<?xml version="1.0" standalone="yes"?>
<Paper uid="A83-1005">
  <Title>IR-NLI : AN EXPERT NATURAL LANGUAGE INTERFACE TO ONLINE DATA BASES (o)</Title>
  <Section position="4" start_page="31" end_page="32" type="metho">
    <SectionTitle>
3NLINPS INFO2MATION RETRIEVAL
</SectionTitle>
    <Paragraph position="0"> in :his section we present an application domain ;here the topic of natural language reasoning plays fundamental role, namely, natural language access to online information retrieval services. As it is well known, online services allow interested users zo solve information problems by selecting and re~riev~ng reievant io~uments stored in very large bibilogr%phic sr f~ctual data bases. 3enerail~&amp;quot; end-users ire unwilling or unable\] ~o serach ~ersonai!y and iirectly access these large files, but they Dften rely 3n the ~ssistanee ~f e skilful information professional, the intermediary, who h.ws how tc select e~prtpriate data bases end hr.. to design good search ~%artegies for the retrieval of the desired information, and how ~o impiement them in e suitable formal luery !an~/age. Usually, the interaction between end~=er %n~ intermediary begins ~ith ~ presearch intervlew aimed e~ precisely clarifying the content and t .e Db.jecti'zes of nhe information need. On tha base zf the information gathered, the inzeremdiary chooses  the most suitable data bases and, with the nei~ :f seraching referr~l aids such ~s thesauri, iirecncties, etc., he devises the search ~trate~# no 0e executed by the information retrieval system. The output of the search is then evaluated by the enduser, who may propose K refinemen~ ~nd an interaction of the search for better matching hi3 requests. We claim that the intermediary's task represents a good example of the issues of natur~i ian~aage reasoning, part~icularly for what concerns the ebliity ~f understanding natural language user'3 :'equest~ an! ;:' reasoning on their linguistic and aemanzlz z'..nn:e~ in order :o fully :~pture user's nears ~nd gczl~.</Paragraph>
    <Paragraph position="1"> Besides, it has to be stressed that ~he intermedizrv should also posses other important skills, nh..: i_ expertise ~nd precise knowledge ~bout ia~a ba~e cantent, organization, and indexing criteria, abcu~ availability and use of searching referral ai!s,abBut system query languages and ~coess procedures, znd last about how %o plot ~nd construct en adequate search strategy. The above illustrated :hrzcteriszlcs motivate the design of a natural Language expert ~ystem for interfacing online ~ata bases. 3n fact, the !R-NLI project has among its long term goals the i~ plementation of a system to be interposed between the end-user and the information retrieval system, capa ble of fully substituting the intermediary's role.</Paragraph>
  </Section>
  <Section position="5" start_page="32" end_page="32" type="metho">
    <SectionTitle>
IV THE IR-NLI SYSTEM
</SectionTitle>
    <Paragraph position="0"> IR-NLI is conceived as an interactive interface to online information retrieval system suppoz~ing English language interaction. It should be able to manage a dialogue with the user on his information needs and to construct an appropiate search strategy.</Paragraph>
    <Paragraph position="1"> More precisely, IR-NLI is aimed at meeting the needs of non-technical users who are not acqua/nted with on line searching. For this purpose three different capabilities are requested. First, the system has to be an expert of online searching,i.e, it must embed knowledge of the intermediary's professional skill.</Paragraph>
    <Paragraph position="2"> Second, it must be able of understanding natural language and of carrying on the dialogue with the user.</Paragraph>
    <Paragraph position="3"> Third, it has to be capable of reasoning on language in order to capture the information needs of the user and to formulate them with appropriate terms in a given formal query language.</Paragraph>
    <Paragraph position="4"> In the current first phase of the project we have considered a set of working hypotheses for IR- null NLI : - it operates on just one data base; - it utilizes only one query language; - it refers to only one subject domain; - it is conceived only for off-line use without in- null teraction with the data base during the search session.</Paragraph>
    <Paragraph position="5"> fire suitable sequences of understanding, dialogue, and reasoning functions until the internal repre.sentation of the user's requests is completely expanded and validated.</Paragraph>
