THE AUTOMATICALLY BUILT UP HOMOGRAPH DICTIONARY A COMPONENT OF A DYNAMIC LEXICAL SYSTEM

Heinz J. Weber

THE AUTOMATICALLY BUILT UP HOMOGRAPH DICTIONARY - A COMPONENT OF A DYNAMIC LEXICAL SYSTEM -

0. Introduction.

Ambiguous word forms (often called " homonyms " or - in written language - " homographs ") are known as obstacles in many fields of computational linguistics, especially in automatic documentation, content analysis or mechanical translation. In this respect two problems must be distinguished:

1) the detection of homographie word forms in the text,

2) their disambiguation by analysis procedures. This paper exclusively deals with the first problem.

1. The Detection Of Homographs.

1.1. In current procedures for the detection of homographs two alternatives can be differentiated:

i) Homographs are identified like monosemic word forms by segmentation and looking up in the standard lexicon. Homographs are detected, if segments of text word forms correspond with more than one lexicon-entry. Lexicon-entries representing homographie items therefore need no special marking.

ii) Homographs are identified by means of a special homograph dictionary, which can be worked out in two versions :

1) the homograph dictionary contains the graphemic shapes of all homographie word forms (full forms) and their possible linguistic specifications. In this case no segmentation procedures are required.

2) the dictionary does not contain full forms but only the respective canonical forms. A special marking gives information about other corresponding dictionary-entries and the extent of their overlapping.

In both cases (1) and (2) the identification of homographie text word forms is separated from the identification of monosemic ones.

Procedures (i) and (ii) have some characteristic advantages and defects, which I will consider rather briefly.

1.2. As already pointed out the first method requires (1) a segmen-tational algorithm, with the help of which the word forms of a text can be parsed into segments (e.g. stems and inflectional affixes), (2) an identificational component composed of a grapheme-sequence-comparing algorithm and the standard lexicon ; thus it can be checked, whether a text segment detected by (1) is the expression-side of one (or perhaps more) lexical unit(s). If this is the case, the content-side (s) of the corresponding unit(s) can be assigned to the text segment.

1.2.1. According to this conception the identification of word forms would offer no problems in the following cases:

i) the word form represents only one sequence of segments (is monosemic), that means that each segment corresponds to one lexicon-entry :

Germ. : .. .0 kindes0...

Engl.: ...0childs0...

Frnc. : ...0enfant0...

Pvuss. : ...0rebënok0.

-/kind/+/es/Kind-/child/+/s/child-/enfant/-|-/0/enfant-/rebënok/+/0/rebënok

ii) The word form can be parsed into more than one set of segments (is homographie), and the possible readings show coinciding segment-boundaries :


Germ. :	...0laut0...	-/laut/+/0/ /laut/+/0/	*Laut (SUB) laut (ADJ)*
Engl. :	...0mean0...	-/mean/+/0/ /mean/+0/	*mean (ADJ) to mean*
Frnc:	...0mort0...	-/mort/+/0/ /mort/+/0/	*mort (SUB) mort (ADJ)*
Russ. :	...0zila0...	/zil/+/a/	*zilarzila%*

If the detected segments are compared with all graphematically corresponding lexicon-entries, that is, if the lexicon-look-up is not stopped after the first correspondence, this case can also easily be coped with. If the entries are arranged in alphabetical order, the respective units will immediately succeed one another.

1.2.2. The detection becomes difficult however, if one word form is parsed into more than one set of segments (is homographie), while

As the segments which have thus been detected do not coincide graphematically (e.g. /trieb/ - /getriebe/), i.e., as the respective lexicon-entries are to be found at different places in the lexicon, for the identification of such homographs enormous parsing - and comparing -procedures are required. As cases of homography with overlapping segment-boundaries in the various readings are encountered quite frequently in languages with extensive inflection (e.g. German, French, Russian), method 1.1. (i) is not the best in any case.

