- Email: [email protected]
An Intelligent Interface for Accessing a Technical Data Base
C()I)lTi ~ h{
©
II:,'. C \1;,11 · \1;" h illl' S , , {clll' ,
\ ' art'st:', llah, I ~1 ~"Fl
AN INTELLIGENT INTERFACE FOR ACCESSING A TECHNICAL DATA BASE F. Bastin*, S. Capobianchi*, Sylviane Carlesso*, G. Mancini *, T. Barbas**, A. D. Koletsos** G. Guida***, M. Somalvico***, and C. Tasso**** *CEC,
JRe
Ispra E ,/ a /!lis/lll1ml, Ispra , 1111 /\'
*o,oC,t.' ,c;, G m lll Ho/fin.1 *"" M,ho, P u/y/n hll i, Arlilirifl/ IlIlrlligma Proj,'(I, ,\li/all o, Ila/y **** / slillllo di ,\lall' l/wliC!l . 11I1() l'IIwlim
aimed at the correct , and effective way . European Reliability Data System - ERDS) in a friendly , The interface incl u des both natural language understanding and expert problem solving capabil i ties and features a modular archit e cture devoted to cope with the exigencies In the paper a brief introduction to ERDS of both technical and non - technical users. is first presented and the problems related to its use are analysed. The requirements of the users of ERDS are then investigated , and the basic archi tecture of an intelligent interface is illustrated. The design c r iteria of the two main modules of the i nterface , name ly the natural language query interface and the expert interface , are later discussed. At last , the present state of t h e project is reported and promising directi o ns for future research are out1 ined . Abstract.
The
paper
deals
with
the
I' Sisll'lll istira. C llil'nslh i di Cdilll', I. 'dilll'. Ila/\'
design
of
an
inte l ligent
int.erface
assisting e n d- users in accessing a large and complex technical data base (namely,
Key words.
I n telligent interfaces ,
natural language processing ,
knowledge-based sys-
tems , data bases .
II~TRODUCTION
One of the major recently emerging fields of ap plication of ar ti ficial intelligence is consti tuted by data base management systems (King , 1983 ; Reiter , Gallaire , King , Myl o pou l os, and 'webber , In this area many studies are currently 1983). carried out to implement systems known as intel1 i gent or expert i n'cer faces (Cohen and Pe rrau 1 t, 1982; Kap 1 an , 1983; Gui da and Tassu , 1983 ; Bra j nik, Guida , and Tasso , 1984; Reichman- Adar , 1984) , aimed at assisting end - users in accessing i nformation stored in a data base in a friend l y , correct, and ef f ective way. Generally , intelligent interfaces encompass c apa bilities of both nat ural language front - ends and expert systems , in order to bridge the linguistic gap between non - technical user and H.e formal query l anguage necessary for accesslng the data base, and to support the users in the complex task of effectively and correctly using the information stored in the data base. CQping wi th such problems becomes more and more
the European Reliability Data System (ERDS) , de velo ped at the Ispra JRC Establishment in t h e frame of the Reactor Safety programme. Th e use of such a 1arge system , composed of many intercon nected subsystems , requires precise knowledge in several domains in order to : - state problems correctly and manage their solution; - i den ti fy whi ch data are needed in the prob 1 eo. solving process ; - navigate inside the system and its multiple data bases to retrieve the desired information ; - validate the extracted data and possibly process them through speclal purpose application p r o grams . In such a case , natural language processing and expert system technologies can effecti ve ly con tribute in bridging the gap between end- users and the data base. This paper is devQted to illustrate the results of a feasibility study concerning an intell i gent in te r face to the ERDS data base recently carried out at Ispra JRC Establishment.
important , and almost mandatory, when one deals with complex technical data bases accessed by several classes of end - users. This is the case of
167
16H
F. Bastin DESCRIPTION OF ERDS
ERDS (European Rel iabi 1 i ty Data System) has been developed at the Ispra JRC Establishment over t h e past years in order to collect, manage , and dis seminate information about Light Water Reactors (LWR) located in t h e European Communi ty , and to a ccumul ate reliability data for probabilistic risk assessments (Mane-ini , Amesz , Capobianc h i , and 8astiani ni, 1982; Capob i anchi , 1983) . The original scope of ERDS h as been enlarged in some cases to in c lude non-Eur opean reactors (e . g. , the LER system o f NRC, U. S . A. ) and a 1 so reacto rs ot her t hen LWRs. Th e data considered in t h e ERDS system mainly concern safety , productiv ity , and reliability, and t h ey are collected by se veral national insti tu bons . One o f t h e most fundamental roles o f ERDS is indeed merging a n d standardizing a ll t h ese data , which, due t o t h e i r different origi n and to the peculiarities of each national r egulation , have a very he terogeneous nature . According to the different types of collected in f o rma t ion, ERDS h as been structured into four main subsys tems (see Fig. 1 ) : Component Event Data Bank (CEDB) The purpose of this s ub system is merg i ng data on reactor component failures coming from di fferent sources , e .g., data provided by already e x isting nat i ona l systems or collected by means of 'ad hoc ' data col lect i on campaigns (Component Eve n t Data Bank Handbook , 1983).
(' I
Ill. CEDB l:ontai ns a collection of events , suc h as failures and r e pairs, whi c h occurred during o peration to all major nuclear power plant components. For eac h comp onent t he technical s p eci ficati o n s type , enginee ring characteristics , (component etc.) , t ogeth er with o peration and e nv ironmental co nditions , are stored as well .
In order t o standardize this information, a set of r e ference c l assifications has been defined and used , wh ene ver poss i ble , for the construction of ERDS. Presently , CEDB conta ins a san,ple of data referring to about 2000 components a nd 1000 events , whi c h hav e been used for test purp oses . A trial o per atio n has now been i nitiated with the co- operation of major European utilities , s u c h as EDF , Vattenfall, ENEL , RWE, a nd KEMA, that should allow reaching a volume of about 5000 comp onents and 3500 events . Abnormal Occurrences Repo rting System (AORS) The purpose of this subsystem is the co ll ectio n of information con tained i n national abnormal occurrences reporting systems, in or der to provid e a ser vi ce and a too l for safety analysis . AORS is constituted of a centra lized data base, in which reported events are sto red according to European format a n d reference classifications , and of a set of sate l lite d ata bases , wh ere data coming from different national reporti ng systems are stored with their original coding and classi fi ca tion . At present , the centralized data base is loaded wi th data com ing from the satel1 i te data bases through a complex computer - aided transcoding process .
(URONET THE PROCESSING NETWORK
UNIPEOE
oseE
UTILITIES
UTIUT .; OPE RAT . REGULAT . AG .
UTILITIES
OPERATORS
EUREOATA
OPERAT ORS
IAEA
UTILITIES OPERATORS USERS ON REG Ul. AG . NEGOCIAT .
Emllngdlllbu./tonnltllom
In prot'N diD basalconnKllom
Fig. 1 - General arc hi tectu re of the ERDS
USERS ON NEGOCIAT .
USERS ON NEGOCIAT
169
.". n Int e llige nt Interface Information stored in AORS includes: - general abnormal event data (the plant status before the event and the main consequences of the event); - occurrences data (cause, way of discovery, actions taken,
systems involved, etc.) concern -
ing the event-originating occurrence and subse-
quent occurrences which happened in the frame of the same event; - fr ee - text event descriptions. At present, in the centralized data base o f AORS almost 10,000 events have been stored, while for other 20 I 000 events, concerni ng reactors in U.S.A., France, Sweden, and Italy, automatic transcoding
procedures
are
being
satellite data bases contain about in total.
studied.
