A Two Level Modular System for Natural Language Understanding

1. INTRODUCriON

The conception of modular and flexible architec tures has been proposed recently in the develo£ ment of natural language understanding systems (2]. Basic exigencies of unification, flexibi­ lity, and intercommunication have motivated the deve lopment of the DONAU (Domain Oriented NAtu­ ral language Unders~ng) system, which ~a~ been successfully experimented on an UNIVAC 1110 at the Milan Polytechnic Artificial Intel­

ligence Project (.3].

The original contribution of DONAU consists in the adoption of a ~ ~modular architec­ tur~ for a system able of understanding natural language.

In the development of the DONAU two level arch! tecture, the morphologic and syntactic analysis has been designed and implemented on the base of the PIAF system which has been conceived, ~ veloped, and successfully tested in the past at the IMAG Laboratory of the University ot G:t·e noble (France) [4]. -

Two versions of DONAU have been already comple­ tely implemented and experimented: a first one, oriented on robotics, for the programming and control of the SUPERSIGMA robot developed and operating at the Milan Polytechn~c Artifi~ial Intelligence Project; a second one, oriented on data bases, for the inquiry of a relational data base about the oil resources in the world proposed by the International Institute for Ap­ plied System Analysis, Laxenburg, Austria. A third version, centered on program synthes.i.s,is presently being designed [1].

2 · 'IHE TWO IEVEL MODULAR STRUCTURE

This section illustrates the modular structure of DONAU and the adopted design criteria. The main criterion which has guided the conception of DONAU has been that one of two level modu­ larity.

The first level, called horizontal level,re­ flects the distinction between the different conceptual activities concerning the natural language understanding process. Eachone of th~ se activities (i.e., morphologic and syntactic ahd.lysis, semantic analysis, information ex­ traction, and legality control) is associated with a bimodular segment which is precisely defined by its input-output characteristics. The second level of the modular architecture, called vertical ~, is related with the di­ stinction of the different nature of the two modules which compose the bimodular segment. The first module, called !£e module, embeds the domain-independent, general-purpose algo­ rithm associated with the conceptual activity proper of the segment. The second module, cal­ led bottom module, consists of a framework con­ taining the~n-dependent, special-purpose information which directly relates with the particular semantic domain in which a version of the DONAU system is conceived to be operat­ ing.

An advantage of this two level architecture lies in an efficient balance between the charac teristic of generality, required by a natural­ language understandin~ system, and the exi~ency of efficiency, implied by an actuallv u~eful system.

Another feature is the simplicity in design­ ing different versionsof the same system, each­ one devoted to a particular instance of a se­ mantic domain. Each new version can be obtaine~ in fact, by simply changing, in each segment, only the bottom module, i.e., the domain orie~ ted information. A further quality is the poss~ bility of having incrementality in the defini­ tion and set up of the system, i.e., the capa­ bility of modifying and increasinq the content of the specification information on the base of the requirements and suggestions which the experimentation of the system might propose.

The fundamental assumption on which the DONAU project is based is that one of artificial understanding of natural language. We ~laim that the understanding of natural language, ha­ ving as ultimate goal the communication between the man and an artificial system, does not re­ quire the capability of following all the pe­ culiarities and the nuances of the possibly dif­ ferent ways of expressing an operative informa­ tion in natural language. The goal of the under­ standing process is to obtain just that kernel of information which is necessary to activate in the desired and correct way the artificial system which is the target of the man-machine interaction and which can perform only a bound­ ed, usuallyquite small, number of activities. This approach to natural language understand­ ing can be useful, in our opinion, for avoiding some of the linguistic difficulties encountered in early research works. Moreover, this stand­ point allows to largely extend the capabilities of the system in accepting a great variety of free natural language sentences, still avoiding the explosion of the complexity of the parsing algorithm consequent to an increased complexity of the input sentences which might actually be encountered in real applications.

3. ~-DERSTANDIN� ~ SENTENCES

The first step of the parsing consist~ of the morphological and syntactic analysis. The clas­ sical methods of ~ state recognizers and dependence grammars @] are adopted for this activity. The result of the syntactic a11alysis, constituted by a (small) set of dependence trees Q], is then supplied ~s input of the semantic analyser.

The role of the semantic analysis is to discri­ minate, among the different dependence trees, that one which is consistent with the semantic description of the natural language supplied to the system.The specification information of this segment is constituted by a complex frame­ work called model-list network.'l'his network is composed of an hierarchical organization of modellistS.The model-list:sare lists of symbols representing morphological types and semantic types, which denote the categories of word~ and elementary constructs considered as relevant ones for defining the semantics of the natural language in a particular application domain. Each model-list is associated to a semantic ty­ pe and constitute therefore a transformation r~ le. A dependence . tree is discriminated by the semantic analyser if to its nonterminal node s model-lists can be associated in such a way to ~ satisfy an appropriate set of matching condit­ ions (3]. The output of the semantic analyser is represented by a semantic tree embedding all the relevant information so far obtained in the parsing.