1.2.3. The advantage of this method is above all to be seen in the fact that the identification of homographs can be managed automatically, and that no special marking of the respective entries is necessary. This is especially important with regard to dynamic lexical systems, where the number of lexicon-entries and their specification can vary ; new entries do not require a change of the detection procedure. The disadvantage consists in the fact that monosemic and ambiguous word forms are submitted to the same procedure, which amounts to an undue delay of the determination of monosemic word forms. Multiple parsing with subsequent lexicon-look-up has always to be applied if the respective text word forms contain grapheme sequences, which correspond to inflectional affixes. Only after this can it be found out whether more than one plausible reading has resulted:


the segment-boundaries in	the possible readings are	overlapping :
Germ. : .. .0getrieben0...	-/trieb/+/ge-en/	*treiben*
	/getriebe/+/n/	*Getriebe*
Engl. : .. .0hearing0...	-/hear/+/ing/	*to hear*
	-/hearing/470/	*hearing*
Frnc. : .. .0pecher0...	-/pêch/4-/er	*pêcher (VRB)*
	/pêcher/4-/0/	*pêcher {SUB)*
Russ.: ...0valax0...	-/val/+/ax/	*val*
	/valax/+/0/	*valax*

Germ. : .. .0getrieben0...

~/trieb/+/ge-en/ /getriebe/+/n/

or only one reading is true: Germ. : .. .0gelagen0...

-/gelage/+/n/

but not:

/lag/+/ge-en/

1.3. Method 1.1. (ii) does not have this disadvantage. Homographs are separately registered and marked according to their readings; thus a considerable acceleration of the identificational procedure is made possible. Monosemic word forms, the stems of which are listed up in the standard lexicon, are identified more easily, as the segmentation and lexicon-look-up can already be stopped after assignment of one reading. Ambiguous word forms are specified more easily, as the extensive segmentation- and comparing-procedures do not have to be applied (as the various readings are registered in the homograph dictionary - version 1.1. (ii) (1) -) or are reduced to a minimum (as the respective lexicon-entries bear a special marking, by which their homography can be derived - version (ii) (2) -). These advantages however entail certain disadvantages: as a rule homograph dictionaries are built up manually and have to be manually complemented, when the standard lexicon is extended; the same has to be stated for the marking of lexicon-entries. Aside from this troublesome and time-consuming business one cannot be sure that all homographies are registered or are marked exhaustively.

2. The Automatically Built Up Homograph Dictionary.

2.1. In this paper a method will be outlined, in which the advantages of the first procedure are combined with those of the second one: the standard lexicon therefore can be extended automatically without delaying the identificational procedure. The homograph dictionary is compiled by analysis of the standard lexicon; all stems representing homographie items are taken away from it and integrated into the homograph dictionary. The same algorithm, which detects homographies incorporated in the standard lexicon, can be used to find out by analysis of both lexica, whether new entries and all inflected forms represented by them are homographs. If this is the case, they are registered in the homograph dictionary, otherwise in the standard lexicon. Thus the number of entries in both lexica can be increased automatically and the specifications of ambiguities in the homograph dictionary always correspond to the current state of information.

Standard Lexicon

New Entries-Algorithm

Homograph Dictionary

2.2. The standard lexicon can be characterized as follows: the lexicon-entries are "stems"; each stem representing a set of inflected forms, which is called a " paradigm " or " part of a paradigm ". In order to abbreviate the graphemic assimilation of text word forms to the graphemic shapes of lexicon-entries during the identificational procedure, morphologically and syntactically determinated allomorphs have also been noted. Stems of complex lexical units (e.g. /kick0the0 bucket/, /zum0zug0komm/) are, however, ignored:

stems of (Ger.)graben (VRB):/grab/, /grub/

beipflichten: /beipflicht/, /pflicht/ ...

stems of (Eng.)to sing:/smg/> /sang/> /sung/

stems of (Frc.) mourir: /mour/, /meur/, /mort/, ...

stems of (Rus.) rebënok: /rebënok/, /rebënk/, ...