The
30,000 events
Operating Unit Status Reports (OURS) The purpose of this subsystem is collection , o rganization, and dissemination of producti vi ty and outage data on European power reactors. Stored information is consituted by: - outage data, including the nature of significant outages and the involved systems; - performance data giving an overall summary of the performance indexes used to evaluate the productivity of the reactor. OURS already contains information on European reactors for the years 1982 , 1983, and 1984. Reliability Parameter Data Bank (RPDB) The purpose of this subsystem is collection and organization of reliability parameters for similar classes of components. The architecture of RP DB is presently being designed. ERDS has been developed on an AMDAHL 470/V8 computer, under OS/MVS , and it is based on ADABAS data base management system of SOFTWARE A.G. (including NATURAL and ADASCRIPT query languages).
STATEMENT OF THE PROBLEM End-users of ERDS are faced with several difficul ties concerning : - formulating meaningful queries, and - correctly interpreting the answers received from the data base. The first type of difficulties involves the following aspects: - the user cannot relay, in formulating his requests, on a high- level , problem oriented view of the data, but he must refer to the lowest level of information corresponding to the data explicitly represented in the data base; - the user must learn and use a formal query lan guage, far from natural language; - the user must be fully acquainted with the logical and physical structure of the data base and with some technicalities of its implementation; - the user must be knowledgeabl e of the information stored in the data base and of its meaning and use; - user's queries generally contain the request both of data that can be directly retrieved in the data base and of information that can only
be obtained through a complex processing (by means of available application programs) of an appropriate set of elementary data extracted from the data base. The second type of difficulties is due to the conceptual gap existing between the expectations of the user and the way In which the data base management systems formulate answers: it is general ly very di fficu I t for the· user to obtai n c I ear and meaningful information . Moreover, the user is often interested in interrogating not on ly a particular subsystem of ERDS, but two or more subsystems at the same time. For examp le, the query "Which fai 1 ures occurred to gate valves for borated water , caused a decrease of safety level and productivity" involves three subsystems, namely : CED8 for engineering charac teristics of components, AORS for safety relevant events, and OURS for productivity. Two different attempts have been made in recent the prob lems years to face, at least in part, above illustrated: 1. An enhanced version of ADASCRIPT (called ADACCL) has been implemented at the JRC devoted to interrogate AORS and OURS. Its major objectives are; a) to free the user from the necessi ty of having precise knowledge of the mul tifi le structure of the data; b) to allow use of synonyms to denote the information fields involved in the query; c) to facilitate the use of special purpose application programs on retrieved data. 2. A specific interactive query language has been implemented for the interrogation of CEDB (Carlesso, Bastin, and Jaarsma, 1983) . This language allows the user to select any desired data sample, to check the validity of the retrieved sample rejecting uncomplete data, to display the sample, and to submit the retrieved data to reliability assessment programs. These tools are currently used, but they satisfy only partially the user needs above mentioned. For this reason, the design of an inte 11 igent interface supporting natural language dialogue, cooperative behaviour, and expert assistance has been undertaken. As a first step, a feasibi li ty study has been carried out in order to assess the practicabi l i ty of the project and to devise the basic technical choices and design criteria.
USER PROFILE AND REQUIREMF.NTS Users of ERDS that could benefit by avai labi li ty of an intelligent interface belong to two classes: Technical users, mainly engineers interested in the design, operation, maintenance, and management of nuclear power plants . These are fully acquainted with the application domain (and therefore with the semantics of the data stored in the data base), and are generally knowledge-
F. Bastill 1'1 !I/.
170
able o f the available app l ication programs and of their use. However, only a very sma l l subs et of them has pr e c ise kno wI e dge o f t h e log i ca 1 structure of the data base, of the avai lable query l anguages , and of the proc edures required to navigate inside EROS. - Non-te c hnical users, mainly high- level managers , who only occasionally access the data base, often through an intermediary. Thes e us ers are not fully acquainted wi t h the application domain and are not kno wledgeable of the t e chnical aspects o f EROS. An analysis o f the needs and expe ctations o f a sample of users ex t ra c t e d from both the above mentioned classes has shown that the EROS interface should include several different c apabilities : - understanding user's requests t o EROS expressed in natural language and mapping the i r meaning into a formal in t ernal represen t ati o n; - managing a bounded sc o pe dial ogue with the user, aimed at: supporting the us e r in the intera c tion wi th EROS in order to enable him to express his information needs in t he most suitab l e way; resolving pos s ible ambiguities (both linguis tic and conceptual) in the user's requests; obtaining necessary information o mi t t e d by the user in expressing his request; supporting the user with information about the c ontent, structure, and operation of EROS and about the available application programs ; - recognizing which EROS subsystems are invol ved
a simple formal language , thus discarding the issue of having a natural language interaction; 3 . designing a natural language interface to one of the avai lable formal query languages for EROS (e.g. , AOACCL) , without paying attention to the intrinsic c onceptual difficulties of using the EROS data base. All the above mentioned approaches show very hard difficulties and disadvantages. More specifically , the second and third approach are unsa t isfactory, as the second one does not consider the requirements of non - technical users , and the third one does not provide an effective interaction with ER OS for technical users. Clearly, the first approach can fully match all the expe c ted requirements , as it is adequate both for technica l and non - technical users. However , the technical difficul ties it is expected to in volve have been estimated so hard by the feasibil ity study that , in order to arrive to a reasonab l e it has been necessary to make design proposal, some further assumptions . The most convenient way out of these difficulties is to split t h e global interface into two simpler and more specific interfaces , each one devoted to tackle separately one of the two major issues involved in the access by technical and no n- technical users to EROS , namely: (i) bridging the conceptua l gap between the exigencies of the users and the complex log ical structure of EROS , a nd (ii) al l owing t he users to interrogate EROS by means of natural l anguage. The resulting proposal comprises therefore t h e design of two systems:
in answering the user's request;
- generating correct search programs in the appro priate formal query language; - recognizing which kind of further processing of the extracted data through appl i c ation pro grams is requested (if any) ; - evaluating the obtained data for significance and usability for further processing through appli c ation programs (if requested); - possible refinement of the search until the obtained data become signi fi c ant and fi t the requirements of the application programs to be executed on them (if any); - activation of the appropriate application programs (if requested) , c ritical evaluation of obtained results (significance, certainty, etc.) , and presentation to the user.
INTERFACE SPECIFICATIONS AND BASIC ARCHITECTURE Several approaches can be identified for designing an intelligent interface to EROS s u pporting most of the capabilities liste d in the previous sectio n . Three basic possibil i ti e s have been investi gated:
- an expert interface (Er), including a formal interaction language suitable for s pecifying the information needs of the users (called prob l em description language - POLl and devoted to support technical users in the access to the ER OS data base ; de _ a natural language query interface (N LQI ) , signed as a front-end to the El , via the POL, and devoted to support non- technica l users through natural language dialogue. The basic archi tecture of the EROS interface is illustrated in Fig . 2 . It can ensure a l l t he bene fits expected by the users , and, besides, it features a flexible and modular structu r e that can greatly contribute to make design and implementation easier and faster . In fact , once t h e i nterme diate language POL has been defined , the El and NLQI can be developed almost independently. Moreboth of them may be implemented gradual l y, over , starting from a simple basic kernel and proceed i ng further stepwise towards higher-level and more challenging issues.
