Deep Sentence Understanding in a Restricted Domain* 
Pierre Zweigenbaum and Marc Cavazza 
DIAM -- INSERM U.194, 91 bd. de l'H6pital, 75634 Paris Cedex 13, France 
e-mail: zweig@frsim51 (bitnet) 
Abstract 
We present here the current prototype of the 
text understanding system HELENE. The ob- 
jective of this system is to achieve a deep un- 
derstanding of small reports dealing with a re- 
stricted domain. Sentence understanding builds 
a model of the state of the world described, 
through the application of several knowledge 
modules: (i) LFG parsing, (ii) syntactic disam- 
biguation based on lexical entry semantic com- 
ponents, (iii) assembly of semantic components 
and instantiation of domain entities, and (iv) 
construction of a world model through activa- 
tion of common sense and domain knowledge. 
1 Introduction 
We present here the current prototype of the 
text understanding system HI~LI~NE. The ob- 
jective of this system is to achieve a deep un- 
derstanding of small reports dealing with a re- 
stricted domain (here, patient discharge sum- 
maries in a medical specialty). This means that 
H~LJ~NE should rely on an extensive description 
of all types of required knowledge. This implies 
of course a deep domain knowledge base, includ- 
ing the needed common sense knowledge. 
Precise understanding must not only rely on 
complete domain knowledge, but also on enough 
syntactic information. This is the reason why 
H~LI~NE includes a full syntactic module, whose 
task is to provide the semantic construction 
module with the (deep) structures of sentences. 
One problem with syntactic processing is that it 
gives rise to numerous ambiguities. These am- 
biguities are filtered on semantic grounds by a 
disambiguation module that does not build any 
semantic representation. 
*supported by AIM project 1003 and PRC Commu- 
nication Homme-Machine, PSle Langage Naturel. 
Semantic construction is concerned with the 
recognition of domain entities that can be ex- 
pressed by word groups. We thus had to adopt a 
lexical semantics approach compatible with de- 
scriptions. Domain entities, once instantiated, 
provide the basis on which a model of the cur- 
rent state of the world (here, the patient state) 
is built. The same lexical semantic in{brmation 
is used both to help syntactic processing, and in 
a more extensive way to access domain models 
in order to build semantic representations. 
The prototype includes the following main 
modules: 
The syntactic module implements the Lexi- 
cal Functional Grammar formalism \[7\]. The 
parser builds c-structure and f-structure 
bottom-up in parallel on a chart, so that f- 
structure validity can constrain c--structure 
construction. 
Ambiguous attachments are submitted to 
evaluation and ranking by the disambigua- 
tion module. This module applies a set of 
general heuristic rules that operate on the 
semantic definition of the LFG predicates. 
Semantic construction relies on dynamic 
domain models that integrate common 
sense. LFG predicates are characterized by 
semantic components that point to parts of 
the knowledge base. 
The prototype runs in Common Lisp (VAX 
Lisp) and K, a proprietary language embedded 
in Common Lisp. The remaining sections de- 
scribe these modules in more detail. 
2 Parsing with a Lexical- 
Functional Grammar 
We chose to implement the LFG fl:amework 
for several reasons. Being a linguistic theory, 
82 1 
it should provide better foundations for prin- 
cipled syntactic coverage. A formalism with a 
context-free backbone was easier to implement. 
Furthermore, LFG extracts from a sentence a 
predicate-argument structure which consitutes 
a good starting point for semantic processing. 
Our implementation of LFG does not include 
yet a schema for long-distance dependencies (or 
functional uncertainty) and coordination. It al- 
lows cyclic f-structures. 
Our parser uses a chart to build both c- 
structure and f-structure. Incomplete and com- 
plete constituents are represented by active and 
inactive cs-edges, while incomplete and com- 
plete fstructures are placed on active and in- 
active fs-edges. The parsing strategy is bottom- 
up, left to right (left corner). Top-down pars- 
ing is also available, as well as right to left 
parsing. LFG schemas are evaluated as soon 
as possible. Equational (construction) schemas 
are evaluated when encountered, and constraint 
schemas (existential, equational and negation of 
those) are kept on fs-edges until they can be 
evaluated. When fs-edges are combined, remain- 
ing constraints are propagated to the resulting 
fs-edge. Each new active f-structure is tested 
for consistency and coherence. Furthermore, the 
value of a closed function is tested for complete- 
ness (this should be revised if a scheme for long- 
distance dependencies is implemented). When 
a constraint is violated, its fs-edge is flagged as 
invalid. 
Grammar rules are described as regular ex- 
pressions which are compiled into (reversible) 
transition networks. Each arc of those networks 
is labelled with a category and a disjunction of 
conjunctions of schemas. A model of hierarchi- 
cal lexical entry representation has been devel- 
oped, with data-driven lexical rules. It is not 
currently coupled to the parser, and will not 
be presented here. The prototype uses a sim- 
ple word list with what would be the result of 
this model as lexical entries. 
The prototype uses a small French gram- 
mar that contains 14 networks, equivalent to 90 
rules. It was assembled by borrowing from the 
literature \[3,10\] and by completing with gram- 
mar manuals and real text confrontation. It has 
the particularity of building cyclic f-structures 
for constructions where a head plays a role inside 
an adjunct. This is how we process attributive 
adjectives, participial phrases, and (in a very 
limited way) relative phrases. 
3 Semantic rules for Syntactic 
Disambiguation 
Structural ambiguity is ubiquitous in our tar- 
get texts, since they contain descriptions that of- 
ten make use of series of prepositional phrases to 
qualify a noun. We have then decided to submit 
ambiguous attachments to semantic approval 
and ranking before building complete parses. 
An ultimate test of semantic validity would 
consist in comparing complete semantic repre- 
sentations built for each attachment proposal 
\[1\]. However, such a method is too expensive to 
allow systematic application. Our system im- 
plements a more tractable approach that gen- 
eralizes selectional restrictions (or preferences). 
Evaluation is performed by executing a set of 
heuristic positive and negative rules that vote 
for or against each proposal. Rule conditions 
embody criteria that refer to the semantic com- 
ponents (see below) of the predicates to be at- 
tached, and include the notion of isotopy \[8\]. 
They apply not only to predicate-argument se- 
lection, but also to predicate-adjunct combina- 
tion. 
4 Semantic Construction 
Semantic processing of a sentence results in 
the activation of a relevant body of domain 
knowledge with related inferences within the 
knowledge base. 
Domain knowledge (here concerning a single 
disease: thyroid cancer) is embedded in a model 
\[6,4\] describing domain objects, actions oper- 
ating on them and specific processes involving 
these objects. Such a model is thus a dynamic 
causal model rather than a memory structure 
devoted to object and event integration \[9\]. It 
is analogous to deep-knowledge models used in 
modern expert systems \[2\]. 
Domain objects are represented in a frame- 
like formalism. Actions and operative aspects of 
processes are described as production rules sim- 
ulating a distributed parallel activation \[4\]. The 
whole model corresponds to a dynamic, data- 
driven environment. 
Some domain concepts specifically represent 
states, relationships between objects, or state 
2 83 
transitions. They can be triggered by their 
occurrence as word meanings in the sentence. 
Implicit occurrence of these concepts may also 
be recognized by observing the evolution of the 
model. In this case default procedures create the 
corresponding concepts inside the model just as 
if these elements were explicitely stated in the 
proposition. These concepts subsume important 
situations in the model and translate them into 
a higher description level, thus allowing output 
to the user for a trace of correct understanding. 
No deep understanding would be possible 
without a treatment, even partial, of common 
sense, which in this application is concerned 
mainly with part-whole relationships \[5\], reason- 
ing about transitions and change, and elemen- 
tary physical actions (e.g., removing, touching). 
Default knowledge on actions, roles and refer- 
ence (e.g., used in the resolution of pragmatic 
anaphora) are associated to the common sense 
module. 
ColImlon sense mechanisms are incorporated 
as production systems similar to those describ- 
ing other active elements of the model, and can 
thus recombine freely with them in order to com- 
plete or modify existing representations. 
Domain representations are built from the as- 
sembly of lexical contents along the syntactic 
structure of the proposition. Words contain se- 
mantic, components \[8\], which are markers refer- 
ring to elements of the knowledge base or prop- 
erties of these representations. The existence of 
explicit colnmon sense concepts in the knowl- 
edge base makes it possible to decompose ho- 
mogeneously technical and ordinary words. 
The lexical contents are assembled by heuris- 
tic rules to form candidate domain objects which 
are recognized as instances of prototypes in the 
representation. Lexical content itself is loosely 
structured; the association of the components is 
completed according to their type (which is de- 
rived t rom the type of entity they refer to) and 
the dependency (predicate-argument, predicate- 
adjunct) relations between the lexemes that con- 
tain such components, as provided by the LFG. 
As such elements of representation are recog- 
nized by the model, the reactive environment is 
triggered and interprets data until new informa- 
tion is analyzed. 
The prototype currently runs on a small set of 
30 sentences taken from patient discharge sum- 
maries. These sentences were selected for the 
linguistic issues they illustrate and the domain 
inferences they trigger. A fully compiled version 
of the program running on a VAX 8810 processes 
a sentence in an average 12 sec. CPU time. 

