<?xml version="1.0" standalone="yes"?>
<Paper uid="C80-1083">
  <Title>References</Title>
  <Section position="1" start_page="0" end_page="0" type="metho">
    <SectionTitle>
GOAL ORIENTED PARSING: IMPROVING THE EFFICIENCY OF
NATURAL LANGUAGE ACCESS TO RELATIONAL DATA BASES
</SectionTitle>
    <Paragraph position="0"> This paper is devoted to present a new approach to natural language understanding which is called here goal-oriented parsing. The interaction in natural language with artificial systems (robots, data base systems, program generators, question-answering systems, etc.) does not require in most cases of actual interest a full (human like) comprehension of natural language in all its details and nuances. A partial understanding is often enough ,wich extracts from the natural language expressions the only significant information which is necessary to construct a correct formal input for the target system. In such a model of comprehension the same meaning is assigned to several different natural language expressions, thus defining a many-to-one mapping between natural language sentences and corresponding formal representations. We argue that a bounded scope, restricted, goal-oriented understanding of natural language may greatly increase the efficency of representation models and parsing algorithms, thus allowing the construction of effective systems. This claim is supported by the design and implementation of a natural language interface to a relational data base called NLI and developed at the Milan Polytechnic Artificial Intelligence Project. In the paper the architecture of the system, the linguistic models, and the parsing algorithms are presented and illustrated through selected examples. Promising directions for future research are outlined as well.</Paragraph>
    <Paragraph position="1"> l.lntroduction The development of natural language understanding systems has received in the last years a growing interest, both in the area of computational linguistics and artificial intelligence. In this field a lot of running systems have been implemented which are based on different linguistic models and parsing algorithms. This paper is devoted to illustrate NLI, a natural language understanding system which has been developed at the Milan Polytechnic Artificial Intelligence Project for the inquiry in Italian of a relational data base. NLI is based on a new approach to natural language understanding called here goal-oriented comprehension. It is claimed that the understanding activity may be correctly defined only if the purpose or goal of the comprehension is clearly specified, and that the same sentence may have several different meanings depending on the purpose for which it is considered. It follows that the complexity of the natural language understanding task depends, in addition to the inner linguistic or structural complexity of the sentence which is considered, on the complexity and extent of the scope of the understanding activity, Therefore, we can argue that, if one confines the understanding activity to well defined and bounded goals (as it is generally possible in most of the natural language processing applications), it should be possible to develop ad hoc linguistic models and parsing algorithms which are particulary fitting and efficent. Such models and algorithms will obviously strictly depend on the particular goal which is considered and will generally not hold for very different scopes. Let us outline that goal-oriented understanding implies a bounded scope, incomplete,and in some way diagonal comprehension, in which any detail and nuance which is not relevant to the goal is ignored.</Paragraph>
    <Paragraph position="2"> In the design of NLI the classical topic of data base inquiry has been chosen as application domain and the goal of the understanding process has been defined as that one of translating the user's natural language queries into the formal query language QBE (Query-By-Examplel5). The adopted liguistic model is strongly semantics based and doesn't take into account,as much as possible, the syntactic aspects of the sentence to be analysed I. The linguistic information needed for the parsing is stored in a lexicon and a base of the heuristics,and the parsing algorithms are strictly dependent on the particular goal which is considered.NLl has been so far developed in two versions,NLl-I and NLI-2; the latter one, which is presently in an advanced development  stage, is the subject of this work.</Paragraph>
    <Paragraph position="3"> The paper is organized in the following way: in section two the version NLI-I is shortly illustrated and the design specifications and criteria for NLI-2 are discussed; section three is devoted to present the overall architecture of the system and the parsing algorithms which have been implemented; section four shows some selected examples of comprehension; in section five some promising directions for future research are illustrated together with some conclusive remarks.</Paragraph>
    <Paragraph position="4"> 2.Previous work and experimental activity In this section the first prototype version of the system NLI(NLI-I) is shortly illustrated.</Paragraph>