THE NATURAL LANGUAGE QUERY INTERFACE 1. designing a global interface wh i ch implements all needed capabilities within a composite sys tem, including a natural language interface and a complex expert system; 2 . designing an expert interface capable o f supporting an effective and reasonably easy use of ER OS by nuc l ear po wer plant specialists through
A preliminary point to be faced in the NLQI is the identification of and concep t ual competence to put in latter is fully determined by the
the design of the linguistic it. Whi l e the definition of
the POL language , the former requires a specific analysis of a br o ad set of sample queries col. € ct e d fr o m the users. The results obtained during
An I Ille llige lll I nt e rfacl' n atur al lan gu age di a~
NA TURAL LAN GUAGE QUERY I!,; TERFA CE
no n·techni· eal users
1
PD L d ia
tee
~
logue
17 1
11
.-.
Eu ropean ReI ia b il jty
Data System I ERDS )
pro bl em descri ption
use rs
! EXPE RT I NTERFACE
I
----.
appl icatio n
prog rams
Fig . 2 - Basic architecture of the ERDS interface the feasibility study allow sketching a prelimi nary list of requirements for linguistic competence . Basic require me nts
- understanding simple direct queries , both in interrogative and imperative form; - dea l ing with simple subordinate phrases and relative pronoun reference j
- dealing with co-ordinate phrases used to express multiple queries; - dealing with simple anophoric references (e.g ., possessi ve and demontrative adjectives or pronouns) (e . g. , " List the failures wi th unavail abili ty time less than ten hours. Among these select ... "); - dealing with usual cases of ellipsis (e.g. , verb, subject) (e.g . , "Which U.S . plants exh ibi ted more t han two abnormal events in 1981? and in 1983? " ) ; - correcting misspel l ed words. Advanc ed requiremen ts - deal i ng with ungrammatical sentences ; - understa ndi ng te l egra ph ic sentences (e . g. , "Events with power loss greater than 50%" , 11 Pumps : sorted by power 1' ) ;
- dealing
with null queries (e.g . , "How many concerned pressuri zed water reactors in Be l gium in 1979?" Answer : "No pressurized
even ts
water
reactor
was
ope r ating
in
Belgium
in
1979 " ) ;
- dealing with incomplete queries; - understanding indirect speech acts; - adding information in concise answers
(e . g. ,
uti 1 i ze component
xx?" -
"Which power reactors
Answer :
"Power reactors number .. , '
i . e. t hose
located in France " ); - managing a bounded scope c l arificat i on dia l ogue with t he user to solve ambiguous , unclear , or incomplete queries; - answering meta - queries ,
i . e. ,
queries
on
the
structur e and content of the data base (e . g. , "What cou l d you tell me about " Is there any 1 ink in data base between . . . and . . . ? " ) ; - recognizing out- of - domain queries (e.g . , " How many people work at the plant YY? " - Answer : "I do not know about employees, only about persons involved in accidents") .
One of t he basic pOints in t he constr uct i on of t he NLQI is the design of an effective a nd robust parser which can meet t he majority of the requirements above stated. Several techniques have been developed in recent years in the fields of computationa l linguistics and art i ficial i nte l l igence , that are worth to be cons i dered for t h e de s ign of the NLQI. Among these , the followi ng approac h es have been considered as specially relevant: - semantic grammars (Hendrix, Sacerdo t i , Sagalowicz, and Slocum , 1978); - goa l - oriented parsing (Guida a nd Tasso , 1982); - word expe r t parsing (Sma l l and Rieger , 1982) ; - activation - based parsing (Jones, 1983) ; - augmented phrase structure grammars (Heidorn , 1975) ; - defini te c l ause grammars (Pe rei ra and Warren , 1980) ; - genera liz ed ph r ase struct ure gra mma r s (Gazdar and Pullu m, 1982); - integrated syntactic- semantic approaches (Bobrow and Webber, 1980; Comino, Geme l lo , Guida , Rul lent, Sisto , an d Soma l vico, 1983 ; Ph illips, 1984) . Each o f t hese approac h es has its own fea t ures a nd can fit specific application requiremen ts. From a preliminary analysis and a limited exper i mentation on a set of sample utterances it can be foreseen that the most suitable technique for imp lementing the NLQI could be a semantics- based approac h (e.g. , sema n tic grammar s, goa l- directed parsing) loca lly supported by a bounded sco pe syntactic analysis (e.g . , augmented phrase structure grammars , activation based parsing) . In fact, user's questions seem to be syntactically simple , so that in actual operation t he relative weakness in syntactic competence wou l d no t af f ect per f ormance. A possible cri tical point in the construction of the parser is the organization and extent of t he vocabulary. In fact , many cases of ambiguity can arise in which the same term may have di fferent possible meanings , since it can refer to more t han one field a n d more than one file in the data base. These cases can genera 11 y be so I ved by means of se,""ntic information about the context in wh ich the word is utilized. the vocabulary is expected t o be very t1or eo ver I
172
F. Bastill r l 0/.
large, since it has to include a lot of technica l terminology. This could raise problems of effi cient retrieval , semantic interpretation , com p leteness, and updating. Another important tech n ical a spect concerned wi t h t he des i gn of the NLQl is the implementation of an effect i ve and gracefu l di al ogue. The user - system interaction cannot be one- way : dialogue is a basic .ssue of any natural language interface . The capabilities of a dialogue system include , in addition to the basic and advanced requirements listed above , echoing user's requests whenever needed for validating understanding or supporting user ' s confidence in the system capabilities , and appropriate presentation of results. These features are clear 1 y beyond the usual competence of a parser, and they have to be implemented by means of a spec i a I-purpose system which can cooperate with the parser for improving the level of interaction with the user.
THE EXPERT INTERfACE The design of the El involves consideration of several classes of requirements . Three main points are considered below:
- globa l role of the El, - requirements of t h e PDL language , - specific capabilities of the El . The El has the main purpose of offering to the technical user a sophisticated problem- oriented view on the data base . Direct visibility of the l ogi ca land phys i ca 1 structure of ER OS should be possible only if explicitly requested and, also in t h is case, access to the fi 1 es and use of data should be supported and appropriately fi 1 tered by' the El . The El shou 1 d allow users to represent i nformation needs at different levels of abstraction and detai l. A crucial point in the definition of requirements for the El is the PD L language for interacting wi th the technical use r s (or the NLQI) . Three aspects seem of basic importance in this respect : a 1 though , pos- the language should be formal, s i bly, with a simple and user- friendly syntactic structu r e;
- interactivity should be very high: long and complex command statements should be avoided, a mixed initiative co-operation including both commands and menu-based i nteraction could be appropriate ; - the language should be high- level and mainly problem- oriented , i.e . the user should be expected to specify only what he needs , not which data have to be r etrieved and processed. The analysi s of the capabi 1 i ties to be put into the El deserves particular attention. Eight major classes of tasks have been identified : - supporting the user in the definition of his information needs i - checking user ' s requests for consistence and completeness; - identifying which data and which application programs are needed to answer the request ;
-
-
-
identifying which subsysteCls of the ERDS are involved in the retrieval of needed data , and which search strategies have to be used; accessing ERDS and extracting data; evaluating extracted data for significancE, also in relation with the application programs (if any) to be executed on them; executing the requested application programs on the appropriate data ; evaluating the obtained resul ts and reporting them to the user .