References 

Birnbaum, L. (1985). Lexical ambiguity as a 
touchstone for theories of language analysis. IJ- 
CAI 9, Los Angeles, 815-820. 

Chandrasekaran, B. and Mittal, S. (1!183). On 
deep versus compiled knowledge approaches to 
medical diagnosis. Int J Man-Machine S¢udies 
19:425-436. 

Orimshaw, J. (1982). On the lexical represen- 
tation of Romance reflexive clitics, in Bresnan 
(ed.), The Mental Representation of Grammat- 
ical Relations, Cambridge, Mass., MIT Press. 

Holland, J.H., Holyoak, K.J., Nisbet~, R.E., 
Thagard, P.R. (1986). Induction: Processes 
of Inference, Learning, and Discovery. Cam- 
bridge: MIT Press. 

Iris, M.A., Litowitz, B.E., and Evens, M. 
(1988). Problems of the part-whole relation. In 
Evens, M.E., ed. Rela$ional Models of the le~i- 
con. Cambridge University Press. 

Johson-Laird, P.N. (1983). Mental Models, 
Cambridge University Press. 

Kaplan, R.M., and Bresnan, J. (1982). Lexical- 
Functional Grammar: A Formal System for 
Grammatical Representation, in Bresnan (ed.), 
The Mental Representation of Grammatical Re- 
lations, Cambridge, Mass., MIT Press. 

Rastier, F. (1987). Sdmantique Interprdtative. 
Paris: Presses Universitaires de France. 

Schank, R.C., and Abetson, R.P. (1977). 
Scripts, Plans, Goals and Understanding. 
Lawrence Erlbaum Associates. 

Waite, J.3. (1986). Orammaire Lexlcale- 
Fonctionnelle -- Dislocation, inversion stylis- 
tique et constructions impersonnelles. PhD 
Thesis, University of Montreal. 
