COLING 82, J. tIoreekp (eel) 
North-Holla~l Publilh~ Company 
@ Aca~¢m~ 1982 
DIALOGIC: A CORE NATURAL-LANGUAGE PROCESSING SYSTEM 
Barbara Grosz, Norman Haas, 1 Gary Hendrix, 1 Jerry Hobbs, Paul Martin, 
Robert Moore, Jane Robinson, Stanley Rosenschein 
Artificial Intelligence Center 
SRI International 
Henlo Perk, California 94025 
U.S.A. 
The DIALOGIC system translates English sentences into 
representations of their literal meaning in the context of an 
utterance. These representations, or "lngical forms," are 
intended to be a purely formal language that is as close as 
possible to the structure of natural language, while providing 
the semantic compositionality necessary for meaning-dependent 
computational processing. The design of DIALOGIC (and of its 
constituent modules) was influenced by the goal of using it as 
the core language-processing component in a variety of 
systems, some of which are transportable to nee domains of 
application. 
OVERVIEW 
The DIALOGIC system translates Engllsh sentences into representations of their 
literal meaning in the context of an utterance. These representations, or 
"logical forms," are intended to be a purely formal language that is as close as 
possible to the structure of natural language, while providing the semantic 
compositionality necessary for meanlng-dependent computational processing. The 
design of DIALOGIC (and of its constituent modules) was influenced by the goal of 
using it as the core language-processlng component in a variety of systems, some 
of which are transportable to new domains of application. 
Currently DIALOCIC is a core component of four systems being developed within 
several different research projects at SRI. One is the TEAM project, ~ whose goal 
Is to provide natural-language access to large data bases through systems that are 
easily adaptable to a wide range of new applications. Another, the RLAUS 
project, ~ is a longer-range effort to address basic research problems in natural- 
language semantics, commonsense reasoning, and the pragmatics. A third,project is 
investigating the problem of providing natural-language access to text. ~ A fourth, 
in which ~IALOGIC also plays an important role, is examining the development of 
formal grammars. 5 
DIALOGIC is divided into five modules coordinated by the DIAMOND executive system. 
DIAMOND is a modification of the executive system used in the SRI speech- 
understanding project \[Walker 1978\] and also in a task-dialogue interpretation 
I Now working at Machine Intelllgence Corporation in Sunnyvale, California. 
2 Sponsored by the Defense Advanced Research Projects Agency under Contract 
N00039-80-C-0645. 
3 Sponsored by the Defense Advanced Research Projects Agency under Contract 
N00039-80-C-0575. 
4 Funded by the Nations1 Library of Medicine under Grant IROI-LM03611. 
5 Sponsored by the National Science Foundation under Grant IST-8103550. 
95 
96 B. GROSZ et .1. 
system \[A. Robinson, 1980\]. It provides the formal language for defining the 
grammar and the control for parsing English sentences and translating them into 
loglcal-form expressions. 
The five modules are (1) the DIAGRAM grammar; (2) s set of semantic translators; 
(3) a set of basic semantic functions; (4) a scoping algorithm (for quantifiers 
and sentence operators); (5) a set of basic pragmatic functions. The remainder of 
this paper describes these components of DIALOCIC and presents an example 
illustrating how they coordinate in the interpretation of an utterance. A 
description of the logical form that is the target of DIALOGIC's interpretation 
processes may be found in \[Moore, 1981\]. 
DIAGRAM 
DIAGRAM is a general grammar of English. It now contains about 125 rule schemata, 
equivalent to about 800 individual rules. These define all common sentence typos, 
complex auxiliaries and models, complex noun phrases, nominalized sentences, all 
the common quantifiers, relative clauses, verbs with sentential complements, 
comparative and measure expressions, subordinate clauses and other adverbial 
modifiers. Conjunction, however, is limited to a few place-holders, pending 
further study of the problems it poses for constraining the number of syntactic 
analyses. A detailed description of DIAGRAM is contained in \[J. Robinson, 1982\]. 