The features of the El outlined above suggest the adoption of a knowledge- based approach (Hayes Roth, Waterman , and Lenat , 1983) with a highly modular architecture , made up of several intercon nected subsystems, each one specialized in a specific facet of the problem domain (Erman , London , and fickas, 1981). Four subsys t ems have been identified : user, prob l em , application , and data base specialist. In the implementation of the El , the design of an appropriate knowledge representation deserves particular attention . Production rules can be used to encode expert knowledge elicited from the domains of each of the sUbsystems in whic h the El has been spl it . However, the El , besides usi n g expert know l edge (empiric, inexact, i ncomplete , fragmentary , etc.), must be capable of deal i ng with general k no wledge (exact , taxonomic , descriptive , etc.) on the world to which data stored in ERDS refer (nuc lear power plan ts, reactor parts and operation, etc . ), and with f ormal knowledge on the mappi ng between the conceptua I prob 1 em- ori ented user view and the real (logical) structure of ERDS (i nc 1 udi ng the query I anguage and access procedures).
CONCLUSIONS The design of ooth NLQI and El systems is very demanding . I t has been necessary to di v i de the project into two phases : A. design ing a tion; B. design taking during
and iplementation of prototypes , includof basic research and experimenta -
lot
and construction of t h e target systems , advantage of the experience acquired the prototyping phase.
The prototypes wi 11 refer only to a small , possibly simplified subset of the problem domain. The following reduced domains will be considered : for the NLQI, a reduced vocabulary (specially for what concerns technical terminology), a subset of ERDS , a reduced set of linguistic features (all basic requi remen ts , but on 1y a few advanced ones and limited dialogue); for the El , a subset of ERDS , reduced knowledge bases , and only a kernel of the PDL language . At present, the design of the two proto t ypes has been started. They will be implemented on an AMDAHL 470/V8 computer under OS/MVS using INTERLISP , with ERDS .
in
order to
allow an
easy connection
:\ 11 [l1t e lli gc l1t [l1t c ria cc Three major directions for future research can be mentioned, that are expected to extend the capabi l ities of the El: - the implementation of learning capabilities that can enable the El to automatically acquire new knowledge from the users and encode it into rules (Carbonell, Michalski , and Mitchell , 1983) ; - the extension of the reasoning capabilities in such a way as to make the El capable of discov -
ering and ana lysing causal re lations among the data stored in ER OS (e.g., accidents, failures, etc. ); - the development of approximate reasoning tech niques that can enable the El to deal with incomplete and uncertain knowledge (Prade , 1983; Sk ala , Termini and Tr i llas , 1984) .
REFERENCES Bobrow , R.J. and B.L. Webber (1980) . Knowledge representati on for syntacti c / semanti c proces sing. Proc 1st Nat . Conf. on Artificial Intell igence, Stanford, CA . Brajnik, G., G. Guida and C. Tasso (1985) . An expert interface for effective man-mach i ne interaction. In L. Bolc (Ed.), Cooperat i ve Inte r active Systems, Springer- Verlag , Berlin , FRG . s. (1983). General informatic Capobianchi , structure and software facilities developed f or the Europea n Reliability Data System Proc . 4th EUREDATA Conf. , Venice , (ERDS) . Italy . Carbonell , J . G. , R.S . Michals k i and T . M. Mitche l l (Eds . ) (1983). Machine Learning , An Artificial Intelligence Approach . Tioga, Palo Alto , CA. Carlesso, S ., F . Bastin and R.D . Jaarsma (1983). Th e i nqu i ry and statistical analysis of t h e CEDB by means of a terminal. CEC, JRC -I spra Establis h ment , TN . I . 05 . 01 . 83.68 . Cohe n, P . R. and C.R. Perrault (1982) . Beyond question answering . In W. G. Lehnert , M.H. Ringle (Eds.) , Strategies for Natural Language Processing , Lawrence Erlbaum , Hillsdale, NJ. Comino , R., R. Geme l lo , G. Guida , C. Rullen t, L . Sisto and M. Somalvico (1983). Understanding natural language through parallel processing of syntactic and semantic information : IIn a pplication to data base query . Proc. 8th Int . Intelligence , J oint Conf. on Artificial Kar l sruhe, FRG , 663 - 667. Component Event Data Bank Handbook (1984). CEC, JRC, Ispra Establishment , TN.I.05 . Cl . 84.66 . Erman, L. D. , P. E. London and S. F . F i ckas (1981). The des ign and an exampl e of use of HEARSAYI l l. Proc. 7th Int. Joint Conf. on Artificial Intel l igence, Vancouver, BC , Can ada; 409 - 415 . Gazdar , G. and G.K. Pu1lum (1982) . Generalized phrase structure grammars : A theoreti ca 1 synopsis . Indiana Unversity Linguistics Club, Bloomington, IN . Guida , G. and C. Tasso (1982). NLI: A robust interface for natural language person- machine communication . Int . J. of Man-Machine Studies 17, 417 - 433 .
[ 7:\
Guida , G. and C. Tasso (1983) . An expert interme diary system for interactive document retrieval. Automatica IFAC 19(6) , 759 - 766 . Hayes- Roth, F. , D. A. Watermann and D. B. Lenat (Eds. ) (1983) . Bui 1 ding Expert Systems . Addison -Wesley, Reading , MA. Heidorn , G. E. (1975). Augmented phrase structure g r ammars . In B. L . Nash- Webber , R. C. Sc h a n k (Eds.) . Theo r etical Issues in Natura l Language Processing , ACL , Menlo Pa r k , CA . E.D . Sacerdoti , D. Saga l owicz and He n drix , G. G. , Develop i ng a natural lanJ. Slocum (1978). guage interface t o complex data. ACM Trans. on Data Base Systems 3(2) , 105-1 47. Jones , M.A . (1983). Activation-based pars l ng. Proc. 8 t h Int . Jo i nt Conf. o n Arti fi cial Intel l igence , Karlsruhe , FRG , 678-682 . Kaplan , S . J. (1983) . Cooperative responses from a portab l e natural language data base query system . In M. Brady, R. Bersw i c k (Eds . ) , Computational Models of Discourse . Th e MIT Pr ess, Cambridge, MA. King , J . J. (Ed . ) (1983) . Specia l issue on AI and data base research . ACM Sigart Newsletter 86 , 32-72 . Mancini , G. , J . Amesz , S. Capobianchi and P . Bastianin i ( 1982). ERDS: An organ ized in f orma tion exchange on t h e operat i on of the European Nuc l ear Reactors. Int. Con f. on Nuclear Power Exper i ence, Vienna , Austria . Pereira , F. C. N. and D.H.D . Warren ( 1980). Definite clause grammars for language analysis A survey of th e formalism and a comparison wi th augmented transi ti on networks . Arti fic ia I Inte ll igence 13 , 23 1- 278 . Phillips , B. (1984) . An object - oriented parser. In B . G. Bar a , G. Guida (Eds . ) , Computationa l Mode l s of Natural Language Processing, NorthHol l and , Amste r dam , NL , 297 - 321 . Prade , H. (1983) . A synthetic view on approx i mate reasoning techniques. Proc. 8th Int . Joint Conf. o n Artificia l I ntelligence , Karlsruhe, FRG , 130- 136. Reichman-Adart , R. (1984) . Extended person-machine interface . Artificia l Intelligence 22, 157218 . Reiter , R., H. Gallaire, J.J. King , J . Mylopoulos and B.L. Webber (1983) . A panel on AI and data bases . Proc . 8th Int. Joint Conf . on Artifi cial Intelligence , Karlsruhe , FRG , 1199- 1206. Sk ala , H.J. , S. Termini and E. Tri ll as (Eds . ) (1984). Aspects of Vagueness . Reidel , Dordrecht , NL . Rieger (1982) . Parsing and Sma 1 1 , S. L . and C. comprehending with word experts (A theory and In W. Lehnert, M. Ring l e its realization). (Eds.) , Strategies for Natural Language Processing , Lawrence Erlbaum , Hillsdale , NJ .