    <Paragraph position="6"> The UNDERSTANDING AND DIALOGUE module is devoted to perform activities mostly of linguistic concern. First, it has to translate the natural language user's requests into a basic formal internal representation (IR). Second, it manages the dialogue with the user by generating appropriate queries and by translating his replays,thus expanding the IR with new information. The UNDERSTANDING AND DIAILGUE ~odule utilizes for its operation a base of lin~uiszic kno.led~e (LK).</Paragraph>
    <Paragraph position="7"> The REASONING module is aimed at reasoning on IR in order to enlarge its content with all the information required to generate an appropiate search strate~\[. It utilizes for this task a base of domain specific knowledge (DSK).</Paragraph>
    <Paragraph position="8"> The FORMALIZER module, after the STRATEGY GEf~-RATOR has completed its activity, constructs from the IR the output search stra~e~ to be executed for accessing the online data base. The FORMALIZER utili zes for its operation knowledge about the formal fan guage needed to interrogate the online data base and operates through a simple syntax-directed schema, it is conceived as a parametric translator capable of producing search strategies in several languages for accessing online services, such as SDC ORBIT, Euronet DIANE, Lockheed DIALOG, etc.</Paragraph>
  </Section>
  <Section position="6" start_page="32" end_page="33" type="metho">
    <SectionTitle>
&amp;quot;! SYSTSM ARCHITECTURE
</SectionTitle>
    <Paragraph position="0"> The general architecture designed for the IR-NLI system is shown in Figure 2. The kernel of the system is constituted by the 3~dEI'ZGY GENERATOR , which is devoted to devise the top-level choices conzerning the overall operation of the syszem and to :cntroi their execution. It utilizes for it~ acti&amp;quot;'ity a base 3f expert knowledge (E~K) which concerns the evaluation of user's requests, the managament 3f the presearch interview, the selection of a suitable approach for ~eneration of the search strategy, and ~uheduling of the activities of the lower Level modules :~ERSTAINDING AND DIALOGUE, REASONING, and FORMALIZER. The operation of the STRATEGY GENERATOR is organized around a basic sequence of steps, each taking into ~ccount a differnt subset of expert rules, that r%ppiy tO different situations and</Paragraph>
    <Paragraph position="2"/>
  </Section>
  <Section position="7" start_page="33" end_page="34" type="metho">
    <SectionTitle>
Vl KNOWLEDGE BASES
</SectionTitle>
    <Paragraph position="0"> In this section we shall illustrate the main features of the three knowledge bases utilized by the IE-NLI system.</Paragraph>
    <Paragraph position="1"> Let us begin with DSK. The purpose of this kno K ledge base is to store information about the domain covered by the online data base to which IR-NLI refers. This information presents two ~spects : a semantic facet concerning what concepts are in the data base and how they relate to each other, and a linguistic one concerning how the concepts are currently eta'pressed through appropriate termS. The structure of DSK proposed reflects and generalizes to ~ome extent that of classical searching referral aids (in particular, thesauri and subject classifications). At a logical level, it is constituted by a labelled directed network in which nodes represent concepts and directed arcs represent relations between concepts. Each node contains a ~erm., a fla~ denoting whether the term is controlled or not, a field that stores the post in~ count, i.e.</Paragraph>
    <Paragraph position="2"> the number of items in the data base in which the term appear, and a level number which represents the degree of specificity of ~he term in a hirar:hical subject classification. Arcs g'n~erai!y denote the usual cross-reference relationships utilized for struc=uring thesauri; e.g., BT (broader term), (narrower term), RT (related term), UF (used for).</Paragraph>
    <Paragraph position="3"> In addition, arcs of type ne~ are provided that al-Low, in connection with the level numbers of nodes, sequential scanning of the knowledge base accoriing to the currently utilized hierarchical subject ~iassification. This s~ructure is conceived to be ilrec~ly obtained (possibly in a partially automatic way ~hrough appropriate data conversion programs) from ~vailable searching referral aids and online thesauri.</Paragraph>
