuncts are optional and 
constrained orfly by linear precedence rules AP and form 
restrictions. The functional uniqueness condition is also 
relaxed a bit to account fur cases of multiple occurrences 
of some adjuncts such as modal at the sentential level, 
and rop_Lg.p_~ ~ on the noun phrase level, by annotating 
Kleene's star on top of the adjunct modal and the adjunct 
3) Feature Propagation Principles 
a. Head feature conve,)tion \[Gazdar 85\], 
b. Foot feature principle \[Gazdar 85\], 
e. Explicit feature paths: 
Explicitly denoting the daughter's feature in 
the mother node by a feature path such as the 
'argument feature', 
d. Conditional features: 
Locally ambiguous semantic features can be 
expressed by conditional features. For 
iJrstance, the preposition bei 'by' can mark 
an agent, a causer, or an instrument 
depending upon whether the arguments are 
animate, nonphysical, or physical inanimate 
respectively. Therefore the feature structure 
of bei 'by' would be (7). 
(7) Bei 'by': 
Scm feamlcs: I)UMMY feature, 
I. AG EN:I', 
2.CAUSEP,, 
3.1NSTRUM IEN'I" 
llrgtllllCIIl: I)UMMY:ffeann es: 1, + aniw~ilte 
L 2. -phvsic:tl 3. -illlilll;lle 
ft~lm: I)IJMMY\[NI'\] 
liP: 'bci" << I)UMMY 
There are many possit)Ic types of senlantic features allowed in 
ICG. qlw.y are differentiated by attribute value pairs. Tim three 
required types of semantic features for Chinese are 1. semantic 
classes e.g. + animate 2. syntactic/semantic feature, e.g. + 
NEG 3. thematic roles, e.g. AGENT. Different types of 
features can be identified simply by their attribute names e.g. we 
know AGENT in <ARGUMENT AGENT> is a thematic role. 
The head feature principle is thc same as the one in GPSG 
\[Gazdar 85\] which states that in any local subtree, the head 
features of the mother are identical to the hcad features of the 
head daul, hter. Similarly, our Foot feature principle also follows 
GPSG. Roughly speaking, foot features are passed up from 
any daughter in a tree, with the upper and lower limits of this 
propagation are determined by prior specification \[Gazdar 85\]. 
In Chinese, +_. Question, + Negation, + Plural, + Definite etc. 
are all considered foot features. 
The semantic class of a thematic role is usually determined by 
its head daughter. However, for the marked cases with the 
syntactic categories of PP/GP, the semantic classes are 
determined by complement daughters. We can not define every 
semantic type ,as a foot feature. Therefore, we propose to 
explicitly state the daughter's feature in the mother node by a 
feature path such as 'DUMMY featm'es' in (7). 
4. What Makes ICG a Good Representational Language 
We think that the simplicity and flexibility of the ICG 
formalism makes it a good representational language. It is simple 
since ICG is a type of context-free gramnmr and the attributes for 
feature structures are nniform for all different types of phrases. 
Preparation of lexical feature structures are straightfonvard. 
Linguists can start with the categorial feature structure of each 
entry by discovering idividual idinsyncracies and then modifying 
the categorial feature structure accordingly. As for flexibility, 
ICG is much more flexible than tile other context-free 
grammatical forms such as BNF, GPSG, etc., due to the wider 
scope of accessibility and the richness of informatioh encoding on 
each thematic constituent. Context dependent and discontinuous 
rdafions such as agreements, coordinations, long-distance 
dependencies, control and binding, can be easily expressed in 
ICG. 
4.1 Agreement altd Coordhmtion 
Agreement and coordination pose similar problems in 
representation (but different proNems in processing). Both have 
to express the relations between daughters. Coordination can he 
viewed as the agreement of syntactic or/and semantic classes 
between two daughter arguments of a conjunction. In contrast, 
other types of agreements are the relations between head 
daughter and complement daughters. Since ICG provides an 
explicit accessing capability to tim daughters features, both types 
of agreements can be easily solved. The only difference is that 
coordination requires a w~riable ranging over a finite domain 
of syntactic classes and/or semantic classes to denote the 
common features of two arguments (8). For instance, the subject 
verb agreement problems are solved by stating agreement 
constraints on each subject role in every basic pattern. For 
example, the AGENT of the verb "persuades" is expressed as, 
AGENT \[NP, + singular, + third\] in active voice. 
(8) 
nlcailil\]g : "and" 
feature : DUIVlMY feature 
arguments: DUMMY 1: feature: class x 
DUMMY 2: feature: class x 
4 57 
4. 2 Long-distattce DepeiMency 
Kaplan and Zaenen \[Kaplan 88, Kaplan 89\] proposed 
functional uncertainty as a new descriptive technique, assuming 
grammatical function in LFG, in order to account for long- 
distance dependencies. The schema of functional uncertainty 
avoid the. problem of infinite specifications so that the attribute 
positions for topicalizable constituet, us cot, ld be realized as a 
regular expression such as (T comp subj\[obj)=(T topic). This 
equation expresses the uncertainty about what the with-in clause 
functional roles of an extraposed topic might be. It offered a 
