"YREB" ORDER LANGUAGES: AN EXPERZM~NTAL LEXICON BASED PARS~ 
Cristlano Castelfranoni, Domenico Parisi, 01ivtero Stock 
Zstituto di Pstoologia, Oonsiglio Nazionale delle Rioex~he 
Via dei Monti Tiburtini, 509, O0157-Roma 
Most of the work on natural language understanding (NLU) 
has been done on English. E~ish is a language with relative- 
ly rigid word order, a characteristic that has influenced all 
NLU systems proposed so far. These systems have dedicated only 
minor attention to problems that are of major importance in 
lemguagee with a freer ordering like Italian. Work on parsing 
partially un~ram~atioal sentences in rigid l~s (Hayes 
and Mo~radtan, 1980l Charniak, 1981) bears some similarity to 
work on p~rsing of "free" order language. Tn both oases an 
exclusively top-down model seems inappropriate. For instance, 
when considering an incomplete sentence it may still be advis. 
able to proceed in building some representation. The result- 
tng structure will then be inserted within a larger oo~xitive 
structure. The sane bottom-up way of proceed£ng must necessa- 
rily be part of an NLU eyste~ for "free" order languages. 
An ATN (Woods, 1970) type systaw, like the one developed 
for Ztalian (Cappelli etal., 1978), shows definite l~nitat- 
ions even when it is furnished with heuristics for strate~ 
selection (Ferrari and 8took, 1980) based on adaptation to a 
coherent text. Other ideas such as passing infox~nation in the 
case of failure (Weisohedel and Black, 1980) or relaxing con- 
ditions on aras in certain oirotvastances (Kwasny and Sondhei- 
met, 1979) do not seem to be satisfying solutions for our 
px~bl~. 
- 65 - 
In other hand, we do not agree that syntax be given a 
subordinate role, as Schank's theoretical approach implies 
(though in Schank and Riesbeck's parser, ELI (Riesbeck and 
Schank, 1978), syntax has a more important role than would be 
expected). In any event, without enough syntax it becomes hard 
to analyze complex sentences and to explain a number of psycho- 
linguistic phenomena. 
2. WEDNESDAY, the system proposed here, is the core of 
an extended mechanism we are developing and implementing in 
LISP (flg.1) (see Pariei and Castelfranchi, in press). It is 
an analyzer with semantic output based on word interpretation. 
The semantic information brought in by each word is progress- 
ively connected to get at the sentence's meaning according to 
syntactic constraints and expectations. Syntax is a set of 
instructions directly concerned with assembling semantic units. 
LEXI C ON 1 STRUCTURED 
SYNTACTIC WORKING 
SPACE 
PROGRESSIVE SEmANTiC 
REPRESENTATION 
Pig. 1 
What is characteristic of WEDNESDAY is that syntax is 
not a separate component, but is distributed throughout the 
lexicon. Lexical entries are composed of a declarative part 
and a procedural part. The procedural part is made up of syn- 
tactic instructions designed to assemble the declarative (se- 
mantic) part of the entry with the declarative parts of the 
other words in the sentence. The syntactic assembling process 
- 66 - 
allows the construction of the sentence's semantic network to 
be carried out progressively and so it can also operate on in- 
complete sentences. Furthermore, the system's output has a 
format which is homogenous to that of the knowledge base 
(encyclopedia). This means that as the sentence analysis pro- 
ceeds, it is natural to have the knowledge base controlling 
the assembly process top-down, in cooperation with the lexioal- 
-syntactic analysis. Clauses in multl-clause sentences are 
conceived as restricted search spaces for the assembly instruc~ 
tions carried by the words, and the same is true for noun 
spaces (noun phrases). Part-of-speech word categories are 
interpreted as procedures for opening and closing spaces as a 
function of context. Therefore, part of the process is depend- 
ent on a recursive (push-down storage) mechanism. On the other 
hand there is a semantic sentence memory that is transparent 
to the closed-level system. 
Being lexically based, WEDNESDAY can deal in a natural 
way with idiosyncrasies typical of many words. It can also 
deal with flexible idioms, i.e. idioms that can vary in morpho- 
logy, word order, syntactic construction, semantic additions, 
and synonyms. Their reoo~ition is governed by the individual 
lexical entries and takes place at +.he assembling level. 
Word disambiguation is treated in a non-deterministic 
way. Syntactically based disambiguation (e.g. the word J~) 
attempts each reading of an ambiguous word in turn and cancels 
inappropriate readings by testing them against syntactic well- 
- formednese criteria. Encyclopaedia based disembiguation 
(e.g. the word b_~) works through the activation of encyclo- 
paedic nodes by the words in the sentence. It selects that 
reading of an ambiguous word which activates the node "better" 
connected to the other activated nodes. It should be noted 
that, our approach, in comparison to Small°s work on Word 
Expert Parsing (Small, 1980), is characterized by (a) non- 
-determinism, (b) a more systematic recourse to syntax, (c) a 
separation of syntactically based word disambiguation from 
encyclopaedia based word disembiguation. 
- 67 - 
_J 

References 

Capelli A., FezTari G., Moretti L., Prodanof I., Stock O. 
"An ATN Parser for Italians Some Experiments". In 
I~oe. of the SeTenth International Conference on Com- 
putational Lin~istics. Bergen 1978 COLING

Charniak Eo "Six topics in Search of a Parser". Proc~ of th~ 
7th IJCAI. Vancouver, 1981 

Ferrari G., Stock O. "Strategy Selection of an ATN Syntactic 
Parser". Proo. of the 18th Meetin~ of the Association 
for Computational Linguistics. Philadelphia, 1980. 

Hayes Pc, Mou~adian Go "Flexible Parsing". I~oo~ of the 18th 
Meetin~ of the Association for Computational Lin~ulet£¢so 
Philadelphia, 1980 

Kwasny 8o and Sondheimer N. "Ungrsmmatioality and Ex~rn- 
~ammaticality in Natural Lant~age Understanding Syst- 
ems"° Proct of the 17th Annual Meeti~ of the ACL~ 
San DieEo, 1979 

Parisi D° • Castelfranohi C° La comprensione del lin~ussJ~io° 
Turin, Boringhieri in press, 

Riesbeck C° and Schank R° "Comprehension by Computer! Expect- 
ation Based Analysis of Sentences in Context"° Research 
Report 78. Yale Univ., 1978o 

Small S. "Word Expert Parsing: A Theor~ of Distributed Word- 
Based Natural L~e Understandi~". TR-954! Dept. 
of Computer Science, University of Maryland, 1980o 

Weischedel R. Rnd Black J. "If the Parser Fails", l~?c~ of 
the 18th Annual Meetin~ of the Association for Comput- 
ational Lin~tistics. Philadelphia, 1980. 

Woods W. "Transition network Grs~nare for Natural Language 
Analysis". Comm. of the ACM, Vol. 13, no. 10, 1970 