2.3. The homograph dictionary is built up by comparing selected entries of the standard lexicon. In order to elucidate the comparing procedure we restrict ourselves to the coordination of just two lexicon-entries. Two stems represent homographie inflected forms, if the following conditions are fulfilled:

i) The graphemic shapes of the stems belonging to the paradigms Pi and p2 are identical :

gs(2\) ^gs(p2).

In this case homography exists, if any inflectional affixes co-occurring with the respective stems are homographie too:

g;^) ^ g;(p2).

ii) The graphemic shape belonging to the stem of paradigm Plsconcatenated with a sequence Gk, is homographie with the stem of paradigm P2:

G.(P0 X G»£ Gf(Ps).

In this case the graphemic shapes of the co-occurring inflectional affixes have to correspond in the following way:

G}(P1)\Gh^G}(P,).

Concerning Gk some restrictions have to be observed:

iii) Gk has to be homographie or partially homographie with any inflectional affix of the respective language :

Gk £ G}, for G} as an element of the finite set <JÇ which contains all inflectional affixes.

iv) G} for its part must be co-occurring with the respective stem of the paradigm :

Gi-G^P,).

The co-occurence of stem and affix is specified by the respective inflection-class-marking of the lexicon-entry.

2.4. Conditions (i) and (ii) determine the selection of stems, (iii) prevents that all partially homographie stems in the lexicon are selected and examined. Thus - for example - a coordination of the German stems /arm/ Arm and /armee/ Armee is not permitted (though they are fulfilling condition (ii): Gs(Pi) X /...ee/ = GS(P2)), as the sequence Gh = /...ee/ does not fulfill the condition Gh s. G).

iv) further reduces the number of the selected stems to those cases, where G'f (2 Gh) is co-occurring with Gs(Px) (e.g. German /schwer/ schwer and /schwert/ Schwert are not combined, though both stems fulfill conditions (ii) and Gk (= /...t/) fulfills condition (iii); but Gh does not fulfill (iv) : j...tj is not homographie with an affix co-occurring with /schwer/ (ADJ)).

The relationship between the graphemic shapes of the affixes co-occurring with the stems of both paradigms is finally specified by the complementary part of condition (ii): G}(Pi) \ Gk = G}(P2). E.g. the German stems /hoer/ hören and /gehoer/ Gehör (SUB) could be combined according to condition (ii): /hoer/ X /ge.../ = /gehoer/. Moreover conditions (iii) and (iv) concerning Gk = /ge.../ would be fulfilled: /ge.../ <= Gj and Gj(= /ge-t/) = G}(PX). But there is no inflectional affix /t/ co-occurring with the stem /gehoer/ corresponds to /ge-t/ in the mentioned way:

2.5. These conditions determining the entries and their examination for homographie into programmed instructions:

startj-rf

paradigm Pv paradigm Pi+1 of paradigm Pa, which /ge-t/\/ge.../^/t/.

^>_—j^take stem of«<-

selection of lexicon-items can be transferred -condition (ii):

stem ofcondition (i):y Gt(Pt)âG,(P{+1)—gs(pl+l)gs(p|)Xg/fcondition (iii):'g* <= g>I

condition (iv):

g; = g;(p,.)

compare infl. aff. ■of both paradigms according to (ii): G^KG^G}^) generate homographie infl. forms of both paradigms p,-, p!+1;

mark stems accord-• ing to the overlap-*-

- stem ofp,+]

I".

compare infl. aff. of both paradigms according to (i) G}(Pt) ^ Gj(Pf+T) ping of both paradigms 2.6. Prerequisites for the outlined algorithm are:

1) a computerized stem-lexicon; the entries, which have to be arranged in alphabetical order, must bear an inflection-class-marking, which makes it possible to generate all inflected forms represented by the respective stems:

/ficht/ -.fechten, VRB, infl.-class 48 /flehte/ : Fichte, SUB, infUclass 5

2) A complete list of inflectional affixes, which bear the possible inflection-class-markings corresponding to those of the stems :

/0/ : SUB, inn-classes 5, ... VRB, infl.-classes 48, ...

