hermann helbig and andreas mertens - fernuni-hagen.de

1

Hermann Helbig and Andreas Mertens

FernUniversit�at HagenPraktische Informatik VII/Arti�cial Intelligence

D{58084 HagenGermany

E{mail:hermann.helbig@fernuni{hagen.deandreas.mertens@fernuni{hagen.de

ABSTRACT

Natural language processing (NLP) is often based on declaratively represented grammarswith emphasis on the competence of an ideal speaker/hearer. In contrast to these broadlyused methods, we present a procedural and performance{oriented approach to the analysis ofnatural language expressions. The method is based on the idea that each word{class can beconnected with a functional construct, the so{called word agent, processing its own part ofspeech. The syntactic{semantic analysis performed by the word{class agents is surveyed by aWord Class Agent Machine (WCAM) used in the bibliographic information retrieval systemLINAS at the University of Hagen.

The language processing is organized into four main levels: on the �rst and second level,elementary and complex semantic constituents are created which have their correspondence inmental kernels built during human language understanding. On the third level, these kernelsare used to construct the semantic representation of the propositional kernel of a sentence.During this process, the subcategorization features of verbs and prepositions play the mainpart. The task of the last level consists in the syntactic analysis and semantic interpretation ofmodal operators in the broadest sense and of the coordinating or subordinating conjunctionswhich connect the propositional kernels.

1This paper has been presented at the Eighth Twente Workshop on Language Technology 1994 (TWLT8).See [Boves, Nijholt 94].

CONTENTS

1 INTRODUCTION 4

2 THE LINAS PROJECT 5

3 THE WORD CLASS AGENT MACHINE 6

3.1 WORD CLASS AGENTS : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 6

3.2 OPEN ACTS AND COMPLETE ACTS : : : : : : : : : : : : : : : : : : : : : : 7

3.3 WCAM STATES : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 7

3.4 AN EXAMPLE : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 8

4 GENERAL ASPECTS 9

4.1 DISAMBIGUATION PRINCIPLES : : : : : : : : : : : : : : : : : : : : : : : : 9

4.2 LEXICAL INFORMATION : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 10

4.3 KNOWLEDGE REPRESENTATION : : : : : : : : : : : : : : : : : : : : : : : 12

5 THE FOUR MAIN LEVELS OF PROCESSING 13

5.1 ELEMENTARY CONSTITUENTS : : : : : : : : : : : : : : : : : : : : : : : : : 13

5.2 COMPLEX CONSTITUENTS : : : : : : : : : : : : : : : : : : : : : : : : : : : 14

5.3 ELEMENTARY PROPOSITIONS : : : : : : : : : : : : : : : : : : : : : : : : : 15

5.4 COMPLEX PROPOSITIONS : : : : : : : : : : : : : : : : : : : : : : : : : : : : 15

6 CONCLUSIONS 16

LIST OF FIGURES

1 Overview of the NL system LINAS : : : : : : : : : : : : : : : : : : : : : : : : : 5

2 WCAM transition diagram : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 7

3 Lexical entry of Buch (book) : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 8

4 Lexical entry of schreiben (to write) : : : : : : : : : : : : : : : : : : : : : : : : : 10

5 Semantic representation of a sentence : : : : : : : : : : : : : : : : : : : : : : : 12

6 Overview of the WCAM levels : : : : : : : : : : : : : : : : : : : : : : : : : : : 13

7 The CWM structure on the �rst level : : : : : : : : : : : : : : : : : : : : : : : 14

8 OList and CList on the second level : : : : : : : : : : : : : : : : : : : : : : : : 14

9 The structure of the CWM after creating an elementary proposition on the thirdlevel : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 14

10 The structure of the CWM after constructing a complex proposition on thefourth level : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 15

LIST OF TABLES

1 Some examples of word{class agents : : : : : : : : : : : : : : : : : : : : : : : : 6

2 Some examples of semantic relations : : : : : : : : : : : : : : : : : : : : : : : : 11

Hermann Helbig and Andreas Mertens

FernUniversit�at HagenPraktische Informatik VII/Arti�cial Intelligence


E{mail:hermann.helbig@fernuni{hagen.deandreas.mertens@fernuni{hagen.de

ABSTRACT

Natural language processing (NLP) is of-ten based on declaratively represented gram-mars with emphasis on the competence of anideal speaker/hearer. In contrast to thesebroadly used methods, we present a pro-cedural and performance{oriented approachto the analysis of natural language expres-sions. The method is based on the ideathat each word{class can be connected witha functional construct, the so{called wordagent, processing its own part of speech.The syntactic{semantic analysis performedby the word{class agents is surveyed by aWord Class Agent Machine (WCAM) usedin the bibliographic information retrievalsystem LINAS at the University of Hagen.

The language processing is organized intofour main levels: on the �rst and secondlevel, elementary and complex semantic con-stituents are created which have their corre-spondence in mental kernels built during hu-man language understanding. On the thirdlevel, these kernels are used to construct the

semantic representation of the propositionalkernel of a sentence. During this process,the subcategorization features of verbs andprepositions play the main part. The task ofthe last level consists in the syntactic anal-ysis and semantic interpretation of modaloperators in the broadest sense and of thecoordinating or subordinating conjunctionswhich connect the propositional kernels.