Formally, DIAGRAM is an augmented phrase-structure grammar. The lexicon 
categorizes words and associates attributes with them that are used in the rules. 
Each rule has associated with it a constructor that expresses the constraints on 
its application and also a translator (described in the next section) that 
produces the corresponding logical form. 
Phrases inherit attributes from their constituents and acquire attributes from the 
larger phrases that contain them. These attributes are used to impose context- 
sensitive constraints upon the acceptance of an analysis. Before constructing a 
node in the parse tree corresponding to the application of a rule, the executive 
invokes the rule's constructor to test for admissibility. In addition to 
accepting or rejecting a rule application, the constructors can assign scores that 
allow listing alternative analyses in a preferred order. The result of applying 
the grammar to analysis of an input is one or more annotated ~arse trees. 
Attributes and annotations are not limited to syntactic information. The 
translators, described next, specify bow the translation of a phrase into logical 
form is to be defined in terms of the attributes of the words and phrases that 
compose it. This coupling of syntax and semantics (for which attribute ~rammars 
\[Tienari 1980\] were originally designed) is convergent with current formal 
theories bf natural language that advocate constructing a syntax and semantics 
that "work in tandem" \[Dowry et el. 1981; Kaplan and Bresnan (to appear); Gazdar 
(to appear); Landsbergen 1976.\] 
Future work on DIAGRAM includes efforts to extend both its coverage and its 
formalism. In extending the formalism, our dual objective is to capture certain 
linguistic generalizations (e.g., dative movement) and to make the task of 
developing a large grammar more manageable. To accomplish this, we are exploring 
the use of metarules \[Gazdar to appear\]. 
TRANSLATORS 
Followlng the syntactic analysis of an utterance, a sequence of semantic 
translators is invoked to -build the loglcal form that corresponds to a llteral 
interpretation of the utterance in context. The translator for each phrase- 
structure rule specifies how the various constituents of the phrase are to be 
combined to form an interpretation of the wbol~ phrase. It prescribes the 
predicate-argument structures that correspond to the grammatical construction or, 
more generally, the operator-operand structures. 
D1ALOGIc, A CORE-NATURAL-LANGUAGE PROCESSING SYSTEM 97 
Although the translators operate top-down (the translator for each node inyokes 
the translators for its children), the translation is in effect built bottom-up-- 
since, typically, the first thing a translator for a nonterminal node does is to 
invoke the translators for each of its constituents, usually left to right. 
However, the top-down nature of the translation process is significant, because it 
means that information located above a node and to its left is available when the 
node is translated. In addition to producing the logical form, the translators 
determine the syntactic constraints upon and preferences for either coreference or 
noncoreference of noun phrases, especially pronouns, following an algorithm 
described in \[Hobbs, 1976\]. 
BASIC SEMANTIC FUNCTIONS 
To insulate changes in the grammar from those that occur in logical form, the 
construction of the latter is isolated from the translator procedures by calls on 
basic semantic functions \[Konollge, 1979\]). The actual construction of a logical 
form is done in two phases: (I) ioglcal-form fragments (iffs) are attached to the 
parse tree by the basic semantic functions; (2) the final logical form is 
assembled from these by the scoplng algorithm. 
Lffs are assigned only to certain" nodes in the parse tree. Usually the iff at an 
NP node will encode the properties held by the entity the NP describes \[e.g., "X 
such that EMPLOYEE(X) & OLD(X)" for "old employee"\] and the fragment for a clause- 
level construction (e.g., a VP) will encode the predlcate-argument structure of 
the clause, 
The basic semantic functions also leave markers on the parse tree to indicate such 
things as the type of quantifier or deter~iner associated with a noun phrase. 