/st/ : VRB, infl.-classes 48, ...

2.7. The selection of stems and the comparison of the co-occurring inflectional affixes could be carried out in a slightly modified way. As already pointed out, the selection of stems is in the main determined by the grapheme sequence Gk (which specifies the graphematic overlapping of non-homographic stems). Further restrictions concern the correspondence between Gk and the inflectional affixes co-occurring with the selected stems (see 2.3. (iii) and (iv). As the inflection-class-markings of stems and affixes (which are similar) are shortened distributional classifications, it is obvious to bring them into a system, according to the respective specifications of Gk. A matrix is built up by which it can be seen whether a Gh -specification restricts the coordination of stems with certain inflection-classmarkings. In this way the detailed examination and comparison of all co-occurring affixes (in accordance to condition 2.3. (iv)) can be substituted by one single operation, at least in a good number of cases.

/wiese/ Wiese; ... VRB 1: /stell/ stellen; ... VRB 48: /ficht/ fechten; ... VRB 55: /wies/ weisen; ...

This matrix forbids the coordination of two stems belonging to the classes VRB 48 and SUB 5 (in this succession), though they may correspond in accordance to condition (ii) :

/ficht/ X /...e/ ^ /fichte/.

On the other hand the coordination of two stems belonging to the classes VRB 1 or VRB 55 and SUB 5 is admissibile:

/stell/ X /...e/ = /stelle/ or /wies/ X /...e/ ^ /wiese/.

The building-up of such matrices seems to be a useful device, as the number of GA-specifications in the respective languages (German, French, Russian) is limited. In German we have found out ten frequent and about thirty extremely rare GA-specifications. In English homographies with graphemetically overlapping stems are without that rather seldom.

2.8. The conceived algorithm selects just two entries (respectively their paradigms), which are examined for homographie word forms. After the first cycle of selecting and comparing - as pointed out in 2.5. -homographs, which are members of more than two paradigms (e.g.


	coordinated	G^-specifications of stem.! :
	infl.-classes of
	stemi/stem.2	Gk = /...e/ ...
	*VRB* 1 I *SUB* 5	+
	*VRB* 48/*SUB* 5	—
	*VRB 55fSUB* 5	+

German 0alben0 : Album, Alba, Albe, Albu Alb2), therefore, are not specified exhaustively :

Album D Alba = {alben} Album D Albe = {alben} Album n Albi = {alben} Album n Alb2 — {alben} Alba fl Albe = {alben} Alba n Albx — {alben} Alba fl Alb2 = {alben} Albe D Alb! = {albe, alben} fl Alba = {alben}

nAlb2={alb, alben}

In this example 0alben0 is described as a member of altogether ten intersection sets (which are the results of the first coordination-cycle). In a second cycle (which will not be dealt with in detail) all intersection sets of the first cycle are examined for identical word forms. 0alben0 now can be described as an intersection set of five paradigms :

The coordination of intersection sets in the second cycle is - as well as the coordination of paradigms in the first one - determined by conditions, which are derived from the graphemic shapes of the respective stems. In all probability stems like /album/, /alba/, /albe/, /alb/ will represent at least one homographie word form, while stems like /album/, /alge/, /alibi/, /altar/ will not.

2.9. As the outlined method of building up a homograph dictionary is, in the main, using facts of the expression-side of lexical units, it can be applied to various lexicon-types. The content-sides of the respective entries (i.e. stems) can bear either semantically, syntactically or otherwise relevant information.

H. Eggers, et al., Elektronische Syntaxanalyse der deutschen Gegenwartssprache, Tübingen 1969.

D. Kkaixmann, Linguistische Datenhank und kumulatives Wörterbuch, in Kolloquium Maschinelle Sprachverarbeitung, Mannheim 1968.

W. Lenders, Static And Dymanic Lexical Systems, Bonn 1969.

H. J. Weber, Mehrdeutige Wortformen -grammatische Beschreibung und lexikographische Erfassung (thesis), Saarbrücken 1973.