1 INTRODUCTION

Although word{based natural language pro-cessing is not a widely used method to-day, it has been of growing interest fora couple of years. Some examples ofword{based approaches are the following:the Word Expert Parsing [Small 81], theWord{Class Controlled Functional Analy-

sis [Helbig 86], the Word{Oriented Pars-

ing [Eimermacher 86] and the analysis withWord Actors [Br�oker et al. 93].

The theory of word expert parsing (WEP)approaches the understanding of natural

2 THE LINAS PROJECT 5

LINAS

WCAM MESNET

TRANS-FER

LEX-BENCH COLEX

CL-ENGINEER

TARGETSYSTEM

USERNL-INPUT

ASSIM/INFERNET

LEX-MORPH

Figure 1: Overview of the NL system LINAS

language as a distributed process of inter-acting words. Each word is connected withan expert process which "actively pursuesits intended meaning in the context of otherword experts and real{world knowledge"[Small 87].

Eimermacher's word{oriented parsing re-lies upon the word expert paradigm, but dis-tinguishes between word{class experts rep-resenting general grammar rules and wordexperts analysing the relations between sin-gle words [Eimermacher 88].

The ParseTalk model is a concurrent,object{oriented parsing system with gram-matical knowledge completely integratedinto the lexicon. Each word of a sentencethe parser is currently working on activatesa word actor which communicates with otherinitialized word actors and other compo-nents of the system [Br�oker et al. 94].

Within the model of word{class controlledfunctional analysis (WCFA), experts are in-troduced for each word{class and the gram-matical function of a word{class is given by aprocedural speci�cation. The characteristic

feature of the word{class functions discern-ing them from other approaches is their par-tition into two di�erent functional compo-nents which are activated at di�erent timesof the sentence analysis. In this paper, theessential aspects of the word{class agent ma-chine (WCAM), which is a further develop-ment of the WCFA, will be presented. Formore detailed information of this approachsee [Helbig, Mertens 94]. A formal descrip-tion of the four main levels of the word agentbased language processing can be found in[Helbig 94].

2 THE LINAS PROJECT

The LINAS project aims at developing anatural language understanding system and,as a practical side e�ect, at providing anatural language interface to bibliographicdatabases. An architectural overview ofthose system components which are relevantto NLP is shown in Figure 1. The fol-lowing major components are distinguished:an interactive lexicographer's workbench(LEXBENCH), a morphological processor(LEXMORPH), a word{class agent machine(WCAM), a knowledge assimilation and in-ference component (ASSIM/INFERNET),and an interface module to database sys-tems (TRANSFER). In addition, there aretwo main knowledge sources: a computerlexicon (COLEX) and a network knowledgebase (MESNET).

For each word of a sentence the WCAMis currently working on the morphologicalanalysis (LEXMORPH) is activated. If thecurrent word form has no lexical entry, thecorresponding entry (the stem) can be de-termined by cutting o� in ectional syllables.


The information included in the basic entrysuch as stem, grammatical speci�cation andthe accompanying word{class agent will bereturned to the WCAM.

word-class agent characterization

*ART articles

*ADJ adjectives

*NOM nouns

*VB verbs

*PREP prepositions

*AV auxiliary verbs

*IPATTRinterrogative pronouns(in attributive use)

*IPNOMinterrogative pronouns(in nominal use)

*ADV adverbs

*GRADgraduators (adverbs ofdegree)

*POSSPR possessive pronouns

*RELPR relative pronouns

*PART participles

*COMP comparative forms

*NUM numerals

*NEG negative forms

*CONJ conjunctions

*COMMA commas

*STOP full stop

Table 1: Some examples of word{classagents

The aim of the WCAM is to analyse theNL input and to represent it by means ofa multi{layered extended semantic network(MESNET). The analysis is done by a set ofword{class agents (cf. Table 1) and the re-sult produced by the agents is stored in the

network knowledge base MESNET. The as-similation of the knowledge is organized bythe ASSIM/INFERNET component. In thecase of a question to a database managementsystem as target sytem (not to the knowl-edge base of LINAS itself), the semantic de-scription of the NL expression is translatedinto the database query language. This isdone by the translation and interface mod-ule TRANSFER.

The lexical entries of the computer lexi-con (COLEX) are organized as feature struc-tures. For typical entries see Figures 3and 4. LEXBENCH is an interactive work-bench supporting the process of creating andchecking lexical entries.

3 THE WORD CLASS

AGENT MACHINE

3.1 WORD CLASS AGENTS

As well as in the theory of WEP, the ba-sic idea of the WCAM is that words playthe main part in NLP. In contrast to WEP,which provides one expert for each word, wepresent a word{class oriented approach. Foreach word{class its grammatical function isde�ned by a procedural speci�cation, the so{called word{class agent, and each elementof a word{class will be associated with thecorresponding procedure. In LINAS morethan 35 word{class agents are distinguished.Some of them are listed in Table 1. In addi-tion to agents such as *ART (article), *ADJ(adjective), *NOM (noun) and *VB (verb),which are suggestive of the word{classes intraditional grammars, there are other agentssuch as *IPATTR and *IPNOM, which rep-resent interrogative pronouns in attributive


and nominal use, respectively. Furthermore,there are special agents responsible for thegrammatical function of punctuation markssuch as *COMMA (comma) and *STOP(full stop).