These markers are used by the scoplng algorithm to determine the final loglcal 
form for the utterance. (Note that the Iffs and markers left by the basic 
semantic functions may be viewed as further annotations to the parse tree.) 
DIALOGIC currently includes eleven basic semantic functions. Six of these do most 
of the work of buildlng lffs for standard noun phrases and clauses. The others 
are concerned with adding such things as mode, degree, and adverbial modification 
to clauses. As more precise specifications are defined for encoding these 
phenomena in logical form, we expect to collapse some of this latter group. 
SCOPING OF QUANTIFIERS AND OTHER SENTENTIAL OPERATORS 
The scoplng algorithm is designed to collect the loglcal-form fragments from the 
parse tree and produce the possible scoplngs of quantlflers and other scoped 
operators. The scoplng algorithm used in DIALOGIC (adapted from that in 
Hendrlx, 1978) produces all the scoplngs thnt do not vlolate the hard rules of 
Engllsh scoplng, and then ranks them according to a score computed by a set of 
speciallst critics. Each critic is a function that returns a score for some 
aspect of the conflicting rules of quantification in Engllsh; e.g., the left-rlght 
scope critic lowers the score of 8coplngs that involve permuting the left- 
outermost default ordering of quantlflers. All critics receive equal weight in 
the present implementation, but the design of the system does allow for 
differential weighting. 
The current set of critics is concerned with such things as changes in sentence 
order and the relatlve scoplng of quantlflers of different strengths. The scoplng 
of nonstandard quantlflers and of the generallzed negative ("not, .... no one," 
"nothing, .... none") remain to be done. 
BASIC PRAGMATIC FUNCTIONS 
Basic pragmatic functions are intended to fulfill several roles in DIALOGIC, all 
concerned with certain kinds of indetermlnacles in logical form whose resolution 
requires pragmatic information. The four primary uses of basic pragmatic 
98 B. GROSZ et -1. 
functions in the current system are (i) ~o provide a context-speclflc 
interpretation of certain terms that have only vague meanings in themselves (e.g., 
prepositions llke "of" and "~n~" or Inherently vague verbs llke "have"); (2) to 
establlsh the specific relationship underlying any given noun-noun combination; 
(3) to identify the referents of pronouns; and (4) to interpret a limited range of 
metonymy (e.g., the use of "blonds" to mean "people with blond hair"). At 
present, only a small core of pragmatic functions is implemented, each of which 
handles only a subset of the cases it is intended to cover. 
EXAMPLE 
To illustrate how the different modules' of DIALOGIC contribute to the 
interpretation of an utterance, we shall consider the example, 
"What SRI employees have children older than 15 years?" 
The logical form for this query--the target for the interpretation processes--Is 
(lowercase is used to indicate variables, uppercase to indicate constants and 
predicates): 
\[QUERY (WH employeel (AND (EMPLOYEE employeel) 
(EMPLOYEES-COMPANY-OF employeel SRI)) 
(SOME child2 (CHILD child2) (AND (CHILD-0F employeel child2) 
((*MORE* OLD) child2 (YEAR 15\] 
This corresponds rou@hly to a formal representation for "who is each employee such 
that the company of the ~nployee is SRI and some child of the employee is older 
than fifteen years?" 
During DIAMOND'S parsing phase, the parse tree in Figure I is constructed. At 
this polnt~ the attributes annotating the tree encode such properties as the type 
of noun (count, mass~ unit) and syntactic number. These attrlbutes have been used 
during the parsing phase to rule out certain alternative structures. 
Once this structure is built, the translators are invoked. In combination with 
the basic semantic functions, the translators assign addltlonal attributes to 
nodes in the tree, encoding such information as the quantlflers (type~ strength, 
and the variables they bind) and heads of noun phrases. For example~ the head of 
the WHNP, "what SRI employees"m is a variable of type EMPLOYEE that 14 bound by a 
wiT-type quantifier. Attributes also encode the underlying predlcate-argument 
structures for verb phrases and adJectlves~ and the iffs to be used in 
constructing the flnal logical form for the utterance. 