    <Paragraph position="4">  Le~ us turn now to LK. This knowledge base is ~imed at supplying all information concerning natural language ~ha~ is needed to understand user's reluests. According to the mode of operation of =he 'XDERSTANDING A~D DIALOGUE module (see section IX), it 2on%ains the lexicon of the application domain which is currently considered. Each record of the lexicon 2ontains ~ word of =he language, its sem:~nt~2 ~2~.~e concept, ~onnec~ive, f'anction~, ~nd its ~e%nin~. The semantic type denotes ~he role ~f % ~ord in a sentence; namely: - ienoting ~ term of zhe da=a base; - iefining z parz icuiar relation between different 2.\]nc~ta in user's requests; - specifying ~ particular function that the user ie~ires ~o obtain from the information re.ri_,al~ =z  ~ystem.</Paragraph>
    <Paragraph position="5"> The meaning 9f R word may be expressed zs ~ pointer  to a term of the DSK in the case of a word of type concept, as a special purpose procedure in ~he case of a connective or a f~nc:ion.</Paragraph>
    <Paragraph position="6"> Let us note that,in order to avoid ~nuseful duplication of informa=ion in =he DSK and LK, a shared directory of en~r y words may be u~iiized f~r boLh bases.</Paragraph>
    <Paragraph position="7"> The purpose of EK is to contain information tha= concerns the professional expertise of the intermediary on how to manage a search session in 0rder to appropriately satisfy the information needs of the end-user. Its contort= is made up sf several classes of rules concerning the different kinds 3f activities performed during a search session. 3he general s~ructure of the rules is of the ciassical type !F-THEN.</Paragraph>
    <Section position="1" start_page="34" end_page="34" type="sub_section">
      <SectionTitle>
Vll STRATEGY GENERATOR
</SectionTitle>
      <Paragraph position="0"> The task of the STRATEGY JENEEATCR can be considered from two differen~ poinzs of view :  - an external one, tha= concerns performing in~ermedia~j's activity; - an in=ernal one, that rela=es to management and control of REASONING and U~DERSTANDiI;G ~ 31ALC-GUE modules.</Paragraph>
      <Paragraph position="1"> On ~he base of these specifications, it mus~ embed exper~c capabilities and behave Rs a consultation system for information retrieval \[?oli:t,~9@: ..... basic mode of operation of this =odtule is ~rganized around the following four main steps tha~ reflect the usual practice of online information searching (Lancaster,~979; Meadow, Cochrane,~Od~.</Paragraph>
      <Paragraph position="2"> ~. perform presearch interview 2. select approach 3. devise search startegy h. construct search s~ar~egy.</Paragraph>
      <Paragraph position="3">  The IN adopted is unique throughou: the whole operation of the system and it is :onszitu:e/ by z frame, initialized by the UNDERSTAS~ING ~ :IALSOUE modu_le,and then further refined and expanded cy the reasoning module. This fr~-me i~ &amp;ErucEured in%z ~ucframes in such % way no :ontain, :!zssi:'~ei un!er different headings, any information ~ha% is relevant for searching an online data base, and ~3 zii;w an effective pattern-matching for the seleczizn cf search approaches and tactics. More ~ecifizziiy, l? encompasses terminology about zoncepts and facetz present in user's requests, c~-ifi=~tizns about search constraints and output forma~, ~nd fi~lres about search objectives such &amp;s recall and ~recisi~n ~Meadow, Cochr~ne,1981~.</Paragraph>
      <Paragraph position="4"> To go further in our description, let us introduce precise definitions of two technical terms above used in an informal way : search approach : the abstract way of facing a search problem, reasoning on it, analyzing its facets, and devising a general mode of opera~ion for having access to desired informazion stored in an online data base; search tactic : a move, a single step or action, in the execution of a search approach.</Paragraph>
      <Paragraph position="5"> Let us recall that a search strafe@D, is a program, written in an appropriate formal query lan~aa~e, for obtaining desired information from an online system; taking into account the two above definitions a search strate~j can be viewed as the result ~PS the execution of a search approach through application of appropriate ~earch tactics.</Paragraph>