3.2 OPEN ACTS AND COM-

PLETE ACTS

The word{class agents are divided into anOPEN{act and a COMPLETE{act.2 Thisdivision corresponds to the two major ac-tivity modes of a word during NLP: on theone hand, a word opens certain valencieswhich may be satis�ed later on by otherwords or constituents of the current sen-tence. On the other hand, the valenciesopened in the �rst activity mode must besatis�ed by suitable candidates. These arethe words or constituents which have beenalready included in the analysis and whichsaturate the opened valencies. In Arti�cialIntelligence (AI) the specifying of conditions(slots) which must be �lled by speci�c data(�ller) satisfying these conditions is knownas the slot{�ller{mechanism of frames.

The two activity modes of a word are rep-resented by di�erent procedural speci�ca-tions: the OPEN{act and the COMPLETE{act of a word{class agent. Accordingto the special tasks of OPEN{act andCOMPLETE{act, the two acts are triggeredby the WCAM at di�erent moments in theanalysis.3 The division into an OPEN{

2This subdivision, however, is not made in eachcase. There are some word{class agents such as*ADV (adverb) and *IPNOM (the interrogative pro-noun in nominal use) without a COMPLETE{act.The elements of these word{classes do not open va-lencies. Their part is to saturate the valencies ofother words in the current sentence.

3The OPEN{act and the COMPLETE{act of a

NOM-OP/1ADJ-OP/1

OPEN

RETURN

COMPLETE

CLOSED

EXIT

VB-OP...

START BREAK RETURN

PREP-CO*

RELPR-CO* PART-CO*

VB-CO*

ART-CO*

NOM-OP/2*

GRAD-CO*

ADJ-OP/2*

ADJ-CO*

NOM-CO*

*

PREP-OP*ART-OP*

PERSPR-OP*

**

RELPR-OP*

Figure 2: WCAM transition diagram

act and a COMPLETE{act is one of themain di�erences to the other word{orientedapproaches mentioned in Section 1. An-other characteristic feature of the WCAMapproach is the integration of semantic inter-pretation processes into the COMPLETE{acts of the word{class agents.

3.3 WCAM STATES

In Figure 2, the WCAM transition diagramis shown. The three states of the WCAMgiven in this Figure have the followingmean-ing:

� OPEN { if a word is analysed by theWCAM, �rst the morphological compo-nent is activated in order to generatethe feature structure of the correspond-ing lexical entry. Then this structure ispassed on to the WCAM. If the currentword opens valencies, the semantic de-

word{class agent are marked by extending the nameof the agent with one of the two endings {OP or {CO,respectively (cf. the WCAM transition diagram inFigure 2).


scription of the word gets the marker OP(open) before it is stored in the workingmemory. Otherwise, if the word does notopen valencies (for instance, in the caseof adverbs, proper names, etc.), the de-scription of the word is marked by CL(closed).

� COMPLETE { in this state the COM-PLETE{act of a word{class agent is acti-vated in order to satisfy the valencies be-ing opened by the corresponding OPEN{act. Thus, in this state the decision thata group of words forms a particular con-stituent is made. For example, comingto the end of a nominal phrase (NP), the*ART{CO{act is triggered in order tosaturate the valencies being opened be-fore in the *ART{OP{act. As a result, anNP{constituent will be constructed andits complex syntactic{semantic structurewill be stored in the working memory.

� CLOSED { this state of the WCAM indi-cates the fact that the current constituentis completely analysed with all its valen-cies being saturated. The semantic struc-ture of this constituent is marked by CL,now being permitted to satisfy the valen-cies of other constituents. At the end of asentence, indicated by a triggered *VB{CO{act, the valencies of the verb willbe satis�ed. Otherwise, if there are noCOMPLETE{acts being left to perform,the WCAM immediately changes into theOPEN{state (cf. the �{transition in Fig-ure 2).

Figure 3: Lexical entry of Buch (book)

3.4 AN EXAMPLE

The word{class agents are triggered and in-spected by the WCAM. Depending on theword the WCAM is currently analysing, thecorresponding word{class agent is activated.For example, consider the following sentenceconsisting of the articles der and ein, the ad-jective jung, the nouns Mann and Buch, andthe verb schreiben:

� Der junge Mann schrieb ein Buch.

� The young man wrote a book.

The initial state of the WCAM is OPEN(cf. Figure 2). When the �rst word der

of the given sentence is included in theanalysis, the WCAM activates the OPEN{act of the word{class agent *ART, whichis called *ART{OP. The system remainsin the OPEN{state. Analysing the nextword, i.e. the adjective jung, the WCAMtriggers the *ADJ{OP{act. If there is nograduator4 between the article and the ad-

4For instance, if the article is followed by anadverb of degree such as sehr (very), the *ADJ{OP/2{act is triggered and the WCAM changes tothe COMPLETE{state.