In the sentence of Figure I, the nodes WHNP and S are annotated as being 
quantlfledp WHNP with a wh-type quantifier and S as a "query." Although every 
rule has an associated translator, only some of these result in iffs being 
attached to nodes. For this example~ the nodes marked with *e in the original 
parse tree are the only ones for which Iffs are produced. 
The fragment attached to each of these nodes is as follows: 
NOUN1 (*NN* employeel SRI) 
NOUN2 (EMPLOYEE employee1 ) 
PREDICATE (*HAVE employee1 child2) 
NOUN3 (CHILD child2) 
NCOMP ((*MORE* OLD) child2 (YEAR 15)) 
EMPLOYEE, CHILD, and OLD are" monadlc predicates that are part of the conceptual 
model of the domain. *MORE e maps a predicate into a comparative along the scale 
corresponding to the predicate. *NN* and *HAVE are dummy predicates that indicate 
the need tO invoke the basic pragmatic functions. 
After the translation process is complete, the final loglcal form is asseLbled by 
a procedure that considers alteruatlve quantifier scoplngs (using the quantifier- 
DIALOGIC, A CORE-NATURAL-LANGUAGE PROCESSING SYSTEM 99 
/ 
I /---- --\ 
MSDET NO~D 
I 
NOUHI ** I 
/ \ 
NHD NOUN2 ** \[ 
-S N 
/ 
VPT 
t 
V 
4 J 
I 
\.WHAT SRI '-S B~q,OYEE HAVE -'8 
/ \ 
SNHQ ENDPUNCT 
J i ? 
/ 
m 
t 
NorM3 ** 
f /----~ 
-S N 
PREDICATE ** 
t 
VP 
I \ 
NP \[ 
\ 
NCONFS 
t 
NCOMP *t 
f 
PREDADJ 
I /~ -------~ 
ADJT 
I i / \ /------ ~\ 
ER ADJ THAN 
f I 
I I 
I I /-- 
I I 
i I I I 
CHILD ER OLD THAN l$ 
HP 
f 
DETP 
J DEI'Q 
I 
ImUN 
i /---- ---~ 
• "-S B 
I i 
-..S YEAR 
Figure I Parse Tree for "What SRI employees have children older than 15 years?" 
related annotations left on the parse tree) and invokes the basic pragmatic 
functions as needed. The basic pragmatic functions use information in the 
conceptual model of the domain to transform (*HN* employeel SRI)--corresponding to 
the noun-noun compound "SRI employee"--into (EMPLOYEE-OF employeel SRI) and (*HAVE 
employeel child2) into (CHILD-OF employeel child2). 
The nodes with either quantifier or ioglcal-form markings are the only ones 
considered by the TEAM scoplng algorithm. Besides the WH quantifying employee1, 
TEAM recognizes that a default existential quantifier must be created for child2~ 
so SOME is added. The scope rules force QUERY to have the widest scope; this 
position 18 contested only if there are multiple sententlal markers. Both 
orderlngs of the WHAT and SOME quantlflers are generated. The two resultlng 
quantified statements correspon d to (WHAT employeel (SOME child ...)...) and 
(SOHE child (WHAT employee ...)...) 
Next the scope critic functions evaluate the different scoplngs; only three of the 
critics are relevant. One critic considers the left-right node ordering and 
prefers the first scoplng because it comes closer to the surface form, One critic 
prefers scoplngs in which WH outecopes an adjacent exlstential; it too upgrades 
the score of the first and downgrades the score of the second. The other critic 
knows that default existential quantlflers need the narrowest possible scope; it 
too selects the first. 
100 B. GROSZ et al. 
SUMMARY 
Because of the modularlzation in D~ALOGIC, changes in one part of the system 
reverberate very little in other components. Changes in the constraints imposed 
on the phrase-structure rules in the grammar have no effect on any other part of 
the system. A change in a rule itself necessitates a change in the corresponding 
translator, but the basic semantic functions do not need to be revised. 
Similarly, a change in the logical form or in the data structures within which it 
is implemented requires a corresponding change in the basic semantic functions, 
but not in the grammar or translators. 
In addition to extending DIALOGIC as mentioned in the foregoing sections, we are 
also investigating possible revisions of the translation phase (as currently 
performed by the translators and basic semantic functions) to allow translation 
into loglcal form to be specified declaratlvely. In this new approach 
\[Rosenscheln and Shleber (to appear)\[, loglcal types are associated with the 
phrasal categories, and the translation of a phrase is synthesized from the 
translations of its immediate constituents according to a local rule, which 
typically involves functional application. 