clearer and mr.~re accurate characterization of long-distance 
dependencies but still could not handle the case of context 
dependency in topicalization. For instance, there are many verbs 
in Chinese which do not allow the object to be topicalized. When 
such a verb is the head of aa embedded sentence, it is clear that 
(comp obj) is not toplicalizable either. There is no way to predict 
the topicalizable constituent in depth by way of regular 
expressions. There are similar examples in English. 
(9) 
a. 
b. 
Who/ did Mary think that Bill saw e i ? 
?Who/ did Mary quip that Bill saw e i ? 
Who/ did John tell you (that) Mary thought 
that Bill saw e i ? 
?Who/ did John tell you (that) Mary quipped 
that Bill saw e i ? 
lIowever we can solve such context dependent 
problenrs in ICG by recursively defining the topicalizable 
constituents in terms of topicalizable constituents of 
embedded sentences. The verb with sentential 
complements usually causes the problems. We may 
encode the topicalized sentence patterns ms a part of BP 
or as derived by iexical rules. The topic of the embedded 
sentence is ope of the topicalizable cot~sdtuents of such 
"¢erb.~< For instance the topicalized sentence pattern of 
example (2) is: 
TOPIC \[TOPIC (THEME)\]< AGENT\[NP\]< • < 
GOAL\[NI'\] < TttEME/TOPIC 
The topics of embedded sentences are defined rccursively 
under the sentcnce patterns of the verbs of embedded 
sentences. Detailed discussion is given in \[Chen 90, 
Huang 90\]. 
4.3 Control and Binding 
Functional control is tile relation that exists between an 
antecedent and the missing subject in an XCOMP or 
XADJUNCT \[Bresnan 82, Sells 85\]. The coindex label adopted in 
the unification-based grammars is a simple solution to such 
problems. We use the same scheme in ICG, e.g. in (2). 
Anaphoric binding was solved in LFG by the concept of F- 
command \[Bresnau 82, Sells 851 . The same concept is also 
applicable to ICG. 
5, CoHclusiotl 
From the brief sketch above, it is clear that e:xch lexical entry 
contains a large amount of redundant syntactic and semantic 
information shared by other entries belonging to the same 
category, llence, a more elegant and economic strategy is to 
form a category hierarchy and to store shared information on 
higher level nodes. Each lexical entry contains only individual 
idiosyncracies. Thus the redundancy in representation can be 
removed and data consistency can also be easily maintained. 
ICG is a kind of unification-based formalism. Therefore all 
the advantages of unification-based formalism were kept in ICG. 
Furthernmre, additional advaqtages are incorporated via the 
following design features of ICG. 
1. Declarative 
"II~e lexical information defines legal sentences or 
phrase. And the changes in the above information does 
not affect processing procedures and results. 
2. Algorithm-free 
Different control processes are allowed for parsing or 
generation. Regardless of whether it is sequential, 
parallel, or heuristic control, the result will be the same. 
3. Suitable for parallel processing 
The processes are head driven. Each possible head can 
initiate a phrasal construction process. Thus, processes 
can be executed in parallel. 
4. Allows a partial parse 
At any moment of the unification processes, the 
accumulated information shows the partial result up to 
that momem even if the input is ungrammatical. 
5. Semantic structure is universal 
The semantic information contained in the feature 
structures of each lexical entry is universal. Different 
languages differ with regard to syntactic of information. 
6. Incorporated with thenmtic role identification schen~e 
The parametrical information for identifying thematic 
roles can bo encoded in ICG such that syntactic parsing 
and semantic analysis can be done in parallel. 
7. Parsing result is a thematic structure 
Recent theoretical linguistic studies are concerned with 
the relationship between thematic structures and 
argument structures (e.g. Brensan and Kanerva 1989). 
Our formalism directly and explicitly represents thematic 
structures. 
Last but not least, we expect future studies of ICG to confirm 
tile plausibility of the following advantages. 
1. Efficient parsing 
ICG has the advantages of lexicon-based parsing 
which is better focused on the relavant syntactic and 
semantic information. 
2. Cognitive reality 
The language capability of a man seems to be 
improved day after day by learning and polishing 
lexical information. ICG is able to reflect this 
phenomenon. 
58 5 
3. Germration power 
ICG provides a way of generating surface 
seIlteilces froili thematic structures. The pragmatic 
consideration of the selection of the sentence 
patterns can be accomplished by incorporating the 
pragmatic features in the basic patterns and lexical 
rules. 
5. Ackeiowledge\]tzet~t 
P, eseaich of this paper was paltially suppoitcd by NSC grant 
//78-0-108-.E001-001 and the Electronic P, esearch and Service 
O~ganization, industrial Technology P, esearch Institute, Talwan, 
l<.O.C, under Contract No. X2-79007. Some of the examples are 
taken from the research results of the Chinese Knowledge 
Piocessin L, Group at the Cotnputing Ceil\[or of Academia Sinica+ 
We alone, of course, are responsible for any possible errors 
contah~ed in this paper. 