    <Paragraph position="5"> The functional requirements and the basic technical decisions are then presented, on which a second version of NLI (NLI-2) is presently being developed.</Paragraph>
    <Paragraph position="6"> NLI-I has been designed and implemented at the Milan Polytechnic Artificial Intelligence Project in the years 1977-784,5 with the following aims: - to evaluate through a sample application the new model of natural language understanding proposed and the designed algorithms for semantic-based parsingl; - to study in a particular case the most critical aspects involved by the developent of actual applications,such as efficiency, memory occupations, creation and management of the dictionaries, evaluation~of the understanding capabilities, etc.</Paragraph>
    <Paragraph position="7"> The inquiry in natural language(Italian) of a toy relational data base representing the catalogue of a library has been chosen as an application domain for NLI-I. The architecture of the system has been structured in three sequential modules. The first module scanns the input sentence, searches in the vocabulary each word which has been recognized through a simplified lexical analysis, and generates an internal representation of the sentence in which the meaning of each word and elementary construct is embedded. The second module is devoted to reduce the ambiguities and to check the correctness of the structure proposed (through an algorithm based on set intersectionl); it produces a semantic tree which represents the formal internal representation of the input sentence. The third module translates the semantic three into an equivalent sentence expressed in a formal query language for relational data bases (SEQUEL).</Paragraph>
    <Paragraph position="8"> NLI-I has been implemented in FORTRAN first on an UNIVAC II00 and, later, on a Digital PDP-II/34 computer. It requires about 15 Kbytes memory with a lexicon of 200 words. The parsing of a simple sentence(a single query of about 10-15 words) needs a few seconds and more complex sentences need up to 20-30 seconds.</Paragraph>
    <Paragraph position="9"> The research activity done with NLI-I has highlighted some relevant aspects concerning the development and actual implementation of NLI, which will be further considered in the design of NLI-2. We list some of these below: - the necessity of developing an high-level diagnostic subsystem to be used by the application designer(often,the end-user himself) in the tuning activity; - the need for an interactive system for the incremental definition and for the management of the vocabulary(to be utilized also in the design of a new application versions of NLI); - the necessity of refining the parsing algorithms and of defining a base of the heuristics to be utilized in the most complex cases of comprehension; - the problem of the optimal content of the vocabulary (increasing the content of the vocabulary doesn't always improve the understanding capabilities of the system!); - the problem concerning the need of inserting in the vocabulary all the words which constitute the content of the data base.</Paragraph>
    <Paragraph position="10"> For greater details about NLI-I the reader is referred to the literature 5.</Paragraph>
    <Paragraph position="11"> The new version NLI-2 has been designed in 1979 on the base of the experience acquired through NLI-I. The main methodological assumptions, funcitonal specifications,and technical decisions on which NLI-2 is based are illustrated  below: - goal-oriented understanding; - semantics directed parsing; - hierarchical organization of the linguistic  information supplied to the system in two modules:the lexicon and the base of the heuristics; null - development of a generalized parsing algorithm independent of the content of the lexicon and of the base of the heuristics, of the structure of the data base, and of the particular natural language which is used; - developent of an high-level diagnostic module; - design of a fully interactive system for the incremental definition and dynamic management of the content of the lexicon and of the base of the heuristics and for the implementation of new versions of the system; implementation of the appropriate human engineering features needed to improving the usability of the system.</Paragraph>
    <Paragraph position="12"> --551 The natural language used for the inquiry is Italian and the output of the system is the translation of the query in QBE (Query-By-Examplel4). The data base adopted for the development of the sample application of NLI-2 is concerned with a department store description and is just the same used for the illustration of QBE (a complete definition of the data base is reported in the Appendix). In designing NLI-2 it has been decided of not inserting anything about the content of the data base in the lexicon of the system,in order to keep its dimensions as small as possible. This choice implies that the user indicates explicitely(through a pair of special symbols&amp;quot;) in the input sentence the words which denote values of domains of the data base. The adequacy and effectiveness of this decision will be evaluated during the experimental activity to be done with NLI-2 and could be removed if considered unappropriate. null The next section will be entirely devoted to the illustration of the global structure of NLI-2 and of the parsing algorithms which have been developed.</Paragraph>