The third segment of DONAU is the information extractor. Its role is to individuate within the semantic tree the operative significant content which is sufficient and not redundant for ex­ pressing the meaning of the original natural language sentence. The elimination of the super­ fluous information is performed by a covering algorithm on the base of the content diagram U] defined on the lexicon. The result of the whole parsing activity is an executable formal statement which can directly be accepted and executed by the interacting artificial system.

Finally, the legality control consists in the evaluation of the proposed operative content within a legality ~ which represents the no~ linguistic context of the artificial system con nected with DONAU.

4. EXPERIMENTAL RESULTS AND CONCLUSIONS

Two versions of DONAU have been developed.

RB-DONAU, centered on the semantic domain of ro­ botics, is running, at present time, on an UNIVAC 1110 computer and is programmed in 1108LISP. The software architecture of the system r~ fleets quite closely the conceptual schema de­ scribed in section ~- The prograrr. occupies about 25 Kwords memory and the specification information, relative to a lexicon of about 200 words, 5 kwords. The comprehension of simple sentences of 5 to 10 words requires about 1 second (with compiled LISP programs).

DB-DONAU, centered on the domain of data base inquiry [11, has been developed with the preci­ se aim of verifying the actual possibility of changing the semantic domain along the lines su~ gested by the vertical modularity of the system. The experimental activity done has shown a good flexibility of the model proposed for achieving the modularity and incrementality goals initial­ ly stated for the project.

Several projects are presently developing around the DONAU system. In particular, the following goals are considered as relevant ones :

• to improve the diagnostic capabilities of oo­ NAU and to develop an high-level tes ting and tracing subsystem;
• to define and implement an experience subsy­ stem, which can guide the user's tuning and extension activities on the base of the past use of the system;
• to integrate the two steps of syntactic and semantic analysis as parallel coroutines for increasing the efficiency of the parser.

APPENDIX

In this appendix we present an example of na­ tural language understanding on a simple sen­ tence :

ELENCA I DEPOSIT! DEL KUWAIT AVENT! RISERVA DI PETROLIO MAGGIORE DI 1500 TONAL 1.9.1978 (list the deposits of Kuwait having an oil re­ serve greater than 1500 ton on 1.9.1978). Two different dependence trees are obtained: ( (ELENCA verb)~( (DEPOSIT! sub) (I art)~ ((KUWAIT sub) (del prep)'ljl;) ((AVENT! verb)lt ( (RISERVA subl'*-( (PETROLIO sub) (DJ prep~) (MAGGIORE adj) ( ( 1500 num) (DI prep) .. (TON sub)) ((AL prep)il' (1.9.1978 date))))))

( (EIENCA verb)~( (DEPOSIT! sub) (I art)'¥ ((KUWAIT sub) (DEL prepJ1')) ( (RISERVA sub) (AVENT! \rerb)'ltf( (PETROLIO sub) (DI prep)t.J (MAGGIORE adj ) ( (1500 num) (DI prep):l (TON sub)) ((AL prep}11((1.9.1978 date))))

The notation (x y · z) is used,where x is the root of the tree, y are· the left subtrees and z are the right subtrees.

Only the first tree is discriminated by the se­ mantic analyzer; the corresponding semantic tree is :

((sentence: (ELENCA verb))~( (object: (DEPOSIT! sub)) (:r art) ... ( (spec: (KUWAIT sub)) (DEL prepl*l ( (compl : (AVENT! verb) l-'-tl·< (compl-spec: (RISERVA sub))*( (spec: (PETROLIO sub)) (DI prep~) (MAGGIORE adj) ((spec: ( 1500 num)) (DI prep)* (TCN sub)) ((da~: (AL prep)):JI(1.9.1978 date))))))

The resulting information tree is then:

((sentence: (EIENCl. verb) )'!t( (9?ject: (DEPOSIT! sub) Ji(spec: (KUWAIT sub)) ( (compl: ) .. ( (compl-spec: (RISERVA sub) )-11\(MAGGIORE adj) ((spec: (1500 · num) >*(TON sub)) ((date :J* (1.9.197R nate)))))) --

ACKNOWLEDGMENTS

We are grateful to the researchers of the !MAG

Laboratory (University of Grenoble, France) which have cooperated to this project and to the students of the Politecnico di Milano for their contributions in the experimental work.