REFERENCES 

Dowty, D. R., Wall, R.E., and Peters, S. \[1981\] Introduction to Montague 
Semantics (D.Reidel Publishing Co., Dordrecht, Holland, 1981). 

Gazdar, G. \[to appear\]. "Phrase Structure Grammar," in P. Jacobson and 
G. K. Pullum, eds. On the Nature of Syntactic Representation. 

Hendrix, G.G. \[1978\[ "Semantic Aspects of Translation," in Understanding 
Language, D. E. Walker, ed., pp. 193-226 (Elsevier, New York, New York, 1978). 

Hobbs, J. \[1976\[ 'Pronoun Resolution," Research Report 76-1, Department of 
Computer Sciences, City College, City University of New York, New York, New 
York (August, 1976). 

Kaplan, R. M. and Bresnan, Jo W. \[to appear\] "Lexical-Functlonal Grammar: A 
Formal System for Grammatical Representation," Occasional Paper 13, Center for 
Cognitive Science, Massachusetts Institute of Technology, Cambridge, 
Massachusetts. 

Konolige, K. G. \[1979\[ "A Framework for a Portable Natural Language Interface to 
Large Data Bases," Technical Note 197, Artificial Intelligence Center, SRI 
International, Menlo Park, California (October, 1979). 

Landsbergen, S. P. J. \[1976\] "Syntax and Formal Semantics of English in PNLIQAI," 
in COLING 76~Preprints of th___ee 6thlnternational Conference on Computatlonal 
Linguistics, No. 21, Ottawa, Ontario, Canada (June 28-July 2, 1976). 

Moore, R. C. \[1981\] "Problems in Logical Form," in Proceedings of the 19th Annual 
Meetlngof the Association for Computational Linguistics, pp. 117-124, The 
Association for Computational Linguistics, SRI International, Menlo Park, 
California (July 1981). 

Robinson, Ann E. \[1980\] "Interpreting Natural-Language Utterances in Dialogs 
About Tasks," Technical Note 210, Artificial Intelligence Center, SRI 
International, Menlo Park, California (March 1980). 

Robinson, J. J. \[1982\] "DIAGRAM: A Grammar for Dialogues," Communications of the 
ACM, Vol. 25, No. I, ppo 27-47 (January 1982). 

Rosenscheln, S. J. and Shieber, S.M. \[To appear\] "Translating English into 
Logical Form," in Proceedlngs of 20th Annual Meetin~ of Association for 
Computational Linguistics Conference, Toronto, Canada (June 1982). 

Tienari, M. \[1980\] "On the Definition of an Attribute Grammar," in Nell D. Jones, 
ed., Semantic-Directed Compiler Generation. Proceedings of a Workshop, Aarhus, 
Denmark (January, 1980). G. Goos and J. Hartmanis, eds., Lecture Notes in 
Computer Science Vol. 94, pp. 408-414, (Sprlnger-Verlag, Berlln-Heldelberg-New 
York, 1980). 

Walker, D.E. (ed.) \[1978\] Understandln~ Spoken Languag____~e (Elsevier North- 
Holland, New York, New York~ 1978). 