    <Paragraph position="13"> 3.System architecture and parsing algorithms NLI-2 is based on a modular structure in which there is a distinct division between the data(vocabulary,data base model, formal query language structure) and the programs(analyser,generator, vocabulary management). The system architecture is illustrated in Figure I.</Paragraph>
    <Paragraph position="14"> The monitor is devoted to open the job(it informs the user about the system capabilities and present to him an appropriate menu of different options), to interprete the user's commands, and to manage the activities of the lower modules.</Paragraph>
    <Paragraph position="15"> The user's control messages accepted by the moni- null tor are: - SHELP : detailed informations and instructions on the operation of the system are supplied; - SU : the understanding cycle of the natural language queries is activated; the system is ready to run in U-mode; - SV : the module for the management of the vocabulary is activated; the system can be used for the definition,updating,and tuning of the vocabulary in the V-mode; - $STOP : returns the control to the monitor for U-mode and V-mode operations; - SEND : closes the job.</Paragraph>
    <Paragraph position="16">  The analyser, which is activated by the monitor when running in the U-mode,accepts the natural language query in input and generates an appropriate internal representation or, if some step of the parsing fails, a diagnostlc message to be displayed to the user. The analyser operates in a cyclic way: when the parsing of a sen- null tence is concluded it is automatically reset and is ready to accept a new query. Queries are not stored by the parser, so that the user is not allowed to carry on a dialog with the system in which new queries can refer to old ones(or to their answers). The module for the vocabulary management can be activated by the monitor or directly by the analyser when a fail occurs. It supplies the user with some basic functions(some of them mainly relating to text editing) for a flexible management of the vocabulary(search, display,delate,add,update,etc.).The generator receives in input the internal representation produced by the analyser and translates it into the formal query language QBE.</Paragraph>
    <Paragraph position="17"> Let us describe now the internal structure of the vocabulary. It is composed of two parts: the lexicon~ which contains the words and the elementary construct referring to the applica.tion domain in which the system operates,and the base of the heuristics , which embeds further linguistic information(sometimes also of syntactic nature) necessary to understand complex and often ambiguous constructs.</Paragraph>
    <Paragraph position="18"> The lexicon is organized in 26 alphabetic groups,each one of them contains all the words beginning by the same letter. Each word (or elementary construct) is bound to all its possible meanings,thus yelding a record of the lexicon. A record is in fact composed of two parts: - the left part contains a word or a (short) sequence of words representing a simple construct; morphology is (generally)not taken into account in the parsing,but the wordsare represented in such a way to be recognized in all possible forms(conjugated verbs,inflected nouns, ect.); - the right part embeds the representation of the semantics of the word within the application domain which is considered.</Paragraph>
    <Paragraph position="19"> The right part of each record has a different structure depending on the semantictype to which the word stored in the left part belongs. In our application a word can denote three different types of information concerning: the identification of the relations and of the domains involved by the query; the logical connection between relations and domains; the specification of the required output.</Paragraph>
    <Paragraph position="20"> Therefore, in relation to the three above presented activities, we define the following semantic types: object, connective, and function, respectively.</Paragraph>
    <Paragraph position="21"> The general structure of an object record is: li I 111 12222 0 N1.LI=N2.M ~ ..... N1.MI=N2.M ~ where:  - P denotes a word or a simple costruct; - 0 denotes that the record is of type object; - each pair N~.M~ denotes that the wo~d P refers  to the domain IN~ in the relation M~ (relations and domains are ~epresented by positive inte gers); the = symbol separates equal domains belonging to different relations; if P refers to a relation without specifying the domain, N~ is replaced by the special symbol $;each f~eld of the record contains a different meaning of the object P.</Paragraph>
    <Paragraph position="22"> The structure of a connective record is: where: P denotes a word or a simple costruct; C denotes that the record is of type connecti-</Paragraph>