References

t;resnan, J., 1982. The ,\Ionia\] ltel;rescutation mf 
Giammatica! l<elations, Cambridge: MiT Press. 

FStcsp+an, J+ ~,.nd J. Kanc~va, 1989. l.ocative Inversion in 
Chichcwa: A Case Study of t:actorization in Graninlar. 
Linguistic Inquisy 20, pp. 1-50. 

Chef+. K. J. and C. S. Clqa, 198& The Design of a 
Coaceptual Str,.!ctuie and Its l.>.e\]atiou to the Parsing of 
Chipese Sentences. 1CCPCOL'88, Toronto. 

(?hen, K.J., C.I<. ttuang, and L.IL Chart+?, 1989. 'l+he 
Idcntificatio~ of Thenaatic Roles in Parsing Mandariu 
Chinese, l'ioceedings of I<:.OCL1NG II (1989), Taipei, 
Taiwaii, Pl;. 121- I.I 6. 

Chert, f<.J., C.I+L l lut,+r',g, and W.P. Chert, 1990. l<esolution 
of l-ong-Distance i)ependencics with Recursive 
Information l!m!;cdding, In preparation. 

Fiilmole, C. i968. The Case for Case. In I:+. Bach and P,. 
I larms (Eds.), Universals in l+irlb'uistic Theoiy, New York: 
I hilt, Rinehart, and Vv'ir~ston. 

Gaz¢lar, G. et al. 1987. Category Structures. CSLI report 
102, Stanford: Center for the Study of Iamguage and 
Information. 

(_iazda~, G., F. Klein, G.K. Pullurn, and 1.A. Sag, 1985. 
Geucralized Phrase Structure Grammar. Cambridge: 
BlackweN, and Cambridge, Mass.: Ihuvard University 
Press. 

l luang, C.I<. 1989. Mandarin Chinese and the bexical 
Mapping Theory. The 1989 International Conference on 
Sine-Tibetan lmnguages and Linguistics. llawaii. 

ttuang, C.P,., K.J. Chert, W.P. Chert and T.Y. Ilu, 1990. 
Resolution of Long-Distance Dependencies in Mandarin 
Chinese-With an Algorithm Based ot+. Functional 
Uncertainty. To appear in the Proceedings of the 1990 
lmernational Conference on Computer Processing of 
Chinese and Oriental Imnguages (ICCPCOL'90). 

Kaplan, R.M. and J. Maxwell, 1988. An Algorithm for 
Functional Uncertainty. Proceedings of Collng '88, 
Budapest, 297-302. 

Kaplan, P,.M. and A. Zaenen, 1989. Long Distance 
l)ependencies, Constituent Structure, and Functional 
Uncertah W. In M. Baltin & A. Kroch (Eds.), Alternative 
Conceptions of Phrase Structure. Chicago: Chicago 
University Press, 17-42. 

Karttunen, L, 1986. P, adical Lexicalism, CSLI Report No. 
CSL1-86-68, Stmffotd: Center for the Study of Language 
and Information. 

Mellish, C.S., 1988. hnplementing Systemic Classification 
by Unification, Computational Linguistics, Voi. 14 #1, 
40-51. 

Pollard, C. and 1. Sag, 1987. Information-based Syntax 
and Semantics, Vol. I. Fundamentals, CSLI Lecture notes 
• "¢13, Stanford: Center for the Study of lmnguage and 
Information. 

Sells, l>., 1985. \[.ectures on Contemporary Syntactic 
Theories. CSLI Lecture Notes no. 3. Stanford: Center for 
tile Study of Language alld Intormation. 

Shieber, S.M., 1986. Introduction to Unification-based 
Approaches to Grammar. Stanford: Center for the Study 
of Ixmguage and Information. 

Tomita, M., 1986. Efficient Parsing for Natural 
Language, Boston: Kluwer Academic. 

Llszkoreit, II., 1986. Categorial Unification Grammars. In 
Proceediws of Coling 1986. Bonn: University of Bonn. 
Also appeared as P, eport No. CSLI-86-66, Stanford: 
Center for the Study of Language and hfformation. 

Winogtad, T., 1983. Language a.s a Cognitive Processes, 
Vol. 1, Syntax, Addison-Wesley. 