4 GENERAL ASPECTS 9

jective, the *ADJ{OP/1{act is activatedand the WCAM remains in the OPEN{statebecause at this moment not any valenciescould be satis�ed.

Analysing the noun Mann, the *NOM{OP/2{act5 is activated and the WCAMchanges to the COMPLETE{state. Nowthe valencies opened by the article andthe adjective can be satis�ed and theCOMPLETE{acts of the adjective andthe article are activated successively. Inthe course of these COMPLETE{acts, theagreement with respect to gender, caseand number of the analysed word forms ischecked. As a result of the *ART{CO{act,the WCAM changes to the state CLOSED.In this state, the NP is completely analysedand a semantic representation of the wholephrase has been created and stored.

If an OPEN{act of a word{class agent isactivated, this information is stored in thecentral working memory (CWM) which isorganized as a stack. Thus, as analysis pro-ceeds, the corresponding COMPLETE{actswill be triggered just the other way round.

Because there are no COMPLETE{actsbeing left to perform (cf. the description ofthe WCAM{states given above), the stateof the WCAM changes into the OPEN{state. Next the verb schreiben is read andits OPEN{act will be activated. The follow-ing NP is processed by analogy to the �rstone. The processing of the NP results inthe CLOSED{state of the WCAM. At thisstage, the *VB{CO{act is triggered and theitems of the case frame on the syntactic andthe semantic level are allocated to the va-

5*NOM{OP/1 is triggered in the case of appo-sition (e.g., Peter der Gro�e, Mrs Jane Smith, theOxford English Dictionary, Archimedes' Law).

lencies opened by the verb (cf. Figure 4).Expecting an agent (AGT) and an object(OBJ), among others, the valencies of theverb are satis�ed by the complex semanticstructures of the �rst and the second NP,playing the AGT{ and OBJ{role, respec-tively. On the syntactic level, the agreementin respect of person, number and case forthe AGT{role and with reference to case forthe OBJ{role is checked. On the semanticlevel, for both roles the agreement in regardof semantic sorts is veri�ed.

4 GENERAL ASPECTS

4.1 DISAMBIGUATION PRIN-

CIPLES

Although a variety of approaches and a lotof systems have been developed, a generalnatural language system is still out of sight.This desideratum is generally explained withthe enormous complexity of natural lan-guage (see, e.g., [St�urmer 93]). One of themain reasons for this complexity lies in theambiguity of natural language expressions.Therefore, NLP must be based on adequatestrategies which support the disambiguationprocesses.6

In order to explain the human pref-erences in natural language understand-ing, researchers in the �elds of compu-tational linguistics (CL) and psycholin-guistics have suggested several principlessuch as Minimal Attachment, Right As-

sociation, Head Attachment and Lexical

Preferences (see, e.g., [Frazier, Rayner 82],[Ford et al. 82] or [Allen 87]). Some of them

6For the resolution of ambiguity see, for example,[Hirst 87], [Hindle, Rooth 93], [Schmitz, Quantz 93].

4 GENERAL ASPECTS 10

have been con�rmed by psycholinguistic ex-periments (see, e.g., [Hemforth et al. 92] or[Hemforth et al. 94]).

In the word agent based language process-ing of the WCAM the following general dis-ambiguation principles are implemented:

� Completion { semantic constituents,which have their correspondence in men-tal kernels built during human languageprocessing, are created and completed assoon as possible.

� Valency { in the broadest sense, the va-lencies of words a�ect the attachmentpreferences. The principles of Compat-

ibility and Incompatibility handle thesepreferences by making use of the subcate-gorization feature SELECT and the com-patibility feature COMPAT in the lexi-con (cf. Figure 4). Compatible featurevalues of two constituents permit the sub-ordination of the second one and, on theother hand, feature values which are notcompatible inhibit the attachment of con-stituents. Furthermore, the features areprocessed in an descending order (princi-ple of Priority). Obligatory valencies ofa word have priority over optional onesand the latter have priority over adver-bial quali�cations.

� Right Association { a new constituenttends to be interpreted as being part ofthe preceding constituent.

� Preference of Reading { this principlestates that, in the case of polysemouswords, one of the alternative meaningstends to be preferred. The preferredreading must be marked in the lexicon.

A detailed description of the disam-

biguation principles used in the WCAMand illustrative examples are given in[Helbig, Mertens 94]. The co{operation ofthe strategies for disambiguation and thelexical information, which plays an impor-tant part in resolving structural ambiguities,is presented in [Helbig et al. 94a].

Figure 4: Lexical entry of schreiben (towrite)

4.2 LEXICAL INFORMATION

A lexical entry in LINAS consists of mor-phological, syntactic and semantic informa-tion. The feature MORPH has as its value afeature structure which itself comprises suchfeatures as WCA, GEN and FLEX. WCA


denotes the corresponding word{class agentsuch as *VB (verb), *NOM (noun) and*PREP (preposition). The feature FLEXcontains the description of the morphosyn-tactic properties of a word, including typeof declension (for articles, adjectives andnouns) and conjugation (for verbs), respec-tively. The types of declension for words inthe nominal group and the types of conjuga-tion for verbs are divided into several groups.For instance, the group S02 represents thenouns with the ending {s in genitive and inall plural forms (e.g.,Auto { car). In the caseof words of the nominal group, the featureGEN designates the gender. An example ofa noun entry is shown in Figure 3.