    <Paragraph position="24"> type of the adjacent words),Y: denotes(through I a pointer) the function which must be applied or the action which must be performed during the parsing to take correctly into account the meaning of the connective P.</Paragraph>
    <Paragraph position="25"> The structure of a function record is: where : - P denotes a word or a simple construct; - F denotes that the record is of type function; - each Z. denotes a possible meaning of the function ~ and represents (through a pointer) an action which must be performed in the parsing. The base of the heuristics is constituted by a bag of heuristic rules(of the type precondition-action), which allow to represent linguistic informations which are not comtained in the lexicon but are still needed for understanding complex sentences and for the resolution of ambiguities. The heuristic rules are selected and activated during the parsing,whenever a critical situation occures,on the base of a pattern directed invocation algorithm.</Paragraph>
    <Paragraph position="26"> Let us outline that the vocabulary(both its structure and content) is strctly dependent on  --553-the particular application to which the system is devoted. This feature,which represents a quite rigid bound to the flexibility of the system, is, on the other hand, a straightforward consequence of the concept of goal-oriented understanding. The issue of designing system architectures which allow a flexible handling of the purpose and domain of the comprehension is considered as a promising and ambitious topic for future research,as it will be illustrated in the conclusions. null Let us illustrate now the activity of the analyser. It accept as input a natural language query and supplies an internal formal representation of it which is not far from a QBE expression; in such away the role of the generator is confined, within NLI-2, to a simple editing activity. The activity of the analyser can be splitted in four steps: I. scanning of the natural language input sentence, search in the lexicon,and generation of  a first-level internal representation; 2. parsing,(partial or full) resolution of the ambiguities, and generation of the second-level internal representation; 3. intersection,i.e, verification of the consistency of the proposed structure and resolution of the possibly remaining ambiguities; 4. generation of the correct formal representa null tion or, if any fail has occured, of the appropriate diagnostics.</Paragraph>
    <Paragraph position="27"> In step I. the input sentence is first scanned and an internal representation of it is constructed( first-level internal representation).The searching of the words in the lexicon is indexed sequencial(with a one-level index); if a word or elementary construct is found in the lexicon it is replaced by all the record to which it refers; on the other hand, if a word is not recognized, it is enclosed between brackets in the first-level internal representation and it is successively ignored in the parsing. The words which denote values of domains,i.e, which appear in the content of the data base, are preceded and followed by the special symbol&amp;quot; in the input sentence and remain unaltered in the internal representation. The first-level internal representation reflects the ordering of the words in the input sentence and embeds all the linguistic information which can be obtained from the lexicon. null In step 2. the relations involved by the user's query are first recognized through a detailed analysis of the words of type object. This activity cannot yeld, in general, a definite and unambiguous result since some objects may refer to different relations. A pattern directed invocation of appropriate heuristics may contribute in eliminating some (or also all)of the ambiguities. After the correct relations are individuated, their logical schemata are extracted from the data base model and a new phase starts aiming to individuate the domains referred to in the input sentence. The possible ambiguities may be resolved through an appropriate use of the heuristics. The domains which have a role in the user's query are then labelled in order to be further processed in the following steps of the parsing. The words enclosed between pairs of &amp;quot; symbols are then considered in order to find the appropriateassignement of these values to the relating domains. Different criteria may be used to perform such an assignement(e.g., the contiguity of a value and an object),but, in any case, the assignement is not a definitive one until it is confirmed by an heuristic or by the following step 3.(intersection), A tentative interpretation of the input sentence(second-level internal representation) is then produced,which will be further refined and completed in the following steps of the parsing.</Paragraph>
    <Paragraph position="28"> The analysis of the connectives constitutes the kernel of step 3.;it is organized in two phases: singling out the correct semantics of a connective on the base of its position in the sentence and of the type of the words to which it refers(appropriate heuristics have to be utilized to resolve possible ambiguities); verification of the proposed interpretation of the sentence segment to which the connective belongs(this activity is performed through set intersection algorithms I, what gives the reason for the name assigned to step 3.).</Paragraph>