©
II:,'. C \1;,11 · \1;" h illl' S , , {clll' ,
\ ' art'st:', llah, I ~1 ~"Fl
AN INTELLIGENT INTERFACE FOR ACCESSING A TECHNICAL DATA BASE F. Bastin*, S. Capobianchi*, Sylviane Carlesso*, G. Mancini *, T. Barbas**, A. D. Koletsos** G. Guida***, M. Somalvico***, and C. Tasso**** *CEC,
JRe
Ispra E ,/ a /!lis/lll1ml, Ispra , 1111 /\'
*o,oC,t.' ,c;, G m lll Ho/fin.1 *"" M,ho, P u/y/n hll i, Arlilirifl/ IlIlrlligma Proj,'(I, ,\li/all o, Ila/y **** / slillllo di ,\lall' l/wliC!l . 11I1() l'IIwlim
aimed at the correct , and effective way . European Reliability Data System - ERDS) in a friendly , The interface incl u des both natural language understanding and expert problem solving capabil i ties and features a modular archit e cture devoted to cope with the exigencies In the paper a brief introduction to ERDS of both technical and non - technical users. is first presented and the problems related to its use are analysed. The requirements of the users of ERDS are then investigated , and the basic archi tecture of an intelligent interface is illustrated. The design c r iteria of the two main modules of the i nterface , name ly the natural language query interface and the expert interface , are later discussed. At last , the present state of t h e project is reported and promising directi o ns for future research are out1 ined . Abstract.
The
paper
deals
with
the
I' Sisll'lll istira. C llil'nslh i di Cdilll', I. 'dilll'. Ila/\'
design
of
an
inte l ligent
int.erface
assisting e n d- users in accessing a large and complex technical data base (namely,
Key words.
I n telligent interfaces ,
natural language processing ,
knowledge-based sys-
tems , data bases .
II~TRODUCTION
One of the major recently emerging fields of ap plication of ar ti ficial intelligence is consti tuted by data base management systems (King , 1983 ; Reiter , Gallaire , King , Myl o pou l os, and 'webber , In this area many studies are currently 1983). carried out to implement systems known as intel1 i gent or expert i n'cer faces (Cohen and Pe rrau 1 t, 1982; Kap 1 an , 1983; Gui da and Tassu , 1983 ; Bra j nik, Guida , and Tasso , 1984; Reichman- Adar , 1984) , aimed at assisting end - users in accessing i nformation stored in a data base in a friend l y , correct, and ef f ective way. Generally , intelligent interfaces encompass c apa bilities of both nat ural language front - ends and expert systems , in order to bridge the linguistic gap between non - technical user and H.e formal query l anguage necessary for accesslng the data base, and to support the users in the complex task of effectively and correctly using the information stored in the data base. CQping wi th such problems becomes more and more
the European Reliability Data System (ERDS) , de velo ped at the Ispra JRC Establishment in t h e frame of the Reactor Safety programme. Th e use of such a 1arge system , composed of many intercon nected subsystems , requires precise knowledge in several domains in order to : - state problems correctly and manage their solution; - i den ti fy whi ch data are needed in the prob 1 eo. solving process ; - navigate inside the system and its multiple data bases to retrieve the desired information ; - validate the extracted data and possibly process them through speclal purpose application p r o grams . In such a case , natural language processing and expert system technologies can effecti ve ly con tribute in bridging the gap between end- users and the data base. This paper is devQted to illustrate the results of a feasibility study concerning an intell i gent in te r face to the ERDS data base recently carried out at Ispra JRC Establishment.
important , and almost mandatory, when one deals with complex technical data bases accessed by several classes of end - users. This is the case of
167
16H
F. Bastin DESCRIPTION OF ERDS
ERDS (European Rel iabi 1 i ty Data System) has been developed at the Ispra JRC Establishment over t h e past years in order to collect, manage , and dis seminate information about Light Water Reactors (LWR) located in t h e European Communi ty , and to a ccumul ate reliability data for probabilistic risk assessments (Mane-ini , Amesz , Capobianc h i , and 8astiani ni, 1982; Capob i anchi , 1983) . The original scope of ERDS h as been enlarged in some cases to in c lude non-Eur opean reactors (e . g. , the LER system o f NRC, U. S . A. ) and a 1 so reacto rs ot her t hen LWRs. Th e data considered in t h e ERDS system mainly concern safety , productiv ity , and reliability, and t h ey are collected by se veral national insti tu bons . One o f t h e most fundamental roles o f ERDS is indeed merging a n d standardizing a ll t h ese data , which, due t o t h e i r different origi n and to the peculiarities of each national r egulation , have a very he terogeneous nature . According to the different types of collected in f o rma t ion, ERDS h as been structured into four main subsys tems (see Fig. 1 ) : Component Event Data Bank (CEDB) The purpose of this s ub system is merg i ng data on reactor component failures coming from di fferent sources , e .g., data provided by already e x isting nat i ona l systems or collected by means of 'ad hoc ' data col lect i on campaigns (Component Eve n t Data Bank Handbook , 1983).
(' I
Ill. CEDB l:ontai ns a collection of events , suc h as failures and r e pairs, whi c h occurred during o peration to all major nuclear power plant components. For eac h comp onent t he technical s p eci ficati o n s type , enginee ring characteristics , (component etc.) , t ogeth er with o peration and e nv ironmental co nditions , are stored as well .
In order t o standardize this information, a set of r e ference c l assifications has been defined and used , wh ene ver poss i ble , for the construction of ERDS. Presently , CEDB conta ins a san,ple of data referring to about 2000 components a nd 1000 events , whi c h hav e been used for test purp oses . A trial o per atio n has now been i nitiated with the co- operation of major European utilities , s u c h as EDF , Vattenfall, ENEL , RWE, a nd KEMA, that should allow reaching a volume of about 5000 comp onents and 3500 events . Abnormal Occurrences Repo rting System (AORS) The purpose of this subsystem is the co ll ectio n of information con tained i n national abnormal occurrences reporting systems, in or der to provid e a ser vi ce and a too l for safety analysis . AORS is constituted of a centra lized data base, in which reported events are sto red according to European format a n d reference classifications , and of a set of sate l lite d ata bases , wh ere data coming from different national reporti ng systems are stored with their original coding and classi fi ca tion . At present , the centralized data base is loaded wi th data com ing from the satel1 i te data bases through a complex computer - aided transcoding process .
(URONET THE PROCESSING NETWORK
UNIPEOE
oseE
UTILITIES
UTIUT .; OPE RAT . REGULAT . AG .
UTILITIES
OPERATORS
EUREOATA
OPERAT ORS
IAEA
UTILITIES OPERATORS USERS ON REG Ul. AG . NEGOCIAT .
Emllngdlllbu./tonnltllom
In prot'N diD basalconnKllom
Fig. 1 - General arc hi tectu re of the ERDS
USERS ON NEGOCIAT .
USERS ON NEGOCIAT
169
.". n Int e llige nt Interface Information stored in AORS includes: - general abnormal event data (the plant status before the event and the main consequences of the event); - occurrences data (cause, way of discovery, actions taken,
systems involved, etc.) concern -
ing the event-originating occurrence and subse-
quent occurrences which happened in the frame of the same event; - fr ee - text event descriptions. At present, in the centralized data base o f AORS almost 10,000 events have been stored, while for other 20 I 000 events, concerni ng reactors in U.S.A., France, Sweden, and Italy, automatic transcoding
procedures
are
being
satellite data bases contain about in total.
studied.