The values of the feature SELECT areobligatory and optional valencies of words.For verbs it represents the case frame onthe syntactic and the semantic level. For in-stance, the German verb schreiben (to write)requires an agent (AGT) and at least oneof the two roles object (OBJ) or a dativerole (DAT). Besides, it allows for an optionalthematic role.7 The shortened lexical entryof this verb is given in Figure 4. In ad-dition, the feature structure COMPAT ex-presses the compatibility between verbs andadverbial quali�cations.

The feature structure SEMREL of a lex-ical entry contains the semantic relationsto other words in the lexicon. For in-stance, if two words have the same mean-ing, they are in the relation of synonymy(e.g., schreiben { to write and verfassen {to pen). Some other examples of semanticrelations included in the lexicon are SUB(hyponymy), PARS (part{whole{relation),

7Another possible role { BENF (bene�ciary) { isomitted here for the sake of brevity.

relation characterization

AGT agent

ASSOC association

CAUS causality

DAT dative role

DIRC direction

EQU equality

INSTR instrument

LOC location

OBJ object involved

ORIGM original material

PARS part-whole-relation

POSS possessive relation

PROP property, attribute

RSLT result

SUB subordination, hyponymy

SUBA subordination of events

TEMP temporal relation

THM thematic role

Table 2: Some examples of semantic rela-tions

ASSOC (association), etc. The informationabout semantic relations plays a crucial partin lexical disambiguation.

The feature SEMSORT has as its valuethe semantic sort of the lexical entry such asV { Vorgang (event), PS { Person (person)or L { Lokation (location). In LINAS theclassi�cation of entities is divided into thefollowing three levels:

� epistemic{ontological level { entities aredivided into sorts which are philosoph-ically motivated and which are used inde�ning the algebraic structure of the


knowledge representation apparatus.

� socio{cultural level { entities are speci�edby a given set of features allowing for across classi�cation and being motivatedby the importance of these categories forthe description of human activities.

� linguistic subtlety level { entities are char-acterized by connecting them to parts ofa semantic network. This information isused for linguistic �ne tuning of selectionrestrictions.

A detailed discussion of the underlyingthree{level classi�cation and illustrative ex-amples can be found in [Helbig et al. 94b].

4.3 KNOWLEDGE REPRESEN-

TATION

The syntactic and semantic analysis aims atrepresenting NL input by means of a multi{layered extended semantic network (MES-NET) and at integrating the results of theinterpretation process in a network knowl-edge base.

SUB

BUCH

AGT

SCHREIBEN

THM

SUBA

PROP

MANN

1990

COMPUTERLINGUISTIK

JUNG

OBJ

TEMP

SUB

Figure 5: Semantic representation of a sen-tence

In LINAS the language interpretation

process involves selecting the appropriate se-mantic relations which encode the connec-tions between the constituents of a given NLsentence, and using them to construct a se-mantic network that represents the completeinformation of the sentence. Some examplesof semantic relations and their characteriza-tions are listed in Table 2. The semantic rep-resentation of the following example is givenin Figure 5.

� 1990 schrieb der junge Mann ein Buch

�uber Computerlinguistik.

� In 1990 the young man wrote a book

about computational linguistics.

The knowledge representation methodused in LINAS extends the basic approachof semantic networks (SN) by accounting forthe information such as the partitioning ofthe SN into shells and layers, the distinctionbetween di�erent levels and dimensions ofthe underlying classi�cation, etc. The maincharacteristics of this new approach are thefollowing:

� strati�cation { the SN is organized in dif-ferent layers, the most prominent of thembeing the intensional and the preexten-sional layers.

� semantic sorts { the nodes of the SN arecharacterized by semantic sorts which areused in de�ning the network's algebraicstructure. In addition, the nodes can beannotated with the semantic features andthe selectional categories of the underly-ing three{level classi�cation (cf. Section4.2).

� semantic relations and functions { thenodes of the SN are linked to other nodesof the net by means of semantic rela-


tions and functions, which represent thesemantic connections between the con-stituents of the sentence.

� semantic shells { in order to representa complex semantic structure, a set ofnodes and semantic relations of an SNcan be combined in a semantic shell.

� semantic dimensions { the semantic di-mensions are formed by a bundle of con-trastive pairs: e.g., determinate vs. in-determinate reference of concepts, vir-tual vs. real concepts, individual vs.generic concepts, and collective vs. non{collective concepts. Furthermore, thenodes of the SN can be speci�ed by in-tensional and preextensional indicationsof quantities.

5 THE FOUR MAIN LEV-

ELS OF PROCESSING

The natural language processing of theWCAM is organized into four main levels(cf. Figure 6):

� on the �rst level, elementary semanticconstituents are created which have theircorrespondence in mental kernels builtduring human language understanding;

� on the second level, complex semanticconstituents are constructed by connect-ing the elementary constituents;

� on the third level, the constituents cre-ated on the �rst and second level are usedto construct elementary propositions of agiven sentence;

� in order to give a semantic representationof the whole sentence, on the fourth level

a complex proposition is built up by con-necting the elementary propositions.