      <Paragraph position="6"> Within IR-NLI, a search approach is represented as an al~orithm that defines which tactics to utilize, among the available ones, and how to use them in the construction of a s~rate~. An approach is not however a fixed procedure, since it does not ~pecify at each step which paz~cicular tactic to execute, but only suggests a set of candidate tactics, whose execution may or may not be fired. Tactics are represented at two different levels of abstraction : - an high-level representation &lt;name, objectives&gt; provided for ~se by the STRATEGY GENERATOR; - a low-level representation &lt;name,reasoning actions&gt; supplied for use by the REASONING modLtle.</Paragraph>
      <Paragraph position="7"> About ~5 tactics ~re considered, taken from the very rich discussion by Bates (\]979) : CORRECT, CUT, SELECT, EXHAUST, .REDUCE, PARALLEL, PINPOINT, BLOCK, $b~ER, SUB, RELATE, NEIGHBOUR, FIX, RESPELL, RESPA2Z, etc.</Paragraph>
      <Paragraph position="8"> The operation of the STRATEGY GENERATOR is basically pattern-directed; namely, the particular activities to be performed and the way in which UNDERSTANDING ~D DIALOGUE and REASONING modules are activated are determined by the content of the current IR *or of some par~cs of i~), which is matched with zn appropriate subset of the exper~ rules. In :his way !~3 mode of ~peration is not strictly determinate : ~ome %ctivities may or may not be fired .r may be perfDrmed in !ifferent ways according ~o  &lt;pattern-cna~ching with the cu.rr~nt It? and selection of aubfr~nes tha~ could be appropriatel~ filled up with new info~naeion&gt; a~iv~e UNDERSTANDING AND DIAbgGUE &lt;engag~nent of suitabZe dialogue for gather~ing additional information about search convent and objectives : concepts, limitations, conattaints, e~uaions, decider precision ~nd reca~Z&gt; &lt;aa~ansion of IR&gt; seZect approach &lt;selection of the approach which best fits search objeccive8 through pattern-marching bemaeen IR and high-level represencation of ~a~ics involved in each approach&gt; devise search strategy &lt;pa~tarn-ma~ohing bevween vhe current IR and tactics involved ~n the selected approach&gt;  As already mentioned in section IV, in the first version of IR-NLI the off-line operation :f the system lead us go consider only the buiidin~ block approach; future versions of the system viii encompass also other classical and ~ommoniy u~ilized approaches such as successive fraction, zita :ion pearl growing, most speclfic facet flrsn, ~nc. {Meadow, Cochr~e,1981), ~hac are more ~uizab!e fJr an on-line operation of the ~ystem in which iicezt interaction with the data base luring the ae~rcn session is allowed.</Paragraph>
    </Section>
  </Section>
  <Section position="8" start_page="34" end_page="35" type="metho">
    <SectionTitle>
VIII REASONING
</SectionTitle>
    <Paragraph position="0"> The .ctivity 3f the STRATEGY GENERATOR can now he repr:sented in % more ~etaiied way through ~he fsllowing high-lave! program : The REASONING module operates on the IR ani za aimed at precisely capturing user's ~cals ~nd neela through deductive and inferenniai processes. More specifically, the REASONING uodule has the main ~a~k of executing the tactics fired by the S.-?..ATE-GY GENERATOR. Hence, from this point cf view, it  represents the actuator of the reasoning process devised by the STRATEGY GENERATOR.</Paragraph>
    <Paragraph position="1"> It utilizes the low-level representation of the tactics, which specifies suitable reasoning actions expressed in terms of : - accessing DSK; - updating the I~ with the new information.</Paragraph>
    <Paragraph position="2"> Among the basic capabilitis of the ~ESONING module we consider generalization to broader terms, extension to related concepts, particularization to narrower terms, analysis of synonymi and homonymi, etc. its operation is based on special-purpose procedures that correspond to the reasoning actions involved in the tactics. Furthermore, when an action has to be performed on IR for extending its content, validation may be requested from the user in order to ensure a correct matzhing betvwen his wants and system proposals. This is done through the U~DERSTANDING AND DIALOGUE molule which has to gather user's agreement about the new terms to be introduced in the IR.</Paragraph>
  </Section>
  <Section position="9" start_page="35" end_page="35" type="metho">
    <SectionTitle>
ZX U~DERSTA/~DING .~D DIALOGUE
</SectionTitle>