    <Paragraph position="29"> In step 4. the functions~which possibly appear in the input sentence, are considered and the domains to which they apply are individuated. If the parsing terminates correctly the formal representation of the sentence is generated which will be later translated in QBE by the generator. In the following section some complete examples of comprehension are illustrated.</Paragraph>
    <Paragraph position="30"> 4.Parsing sample sentences In this section we are going to present some examples of parsing extracted from the sample application in which NLI-2 is presently working (the department store data base which is described in the Appendix).</Paragraph>
    <Paragraph position="31">  In step 2. the relations involved by the user's request are first recognized,</Paragraph>
  </Section>
  <Section position="2" start_page="0" end_page="0" type="metho">
    <SectionTitle>
&lt;VENDITE (REPARTO,ARTICOLO) &gt;
&lt; FORNITORI (ARTICOLO,FORNITORE)&gt;
</SectionTitle>
    <Paragraph position="0"> and, then, the domains referred to in the input sentence are labelled through the letters A and  PV is a pointer to a routine which puts the special symbol P in the domain which follows the function TR~A. PA is a pointer to a routine which assigns X; to domain 2 of relation 2 (in the case VENDONO),or to domain 2 of the relation 3 (in the case FORNITI DA); PA is activated if the connective to which it is bound connects an object (0) to a particular value of a domain(V), as it arrives in both cases VENDONO and FORNITI DA. The second meaning of these two connectives, (I) find the departments which sell products supplied by Parker represented by the pointer PR, is not considered in this parsing.</Paragraph>
    <Paragraph position="1">  In this representation there are two domains(SA-LARIO and DIRIGENTE) whose role is not yet understood; this pattern activatesaparticular heuristics which allows two obtain the following cor- null Step 4. fails since the assignement of the special symbol P to the domain NOME, which is required by the function VORREI (pointer PV), can not be executed, being the value &amp;quot;CANCELLERIA&amp;quot; already present in the domain NOME. The following diagnostic message is therefore generated: &gt;analisi fallita nella fase 4 &gt;impossiblit~ di effettuare un corretto assegna&gt;mento del valore &amp;quot;CANCELLERIA&amp;quot; ad un opportuno dominio &gt;termini ignorati nell'analisi: I, CHE, LAVORANO, NELLA, DIVISIONE &gt;rappresentazioni interne generate: (4) (vedi sopra) Let us note that,if we insert in the lexicon the new record DIVISIONE/O/I.4=I.2, it is possible to obtain the following correct parsing:  assigned to an appropriate domain words ignored in the parsing: I, CHE, LAVORANO, NELLA, DIVlSIONE internal representations generated:</Paragraph>
  </Section>
  <Section position="3" start_page="0" end_page="0" type="metho">
    <SectionTitle>
&lt;FORNITORI (ARTICOLO, FORNITORE:&amp;quot;OROLOGI&amp;quot;)&gt;
</SectionTitle>
    <Paragraph position="0"> Step 4. fails because of the same reasons as in Example 3. In this case, however,the system can correctly parse the input sentence only by inserting in the lexicon the word OROLOGI which, on the other hand,is part of the content of the data base.</Paragraph>
    <Paragraph position="1"> 5.Conclusions In the paper the NLI system has been illustrated, with particular attention to the NLI-2 version. It is devoted to the inquiry in natural language (Italian) of a relational data base. NLI is based on a modular architecture which allows the designer , or even the user, to define his own application in a fully inte.ractive and incremental way. The linguistic model supported by the system is that one of goal-oriented understanding,and the parsing algorithms adopted are mainly semantics directed.</Paragraph>
    <Paragraph position="2"> The implementation of NLI-2 is now being completed and the system will be used in the future for developing the following research activities: null definition of performance evaluation criteria for measuring the understanding capabilities of natural language systems; evaluation of the complexity of the parsing algorithms in relation to the complexity of the input sentence and of the scope of the comprehension; null investigation on the flexibility of the system in relation to small variations in the structure and content of the data base and in the adopted natural language; definition of new system architectures allowing a full flexibility in relation to the purpose of the understanding and to the application doamin.</Paragraph>
    <Paragraph position="3"> Appendix This Appendix is devoted to present the sample relational data base which has been adopted in the development of NLI-2. This is concerned with the activity of a department store and it is just the same utilized in the definition of QBE 15. The  --556-data base is sketched below through a set of tables in which both the names of relations and domains and their internal representation(pairs of integers) is given.</Paragraph>
  </Section>
class="xml-element"></Paper>