The levels of the WCAM will be illus-trated by analysing the following variationof the example which has been introducedin Section 4.3:

� Der junge Mann schrieb ein Buch �uber

Computerlinguistik.

� The young man wrote a book about com-

putational linguistics.

Sentence

Semantic Representation

WCAM: Fourth Level

WCAM: First Level

WCAM: Second Level

WCAM: Third Level

*VB-OP

*ART-OP

*ART-CO

. . .

*PREP-CO

*CONJ-CO

*VB-CO

Central

Memory

(CWM)

Word-Class

ReferenceMemory(RM)

Word-Class

ComputerLexicon

(COLEX)

LEXMORPH

elementary constituents

elementary propositions

complex constituents

complex propositions

words

Agent Machine(WCAM)

Working

Agents

Figure 6: Overview of the WCAM levels

5.1 ELEMENTARY

CONSTITUENTS

On the �rst level, for each word of a sen-tence the morphological analysis is activatedin order to read the feature values of thecorresponding lexical entries. If the wholesentence has been analysed on the �rst level


<jung> <Mann> <Buch>

<schreiben> <über>

<Computerlinguistik>

CWM

SUBSUB

K1 K2 K3

SUBPROP

temporal restriction: PAST

[REF: DET] [REF: INDET] [REF: INDET]

<young> <man> <book> <computational linguistics>

<write> <about>

Figure 7: The CWM structure on the �rstlevel

(the activation of the word{class agents andthe changing of the WCAM states are de-scribed in Section 3), three elementary se-mantic constituents have been created whichhave their correspondence in mental kernels

built during human language understanding.In Figure 7, which shows the structure of theCWM at the end of the �rst level, these ker-nels are labelled K1 to K3.8

<junger Mann> <schreiben> <Buch> <über> <Computerlinguistik>CWM

CLIST

OLIST

K1 *VB K2 *PREP K3

<young man> <write> <book> <about> linguistics><computational

Figure 8: OList and CList on the secondlevel

5.2 COMPLEX

CONSTITUENTS

The task of the WCAM on the secondlevel consists in constructing complex con-

8Determinate vs. indeterminate reference of con-

cepts is marked by DET and INDET, respectively.

This contrastive pair forms one of the semantic di-

mensions mentioned in Section 4.3.

stituents by connecting the elemenary con-stituents determined on the �rst level. Adi�cult question concerns whether the el-ementary semantic constituents should besubordinated to a preceding elementary con-stituent or to the verb.

The elementary constituents K1 to K3 aremarked by CL (closed) because they are per-mitted to satisfy the valencies of other con-stituents of the sentence. The verb schreiben

and the preposition �uber are marked by OP(open) because these words open valencies(cf. Section 3). The constituents marked byCL are stored in the CList and the wordsmarked by OP in the OList.

<jung> <Mann> <Buch><schreiben> <Computerlinguistik>

CWM

SUBASUBPROP

temporal restriction: PAST

THMSUB

AGT OBJ

<young> <man> <book><write> linguistics>

<computational

Figure 9: The structure of the CWM aftercreating an elementary proposition on thethird level

Given that the elements of the CList andthe OList are arranged in the order of thecurrent sentence, a useful technique thattakes advantage of the structure of the twolists is to analyse the sentence in reverse or-der. In our example, the kernel K3 may besubordinated to the kernel K2 or to the verb(see Figure 8). The lexical entries of thenoun Buch and the verb schreiben shownin Figures 3 and 4 allow both subordina-tions. Following the disambiguation princi-ple of Right Association (see Section 4.1),


K3 will be attached to the preceding kernelK2. Thus, a complex kernel is built up byconnecting the kernels K2 and K3.

Semantic Shell

Semantic Shell

TEMPPAST1

PROP

<verärgert>

SUB

<Verleger>

CAUS

<schreiben>

SUBA

AGT

PROP SUB

PAST

TEMP

<jung> <Mann>

OBJ SUB THM

<Buch> linguistik><Computer-

1

<publisher>

<angry>

<young> <man>

<write> <book>linguistics>

<computational

Semantic Shell

Semantic Shell

Figure 10: The structure of the CWM afterconstructing a complex proposition on thefourth level

5.3 ELEMENTARY

PROPOSITIONS

On the third level, the kernels determinedon the levels before are subordinated to theverb in order to saturate its open valencies.The valencies are processed in a descendingorder. First the obligatory valencies of theverb and then the optional valencies are sat-

is�ed. Finally the adverbial quali�cationsare subordinated to the verb. The subordi-nation of the kernels is done by making useof the subcategorization feature SELECTfor the obligatory and optional valencies andthe compatibility feature COMPAT for theadverbial quali�cations (cf. Figure 4).

The case frame of the verb schreiben re-quires an AGT and at least one of thetwo roles OBJ or DAT. In our example,the AGT{role and the OBJ{role are ex-pressed by the elementary kernel <jungerMann> and the complex kernel <Buch �uber

Computerlinguistik>, respectively. At thisstate, the semantic representation of thegiven sentence is stored in the CWM (cf.Figure 9).