    <Paragraph position="0"> Zdr.e purpose of the UNDERSTANDING AND DIALOGUE module is twofold : - =o ~ranslate user~ requests into IR; - to generate ~ueries to the user and to understand his ~nswers, i.e. to manage a bounded scope diaLogue. null The conception of this module strongly relies, fDr what ~oncerns ~he ,anderstanding function, un ~he experience previously developed by the authors vith \]\[LI project, and is organized around the con:epts of semantics-directed and goal-oriented parsing ~Guida,Tasso,IQ82a).its mode of operation is ma~niy rule-based : ~ main parsing algorithm perfor=Is the most elementa~/ steps of the ~nalysis search in the lexicon, construction of a basic  ~%lon iz the ability to solve critical situations by engaging the user in a clarification iizlDgue +-ccivated by some of the above mentioned heuristic rules, to gather additional information which is zecessar}&amp;quot; to correctly ~inderstand the input natur%i \[~nguage requests.</Paragraph>
    <Paragraph position="1"> For what concerns the dialogue function, it relies on two strictly connected activities : - generation of a lue~j, according to some requests from the STRATEGY GENERATOR ~r HEASONIIIG modules, through assembly and completing of parametric =e:~ fragments stored in the UNDERST.~ND!NG '~D DIALOGb~ module% - understanding of the user's answer and refinement, i.e. validation, updating or completing, of the current IR.</Paragraph>
    <Paragraph position="2"> Let as s~ress that, according to the basic goal-oriented conception of the parsing mechanism of</Paragraph>
  </Section>
  <Section position="10" start_page="35" end_page="35" type="metho">
    <SectionTitle>
&amp;quot;~TDERSTANDING AND DIALOGUE module, the ur.ders~u-
</SectionTitle>
    <Paragraph position="0"> ding activity performed in the frame of :he diaiogue function is strongly directed by knowledge of the query tha~ the system has asked the user ~nd, therefore, of expected information to be zap~ured in the answer.</Paragraph>
  </Section>
  <Section position="11" start_page="35" end_page="36" type="metho">
    <SectionTitle>
X A SAMPLE SEARCH SESSION
</SectionTitle>
    <Paragraph position="0"> In this section we present % short example of the basic mode of operation of IE-ZLI. Figure shows a sample ~ession in which, in ~ddition to the user-system dialog/e, parts of the .'R and the search strategy generated (in Euronet DI.L\'~ EL~O-LANGUAGE) ~re reported. The -~xample refers to the domain of computer science.</Paragraph>
    <Paragraph position="1"> iCi CONCLUSIOn; In the paper the main features of =he ZR-~;LI system have been presented. The projec~ is now entering the experimental phase =hat will be carried on on a VAX 11,'7~0 syszem.</Paragraph>
    <Paragraph position="2"> The design activity so far !evelc~ed ,3uiia, Tasso,1982b,1982c) has reached, 2n cur mind, a quite assessed point so that fuzure work :n -his norA-&amp;quot; will oe mainly concerned wi-h r~mo':~i -f -he restric =ions and working hypotheses -'onsiaered in the current first ~nase and with refinemenr. ::' \[z~/emen=~%ion de=aLia. ~he authors also g~+-n %: \[aziememt _n the next future a-'omple~.e prototype versicr. ~f =::z system to be -'onnected to a real ;nlane S~:s=e~ in &amp;quot;-he fro-me .~f a strictly application ~rle..ted interest .</Paragraph>
    <Paragraph position="3"> The research activity will be focused, .'n -he ocher hand, on several issues c'::z= \[i,~-r'--~ f-~her</Paragraph>
  </Section>
  <Section position="12" start_page="36" end_page="37" type="metho">
    <SectionTitle>
&gt; SEARCH STRATEGY \]ENERATED :
I. FI?D I:~ELLIGENT.TERMINALS
2. FInD ONLINE OR ON.LI~ O__RR INTERACTIVE O__RR REALTIME OR REAL.TIME
FInD INFORMATION OR REFERENCE OR BIBLIOGRAPHICS OR RETRIEVS OR SEARCH$
2. FInD COSTS OR PRICE
5. FI~ S=~ A~ S=2 ;~ND S=3 A~ S=4
6. ?RI2~ FI
&gt; END OF SEARCH SESSION, BYE
</SectionTitle>
    <Paragraph position="0"> investigation. Among these we mention : - the development of more flexible and robust dialogue capabilities, including limited justification of the mode of operation of ~he system (Webbet,1982); null - the study of advanced representations of ~actics through generalized rule structures that will allow more refined matching and firing mechanisms (Winston,198E); - the design of new tactics (e.g., PATTERN, ~ECORD, BIBBLE (Bates,~979)) ~nd reasoning actions, that enable the system to keep track of previous search sessions ~nd to ~nalogize from experience in devising and executing a search approach.</Paragraph>
  </Section>
class="xml-element"></Paper>