The
30,000 events
Operating Unit Status Reports (OURS) The purpose of this subsystem is collection , o rganization, and dissemination of producti vi ty and outage data on European power reactors. Stored information is consituted by: - outage data, including the nature of significant outages and the involved systems; - performance data giving an overall summary of the performance indexes used to evaluate the productivity of the reactor. OURS already contains information on European reactors for the years 1982 , 1983, and 1984. Reliability Parameter Data Bank (RPDB) The purpose of this subsystem is collection and organization of reliability parameters for similar classes of components. The architecture of RP DB is presently being designed. ERDS has been developed on an AMDAHL 470/V8 computer, under OS/MVS , and it is based on ADABAS data base management system of SOFTWARE A.G. (including NATURAL and ADASCRIPT query languages).
STATEMENT OF THE PROBLEM End-users of ERDS are faced with several difficul ties concerning : - formulating meaningful queries, and - correctly interpreting the answers received from the data base. The first type of difficulties involves the following aspects: - the user cannot relay, in formulating his requests, on a high- level , problem oriented view of the data, but he must refer to the lowest level of information corresponding to the data explicitly represented in the data base; - the user must learn and use a formal query lan guage, far from natural language; - the user must be fully acquainted with the logical and physical structure of the data base and with some technicalities of its implementation; - the user must be knowledgeabl e of the information stored in the data base and of its meaning and use; - user's queries generally contain the request both of data that can be directly retrieved in the data base and of information that can only
be obtained through a complex processing (by means of available application programs) of an appropriate set of elementary data extracted from the data base. The second type of difficulties is due to the conceptual gap existing between the expectations of the user and the way In which the data base management systems formulate answers: it is general ly very di fficu I t for the· user to obtai n c I ear and meaningful information . Moreover, the user is often interested in interrogating not on ly a particular subsystem of ERDS, but two or more subsystems at the same time. For examp le, the query "Which fai 1 ures occurred to gate valves for borated water , caused a decrease of safety level and productivity" involves three subsystems, namely : CED8 for engineering charac teristics of components, AORS for safety relevant events, and OURS for productivity. Two different attempts have been made in recent the prob lems years to face, at least in part, above illustrated: 1. An enhanced version of ADASCRIPT (called ADACCL) has been implemented at the JRC devoted to interrogate AORS and OURS. Its major objectives are; a) to free the user from the necessi ty of having precise knowledge of the mul tifi le structure of the data; b) to allow use of synonyms to denote the information fields involved in the query; c) to facilitate the use of special purpose application programs on retrieved data. 2. A specific interactive query language has been implemented for the interrogation of CEDB (Carlesso, Bastin, and Jaarsma, 1983) . This language allows the user to select any desired data sample, to check the validity of the retrieved sample rejecting uncomplete data, to display the sample, and to submit the retrieved data to reliability assessment programs. These tools are currently used, but they satisfy only partially the user needs above mentioned. For this reason, the design of an inte 11 igent interface supporting natural language dialogue, cooperative behaviour, and expert assistance has been undertaken. As a first step, a feasibi li ty study has been carried out in order to assess the practicabi l i ty of the project and to devise the basic technical choices and design criteria.
USER PROFILE AND REQUIREMF.NTS Users of ERDS that could benefit by avai labi li ty of an intelligent interface belong to two classes: Technical users, mainly engineers interested in the design, operation, maintenance, and management of nuclear power plants . These are fully acquainted with the application domain (and therefore with the semantics of the data stored in the data base), and are generally knowledge-
F. Bastill 1'1 !I/.
170
able o f the available app l ication programs and of their use. However, only a very sma l l subs et of them has pr e c ise kno wI e dge o f t h e log i ca 1 structure of the data base, of the avai lable query l anguages , and of the proc edures required to navigate inside EROS. - Non-te c hnical users, mainly high- level managers , who only occasionally access the data base, often through an intermediary. Thes e us ers are not fully acquainted wi t h the application domain and are not kno wledgeable of the t e chnical aspects o f EROS. An analysis o f the needs and expe ctations o f a sample of users ex t ra c t e d from both the above mentioned classes has shown that the EROS interface should include several different c apabilities : - understanding user's requests t o EROS expressed in natural language and mapping the i r meaning into a formal in t ernal represen t ati o n; - managing a bounded sc o pe dial ogue with the user, aimed at: supporting the us e r in the intera c tion wi th EROS in order to enable him to express his information needs in t he most suitab l e way; resolving pos s ible ambiguities (both linguis tic and conceptual) in the user's requests; obtaining necessary information o mi t t e d by the user in expressing his request; supporting the user with information about the c ontent, structure, and operation of EROS and about the available application programs ; - recognizing which EROS subsystems are invol ved
a simple formal language , thus discarding the issue of having a natural language interaction; 3 . designing a natural language interface to one of the avai lable formal query languages for EROS (e.g. , AOACCL) , without paying attention to the intrinsic c onceptual difficulties of using the EROS data base. All the above mentioned approaches show very hard difficulties and disadvantages. More specifically , the second and third approach are unsa t isfactory, as the second one does not consider the requirements of non - technical users , and the third one does not provide an effective interaction with ER OS for technical users. Clearly, the first approach can fully match all the expe c ted requirements , as it is adequate both for technica l and non - technical users. However , the technical difficul ties it is expected to in volve have been estimated so hard by the feasibil ity study that , in order to arrive to a reasonab l e it has been necessary to make design proposal, some further assumptions . The most convenient way out of these difficulties is to split t h e global interface into two simpler and more specific interfaces , each one devoted to tackle separately one of the two major issues involved in the access by technical and no n- technical users to EROS , namely: (i) bridging the conceptua l gap between the exigencies of the users and the complex log ical structure of EROS , a nd (ii) al l owing t he users to interrogate EROS by means of natural l anguage. The resulting proposal comprises therefore t h e design of two systems:
in answering the user's request;
- generating correct search programs in the appro priate formal query language; - recognizing which kind of further processing of the extracted data through appl i c ation pro grams is requested (if any) ; - evaluating the obtained data for significance and usability for further processing through appli c ation programs (if requested); - possible refinement of the search until the obtained data become signi fi c ant and fi t the requirements of the application programs to be executed on them (if any); - activation of the appropriate application programs (if requested) , c ritical evaluation of obtained results (significance, certainty, etc.) , and presentation to the user.
INTERFACE SPECIFICATIONS AND BASIC ARCHITECTURE Several approaches can be identified for designing an intelligent interface to EROS s u pporting most of the capabilities liste d in the previous sectio n . Three basic possibil i ti e s have been investi gated:
- an expert interface (Er), including a formal interaction language suitable for s pecifying the information needs of the users (called prob l em description language - POLl and devoted to support technical users in the access to the ER OS data base ; de _ a natural language query interface (N LQI ) , signed as a front-end to the El , via the POL, and devoted to support non- technica l users through natural language dialogue. The basic archi tecture of the EROS interface is illustrated in Fig . 2 . It can ensure a l l t he bene fits expected by the users , and, besides, it features a flexible and modular structu r e that can greatly contribute to make design and implementation easier and faster . In fact , once t h e i nterme diate language POL has been defined , the El and NLQI can be developed almost independently. Moreboth of them may be implemented gradual l y, over , starting from a simple basic kernel and proceed i ng further stepwise towards higher-level and more challenging issues.