In addition, the case frame of schreiben al-lows for an optional thematic role (THM)which may be satis�ed by the kernel K3.This saturation is blocked, however, for thereasons mentioned above.

5.4 COMPLEX PROPOSITIONS

On the fourth level, the elementary proposi-tion which has been determined in the pre-ceding level may be connected to anotherelementary proposition. This process of theWCAM will be illustrated by analysing an-other variation of the example introducedbefore:

� Der Verleger war ver�argert, weil der

junge Mann ein Buch �uber Computerlin-

guistik geschrieben hatte.

� The publisher was angry because the

young man had written a book about com-

putational linguistics.

REFERENCES 16

Now a clause of reason being added to theexample, two elementary propositions canbe determined by the WCAM on the thirdlevel and, in each case, the nodes and re-lations of the propositions are combined inseparate semantic shells (cf. Figure 10).

On the fourth level, �rst the temporalinformation will be added to the SN andthen the main task of the WCAM consistsin analysing the subordinating conjunctionweil. In order to guarantee the interpreta-tion of the conjunction, the COMPLETE{act of the word{class agent *CONJ is acti-vated and, as a result of this process, thetwo propositions are combined by the corre-sponding semantic relation CAUS.

6 CONCLUSIONS

In this paper we summarized the fundamen-tals of our approach to word agent basedNLP. We further discussed the syntactic{semantic analysis of the WCAM and the dif-ferent levels of language processing. In addi-tion, we introduced the system componentswhich are directly connected to the WCAM.

The word based approach presented in thispaper can be characterized brie y by the fol-lowing essential properties:

� the word and its combinatory potentialplays the central part in natural languageunderstanding ;

� each word{class is connected to a corre-sponding word{class agent representingits grammatical function;

� integration of syntactic and semantic in-terpretation;

� the word{class agents are divided intotwo di�erent acts corresponding to theopening and satisfying of valencies;

� the word{class agents are triggered andinspected by the word{class agent ma-chine;

� the language processing is organized infour main levels: elementary and com-plex constituents, elementary and com-plex propositions;

� the disambiguation processes are sup-ported by a group of strategies and, inaddition, by special information in thelexicon;

� the result of the analysis is representedby means of a multi{layered extended se-mantic network.

Our approach is practically applied in anatural language understanding system andthe components described in the previoussections are implemented and successfullyused in the bibliographic information re-trieval system LINAS at the University ofHagen.

REFERENCES

[Allen 87] Allen, J.: Natural Language Un-derstanding. Menlo Park, California:Benjamin/Cummings 1987

[Boves, Nijholt 94] Boves, L., Nijholt, A.(eds.): Speech and Language Engeneer-ing, Proceedings of the eighth TwenteWorkshop on Language Technology, En-schede, 1 and 2 December 1994, Univer-siteit Twente, Faculteit Informatica 1994

REFERENCES 17

[Br�oker et al. 93] Br�oker, N., Hahn, U.,Schacht, S.: Ein Pl�adoyer f�ur Perfor-manzgrammatiken. In: [DGfS{CL 93],pages 6{11

[Br�oker et al. 94] Br�oker, N., Hahn, U.,Schacht, S.: Concurrent LexicalizedDependency Parsing: The ParseTalkModel. In: COLING 94 { Proceedingsof the 15th International Conference onComputational Linguistics, August 5{9, 1994, volume I, Kyoto, Japan 1994,pages 379{385,

[DGfS{CL 93] Deklarative und prozedu-rale Aspekte der Sprachverarbeitung,Tagungsband der 4. Fachtagung derSektion Computerlinguistik der Deut-schen Gesellschaft f�ur Sprachwissen-schaft, Universit�at Hamburg 1993

[Eimermacher 86] Eimermacher, M.: Wort-orientiertes Parsing mit erweiterterChart{Repr�asentation. In: Rollinger,C.{R., Horn, W. (eds.): GWAI{86 und2. �Osterreichische Arti�cial{Intelligence{Tagung, 22.{26. September 1986, Ot-tenstein/Nieder�osterreich. Berlin etc.:Springer{Verlag 1986, pages 131{142

[Eimermacher 88] Eimermacher, M.: Wort-orientiertes Parsen. Dissertation. Tech-nische Universit�at Berlin, FachbereichInformatik 1988

[Ford et al. 82] Ford, M., Bresnan, J., Ka-plan, R. M.: A competence{based the-ory of syntactic closure. In: Bresnan,J. W. (ed.): The Mental Representationof Grammatical Relations. Cambridge,MA: MIT Press 1982, pages 727{796

[Frazier, Rayner 82] Frazier, L., Rayner, K.:Making and correcting errors during sen-tence comprehension: eye movements inthe analysis of structurally ambiguoussentences. In: Cognitive Psychology, 14,pages 178{210

[G�orz 92] G�orz, G. (ed.): KONVENS 92{ 1. Konferenz Verarbeitung nat�urlicherSprache, N�urnberg, 7.{9. Oktober 1992.Berlin etc.: Springer{Verlag 1992

