A SEMANTIC ANALYSER OF NATURAL ITALIAN SENTENCES 
M. Del Canto, F. Fusconl, L. Stringa 
Elettrohica San Giorgio - ELSAG S.p.A. 
Via Hermada, 6 - 16154 Genova - Italy 
This paper presents the analyser for type-written Ital- 
ian sentences used in the LISA system. 
This system has been developed in ELSAG within the 
researches on Natt~rel Language Processing aimed at making the 
dialogue with the user easier and more graceful. 
The analyser was designed to accept input sentences 
without any constraint on how they are formed (e.g. in active 
or passive form, with a variable number and position of compl- 
ements, etc.), and to accept such irregularities as ellipsis, 
idioms and small grammatical errors. The present version works 
for simple sentences that are introduced one at a time. 
The output of the system is an internal conceptual re- 
presentation that we defined, according to Schank, "concept- 
ualization". It represents the meaning of the sentence in a 
non-ambiguous way and we can define it as an organized set of 
predioate~argument lists, each of which can have some modif- 
iers. One feature of the conceptualization format is its 
generality, which permits to easily adapt the system to any 
new application. 
The analyser makes use of a dictionary that includes a 
lexicon, a set of syntactic descriptions and a set of semantic 
descriptions. 
- 82 - 
The lexicon, in the version presently implemented, con- 
tains about 2500 words that permit the user to introduce 
sentences (statements or questions or answers) con erning 
people characteristics like age, profession, health and 
their actions and relations with the external world, like to 
go, to come, to travel, to give, to take (TRANSFER), to know 
(~OrhEDGE), to speak, to say (COm~JNICATION), to own, to 
contain (RELATION), etc. The lexicon is broken down into 
several sections in order to optimize the memory usage aocord- 
ing to an exhaustive study of a root-ending representation 
of italian words. 
The syntactic descriptions are directly related to 
lexioal entries and contain attributes like gender and number 
of the nouns or like mood and tense of the verbs. 
The semantic descriptions represent the conceptual 
entities related to lexioal entries; they mainly contain 
informations about the consistency between predicates (named 
"operators") and their ar~mente and provide a classification 
of the conceptual entities. 
The analysis process works in three main steps: the 
lexical analysis, a bottom-up reoo~lition of "syntactic and 
semantic Groups" and a top-down insertion of these Groups in 
the semantic structure (conceptualization). 
The lexical analysis recognizes the single words in 
the input string and compares them with the lexicon compon- 
ents in order to recover ell possible interpretations. The 
sequence of the syntactic and semantic descriptions obtained 
forms the output of this step. Unknown terms are neglected 
at this point, but the dialogue oontrolling module is inform- 
ed about that. 
All the characteristics derived from the input ere 
analysed in the next-step end organized in Groups based on 
the most meaningful terms (Verbs, Nouns, Adverbs). Each 
- 83 - 
particular t~pe of construction, corresponding to a Group, is 
recognized by an independent "specialist subgremmar", which 
selects only the relevant portion of the inputo 
The last step of the analysis relates the elements in 
the previously built Groups with the roles in the conceptual 
structure (conceptualization). 
This structure in its main peat ("nucleus") is determin- 
ed by the verbs in the Verb-Group; some other roles, su88ested 
by the "expectations" in the Noun-Groups and in the Adverb- 
Groups can be associated to the nucleus through RELATION oper- 
ators. The main orite~on that guides the association between 
Groups ~d conceptualization roles is based on expectations 
pertly related to the verb and partly pre-established. 
A semantic approach characterizes this method of analys- 
is: in fact a syntactic representation of the input is not 
attempted and the semantic descriptions of concepts are direct- 
ly accessible. 
This makes it easier to ~£de the analysis towards the 
internal representation, at the same time reducin 8 the number 
of the alternatives generated and consequently also the prob- 
lem of dealing with ambiguities. 
A further fundamental feattLre is the co-operation bet~ 
ween bottom-up and top-down techniques in the organization of 
the input in Groups and in the filling of the output structu- 
re. The first of them is best suited to deal with grem~atic~ 
deviations, incorrect infle~ons and fragmentary utter~uces, 
also because of the manifold scanning and non rigorous constr- 
a£nts on the word positions. 
At this level pattern-matc~Lng mechanisms can also be 
used to handle idioms and others fixed phrases; the reco~Lt- 
ion of these forms, may need both syntactic and semantic de- 
scriptions. 
- 84 - 
The top-down approach in the last step is suitable to 
make the conceptualization a stande~d, m~u~eable, structure, 
which could be easily adapted, if necessary, to spectfio 
ne ed8. 
This method of analysis proved %o be a suitable one 
within a system capable of conducting a dialogue through non- 
-oonstreined input sentences, related %o a general and flexib- 
le knowledge representation. 
- 85 - 