THE NATURAL LANGUAGE QUERY INTERFACE 1. designing a global interface wh i ch implements all needed capabilities within a composite sys tem, including a natural language interface and a complex expert system; 2 . designing an expert interface capable o f supporting an effective and reasonably easy use of ER OS by nuc l ear po wer plant specialists through
A preliminary point to be faced in the NLQI is the identification of and concep t ual competence to put in latter is fully determined by the
the design of the linguistic it. Whi l e the definition of
the POL language , the former requires a specific analysis of a br o ad set of sample queries col. € ct e d fr o m the users. The results obtained during
An I Ille llige lll I nt e rfacl' n atur al lan gu age di a~
NA TURAL LAN GUAGE QUERY I!,; TERFA CE
no n·techni· eal users
1
PD L d ia
tee
~
logue
17 1
11
.-.
Eu ropean ReI ia b il jty
Data System I ERDS )
pro bl em descri ption
use rs
! EXPE RT I NTERFACE
I
----.
appl icatio n
prog rams
Fig . 2 - Basic architecture of the ERDS interface the feasibility study allow sketching a prelimi nary list of requirements for linguistic competence . Basic require me nts
- understanding simple direct queries , both in interrogative and imperative form; - dea l ing with simple subordinate phrases and relative pronoun reference j
- dealing with co-ordinate phrases used to express multiple queries; - dealing with simple anophoric references (e.g ., possessi ve and demontrative adjectives or pronouns) (e . g. , " List the failures wi th unavail abili ty time less than ten hours. Among these select ... "); - dealing with usual cases of ellipsis (e.g. , verb, subject) (e.g . , "Which U.S . plants exh ibi ted more t han two abnormal events in 1981? and in 1983? " ) ; - correcting misspel l ed words. Advanc ed requiremen ts - deal i ng with ungrammatical sentences ; - understa ndi ng te l egra ph ic sentences (e . g. , "Events with power loss greater than 50%" , 11 Pumps : sorted by power 1' ) ;
- dealing
with null queries (e.g . , "How many concerned pressuri zed water reactors in Be l gium in 1979?" Answer : "No pressurized
even ts
water
reactor
was
ope r ating
in
Belgium
in
1979 " ) ;
- dealing with incomplete queries; - understanding indirect speech acts; - adding information in concise answers
(e . g. ,
uti 1 i ze component
xx?" -
"Which power reactors
Answer :
"Power reactors number .. , '
i . e. t hose
located in France " ); - managing a bounded scope c l arificat i on dia l ogue with t he user to solve ambiguous , unclear , or incomplete queries; - answering meta - queries ,
i . e. ,
queries
on
the
structur e and content of the data base (e . g. , "What cou l d you tell me about " Is there any 1 ink in data base between . . . and . . . ? " ) ; - recognizing out- of - domain queries (e.g . , " How many people work at the plant YY? " - Answer : "I do not know about employees, only about persons involved in accidents") .
One of t he basic pOints in t he constr uct i on of t he NLQI is the design of an effective a nd robust parser which can meet t he majority of the requirements above stated. Several techniques have been developed in recent years in the fields of computationa l linguistics and art i ficial i nte l l igence , that are worth to be cons i dered for t h e de s ign of the NLQI. Among these , the followi ng approac h es have been considered as specially relevant: - semantic grammars (Hendrix, Sacerdo t i , Sagalowicz, and Slocum , 1978); - goa l - oriented parsing (Guida a nd Tasso , 1982); - word expe r t parsing (Sma l l and Rieger , 1982) ; - activation - based parsing (Jones, 1983) ; - augmented phrase structure grammars (Heidorn , 1975) ; - defini te c l ause grammars (Pe rei ra and Warren , 1980) ; - genera liz ed ph r ase struct ure gra mma r s (Gazdar and Pullu m, 1982); - integrated syntactic- semantic approaches (Bobrow and Webber, 1980; Comino, Geme l lo , Guida , Rul lent, Sisto , an d Soma l vico, 1983 ; Ph illips, 1984) . Each o f t hese approac h es has its own fea t ures a nd can fit specific application requiremen ts. From a preliminary analysis and a limited exper i mentation on a set of sample utterances it can be foreseen that the most suitable technique for imp lementing the NLQI could be a semantics- based approac h (e.g. , sema n tic grammar s, goa l- directed parsing) loca lly supported by a bounded sco pe syntactic analysis (e.g . , augmented phrase structure grammars , activation based parsing) . In fact, user's questions seem to be syntactically simple , so that in actual operation t he relative weakness in syntactic competence wou l d no t af f ect per f ormance. A possible cri tical point in the construction of the parser is the organization and extent of t he vocabulary. In fact , many cases of ambiguity can arise in which the same term may have di fferent possible meanings , since it can refer to more t han one field a n d more than one file in the data base. These cases can genera 11 y be so I ved by means of se,""ntic information about the context in wh ich the word is utilized. the vocabulary is expected t o be very t1or eo ver I
172
F. Bastill r l 0/.
large, since it has to include a lot of technica l terminology. This could raise problems of effi cient retrieval , semantic interpretation , com p leteness, and updating. Another important tech n ical a spect concerned wi t h t he des i gn of the NLQl is the implementation of an effect i ve and gracefu l di al ogue. The user - system interaction cannot be one- way : dialogue is a basic .ssue of any natural language interface . The capabilities of a dialogue system include , in addition to the basic and advanced requirements listed above , echoing user's requests whenever needed for validating understanding or supporting user ' s confidence in the system capabilities , and appropriate presentation of results. These features are clear 1 y beyond the usual competence of a parser, and they have to be implemented by means of a spec i a I-purpose system which can cooperate with the parser for improving the level of interaction with the user.
THE EXPERT INTERfACE The design of the El involves consideration of several classes of requirements . Three main points are considered below:
- globa l role of the El, - requirements of t h e PDL language , - specific capabilities of the El . The El has the main purpose of offering to the technical user a sophisticated problem- oriented view on the data base . Direct visibility of the l ogi ca land phys i ca 1 structure of ER OS should be possible only if explicitly requested and, also in t h is case, access to the fi 1 es and use of data should be supported and appropriately fi 1 tered by' the El . The El shou 1 d allow users to represent i nformation needs at different levels of abstraction and detai l. A crucial point in the definition of requirements for the El is the PD L language for interacting wi th the technical use r s (or the NLQI) . Three aspects seem of basic importance in this respect : a 1 though , pos- the language should be formal, s i bly, with a simple and user- friendly syntactic structu r e;
- interactivity should be very high: long and complex command statements should be avoided, a mixed initiative co-operation including both commands and menu-based i nteraction could be appropriate ; - the language should be high- level and mainly problem- oriented , i.e . the user should be expected to specify only what he needs , not which data have to be r etrieved and processed. The analysi s of the capabi 1 i ties to be put into the El deserves particular attention. Eight major classes of tasks have been identified : - supporting the user in the definition of his information needs i - checking user ' s requests for consistence and completeness; - identifying which data and which application programs are needed to answer the request ;
-
-
-
identifying which subsysteCls of the ERDS are involved in the retrieval of needed data , and which search strategies have to be used; accessing ERDS and extracting data; evaluating extracted data for significancE, also in relation with the application programs (if any) to be executed on them; executing the requested application programs on the appropriate data ; evaluating the obtained resul ts and reporting them to the user .