[Helbig 86] Helbig, H.: Syntactic{semanticanalysis of natural language by a newword{class controlled functional analysis(WCFA). Bratislava: Computers and AI5 (1986) 1, pages 53{59

[Helbig, Mertens 94] Helbig, H., Mertens,A.: Der Einsatz von Wortklassenagen-ten f�ur die automatische Sprachverar-beitung. Teil I { �Uberblick �uber dasGesamtsystem. Informatik Berichte 158.FernUniversit�at Hagen 1994

[Helbig 94] Helbig, H.: Der Einsatz vonWortklassenagenten f�ur die automa-tische Sprachverarbeitung. Teil II {Die vier Verarbeitungsstufen. InformatikBerichte 159. FernUniversit�at Hagen1994

[Helbig et al. 94a] Helbig, H., Mertens, A.,Schulz, M.: Die Rolle des Lexikonsbei der Disambiguierung. In: [Trost 94],pages 151{160

[Helbig et al. 94b] Helbig, H., Herold, C.,Schulz, M.: A Three Level Classi�cationof Entities for Knowledge RepresentationSystems. Informatik Berichte 162. Fern-Universit�at Hagen 1994

REFERENCES 18

[Hemforth et al. 92] Hemforth, B., Koniecz-ny L., Scheepers, C., Strube, G.:SOUL{Processing: Semantik{orientiertePrinzipien menschlicher Sprachverar-beitung. In: [G�orz 92], pages 198{208

[Hemforth et al. 94] Hemforth, B., Koniecz-ny L., Scheepers, C.: Principle{basedor Probabalistic Approaches to HumanParsing: How Universal is the HumanLanguage Processor? In: [Trost 94],pages 161{170

[Hindle, Rooth 93] Hindle, D., Rooth, M.:Structural ambiguity and lexical rela-tions. In: Computational Linguistics, 19,1993, pages 103{120

[Hirst 87] Hirst, G.: Semantic interpreta-tion and the resolution of ambiguity.Cambridge University Press: Londonetc. 1987

[Schmitz, Quantz 93] Schmitz, B., Quantz,J. J.: Defaults in machine translation.KIT{Report 106, Technische Universit�atBerlin 1993

[Sikkel, Nijholt 93] Sikkel, K., Nijholt, A.(eds.): Natural Language Parsing {Methods and Formalisms, Proceedings ofthe sixth Twente Workshop on LanguageTechnology, Enschede, 16 and 17 Decem-ber 1993, Universiteit Twente, FaculteitInformatica 1993

[Small 81] Small, S.: Viewing word expertparsing as linguistic theory. Proceedingsof the Seventh International Joint Con-ference on Arti�cial Intelligence (IJCAI{81), 24{28 August 1981, volume I, Uni-versity of British Columbia, Vancouver1981, pages 70{76

[Small 87] Small, S. L.: A DistributedWord{Based Approach to Parsing. In:Bolc, L. (ed.): Natural Language ParsingSystems. Berlin etc.: Springer{Verlag1987, pages 161{201

[St�urmer 93] St�urmer, T.: Semantic{Ori-ented Chart Parsing with Defaults. In:[Sikkel, Nijholt 93], pages 99{106

[Trost 94] Trost, H. (ed.): KONVENS '94 {Verarbeitung nat�urlicher Sprache, Wien,28.{30. September 1994. Berlin: Sprin-ger{Verlag 1994

FernUniversit�at Hagen

Praktische Informatik VII/Arti�cial Intelligence


The following reports of the Praktische Informatik VII/Arti�cial Intelligence can be obtainedby anonymous FTP from ftp.fernuni-hagen.de (132.176.114.150). After login with username anonymous change to the directory /fachb/inf/pri7. If you should have any problemsplease contact [email protected].

IB 158 Hermann Helbig, Andreas Mertens:Der Einsatz von Wortklassenagenten f�ur die automatische Sprachverarbeitung { Teil I:�Uberblick �uber das Gesamtsystem. Informatik Berichte 158 (4/1994), Fachbereich Infor-matik, FernUniversit�at Hagen, 1994

IB 159 Hermann Helbig:Der Einsatz von Wortklassenagenten f�ur die automatische Sprachverarbeitung { Teil II:Die vier Verarbeitungsstufen. Informatik Berichte 159 (4/1994), Fachbereich Informatik,FernUniversit�at Hagen, 1994

IB 162 Hermann Helbig, Claus Herold, Marion Schulz:A Three Level Classi�cation of Entities for Knowledge Representation Systems. InformatikBerichte 162 (8/1994), Fachbereich Informatik, FernUniversit�at Hagen, 1994

IB 168 Hermann Helbig, Andreas Mertens, Marion Schulz:Disambiguierung mit Wortklassenagenten. Informatik Berichte 168 (12/1994), FachbereichInformatik, FernUniversit�at Hagen, 1994

IB 169 Hermann Helbig, Andreas Mertens:The Word Class Agent Machine. Informatik Berichte 169 (1/1995), Fachbereich Informatik,FernUniversit�at Hagen, 1995

hermann helbig and andreas mertens - fernuni-hagen.de

Documents