The features of the El outlined above suggest the adoption of a knowledge- based approach (Hayes Roth, Waterman , and Lenat , 1983) with a highly modular architecture , made up of several intercon nected subsystems, each one specialized in a specific facet of the problem domain (Erman , London , and fickas, 1981). Four subsys t ems have been identified : user, prob l em , application , and data base specialist. In the implementation of the El , the design of an appropriate knowledge representation deserves particular attention . Production rules can be used to encode expert knowledge elicited from the domains of each of the sUbsystems in whic h the El has been spl it . However, the El , besides usi n g expert know l edge (empiric, inexact, i ncomplete , fragmentary , etc.), must be capable of deal i ng with general k no wledge (exact , taxonomic , descriptive , etc.) on the world to which data stored in ERDS refer (nuc lear power plan ts, reactor parts and operation, etc . ), and with f ormal knowledge on the mappi ng between the conceptua I prob 1 em- ori ented user view and the real (logical) structure of ERDS (i nc 1 udi ng the query I anguage and access procedures).
CONCLUSIONS The design of ooth NLQI and El systems is very demanding . I t has been necessary to di v i de the project into two phases : A. design ing a tion; B. design taking during
and iplementation of prototypes , includof basic research and experimenta -
lot
and construction of t h e target systems , advantage of the experience acquired the prototyping phase.
The prototypes wi 11 refer only to a small , possibly simplified subset of the problem domain. The following reduced domains will be considered : for the NLQI, a reduced vocabulary (specially for what concerns technical terminology), a subset of ERDS , a reduced set of linguistic features (all basic requi remen ts , but on 1y a few advanced ones and limited dialogue); for the El , a subset of ERDS , reduced knowledge bases , and only a kernel of the PDL language . At present, the design of the two proto t ypes has been started. They will be implemented on an AMDAHL 470/V8 computer under OS/MVS using INTERLISP , with ERDS .
in
order to
allow an
easy connection
:\ 11 [l1t e lli gc l1t [l1t c ria cc Three major directions for future research can be mentioned, that are expected to extend the capabi l ities of the El: - the implementation of learning capabilities that can enable the El to automatically acquire new knowledge from the users and encode it into rules (Carbonell, Michalski , and Mitchell , 1983) ; - the extension of the reasoning capabilities in such a way as to make the El capable of discov -
ering and ana lysing causal re lations among the data stored in ER OS (e.g., accidents, failures, etc. ); - the development of approximate reasoning tech niques that can enable the El to deal with incomplete and uncertain knowledge (Prade , 1983; Sk ala , Termini and Tr i llas , 1984) .
REFERENCES Bobrow , R.J. and B.L. Webber (1980) . Knowledge representati on for syntacti c / semanti c proces sing. Proc 1st Nat . Conf. on Artificial Intell igence, Stanford, CA . Brajnik, G., G. Guida and C. Tasso (1985) . An expert interface for effective man-mach i ne interaction. In L. Bolc (Ed.), Cooperat i ve Inte r active Systems, Springer- Verlag , Berlin , FRG . s. (1983). General informatic Capobianchi , structure and software facilities developed f or the Europea n Reliability Data System Proc . 4th EUREDATA Conf. , Venice , (ERDS) . Italy . Carbonell , J . G. , R.S . Michals k i and T . M. Mitche l l (Eds . ) (1983). Machine Learning , An Artificial Intelligence Approach . Tioga, Palo Alto , CA. Carlesso, S ., F . Bastin and R.D . Jaarsma (1983). Th e i nqu i ry and statistical analysis of t h e CEDB by means of a terminal. CEC, JRC -I spra Establis h ment , TN . I . 05 . 01 . 83.68 . Cohe n, P . R. and C.R. Perrault (1982) . Beyond question answering . In W. G. Lehnert , M.H. Ringle (Eds.) , Strategies for Natural Language Processing , Lawrence Erlbaum , Hillsdale, NJ. Comino , R., R. Geme l lo , G. Guida , C. Rullen t, L . Sisto and M. Somalvico (1983). Understanding natural language through parallel processing of syntactic and semantic information : IIn a pplication to data base query . Proc. 8th Int . Intelligence , J oint Conf. on Artificial Kar l sruhe, FRG , 663 - 667. Component Event Data Bank Handbook (1984). CEC, JRC, Ispra Establishment , TN.I.05 . Cl . 84.66 . Erman, L. D. , P. E. London and S. F . F i ckas (1981). The des ign and an exampl e of use of HEARSAYI l l. Proc. 7th Int. Joint Conf. on Artificial Intel l igence, Vancouver, BC , Can ada; 409 - 415 . Gazdar , G. and G.K. Pu1lum (1982) . Generalized phrase structure grammars : A theoreti ca 1 synopsis . Indiana Unversity Linguistics Club, Bloomington, IN . Guida , G. and C. Tasso (1982). NLI: A robust interface for natural language person- machine communication . Int . J. of Man-Machine Studies 17, 417 - 433 .
[ 7:\
Guida , G. and C. Tasso (1983) . An expert interme diary system for interactive document retrieval. Automatica IFAC 19(6) , 759 - 766 . Hayes- Roth, F. , D. A. Watermann and D. B. Lenat (Eds. ) (1983) . Bui 1 ding Expert Systems . Addison -Wesley, Reading , MA. Heidorn , G. E. (1975). Augmented phrase structure g r ammars . In B. L . Nash- Webber , R. C. Sc h a n k (Eds.) . Theo r etical Issues in Natura l Language Processing , ACL , Menlo Pa r k , CA . E.D . Sacerdoti , D. Saga l owicz and He n drix , G. G. , Develop i ng a natural lanJ. Slocum (1978). guage interface t o complex data. ACM Trans. on Data Base Systems 3(2) , 105-1 47. Jones , M.A . (1983). Activation-based pars l ng. Proc. 8 t h Int . Jo i nt Conf. o n Arti fi cial Intel l igence , Karlsruhe , FRG , 678-682 . Kaplan , S . J. (1983) . Cooperative responses from a portab l e natural language data base query system . In M. Brady, R. Bersw i c k (Eds . ) , Computational Models of Discourse . Th e MIT Pr ess, Cambridge, MA. King , J . J. (Ed . ) (1983) . Specia l issue on AI and data base research . ACM Sigart Newsletter 86 , 32-72 . Mancini , G. , J . Amesz , S. Capobianchi and P . Bastianin i ( 1982). ERDS: An organ ized in f orma tion exchange on t h e operat i on of the European Nuc l ear Reactors. Int. Con f. on Nuclear Power Exper i ence, Vienna , Austria . Pereira , F. C. N. and D.H.D . Warren ( 1980). Definite clause grammars for language analysis A survey of th e formalism and a comparison wi th augmented transi ti on networks . Arti fic ia I Inte ll igence 13 , 23 1- 278 . Phillips , B. (1984) . An object - oriented parser. In B . G. Bar a , G. Guida (Eds . ) , Computationa l Mode l s of Natural Language Processing, NorthHol l and , Amste r dam , NL , 297 - 321 . Prade , H. (1983) . A synthetic view on approx i mate reasoning techniques. Proc. 8th Int . Joint Conf. o n Artificia l I ntelligence , Karlsruhe, FRG , 130- 136. Reichman-Adart , R. (1984) . Extended person-machine interface . Artificia l Intelligence 22, 157218 . Reiter , R., H. Gallaire, J.J. King , J . Mylopoulos and B.L. Webber (1983) . A panel on AI and data bases . Proc . 8th Int. Joint Conf . on Artifi cial Intelligence , Karlsruhe , FRG , 1199- 1206. Sk ala , H.J. , S. Termini and E. Tri ll as (Eds . ) (1984). Aspects of Vagueness . Reidel , Dordrecht , NL . Rieger (1982) . Parsing and Sma 1 1 , S. L . and C. comprehending with word experts (A theory and In W. Lehnert, M. Ring l e its realization). (Eds.) , Strategies for Natural Language Processing , Lawrence Erlbaum , Hillsdale , NJ .