According to the description of Task 1 (Selection of lexical entries) given in the paragraph 3.1.3.7 (LEXICON WORKPACKAGE STRUCTURE) of the Technical Annex, Version 1.9, two options are specified for extracting our lexical entries to be included in the LE-PAROLE LEXICON : either "from existing resources or, on the base of their frequency, from corpora". We rather tried to combine these two strategies in the following way :

• Existing resources

We have an existing resource, the ILSP Morphological Lexicon of about 57,000 lemmata, which could offer us a great amount of lexical entries. At the beginning of the conversion procedure, thus, we have tried to exploit in the most efficient way the information already included in this Lexicon :

• We have opted for the conversion to the LE-PAROLE MODEL of the grammatical/function words, already encoded in the ILSP Morphological Lexicon. The idea for this decision was that the grammatical words per se are actually restricted and predictable, and, even if we had followed some other procedure for their selection, the result would be the same.
• The main criterion for the selection of entries of the other morphosyntactic categories, was that of Inflection : we have decided to convert for the LE-Parole requirements the most "populated" inflectional paradigms, that is to say, those inflectional paradigms that cover the inflection for the most of the Greek words. In other words, our conversion strategy focused on whole inflectional paradigms rather than lexical entries. This decision was forced by the time requirements as well as by the theoretical hypothesis that the entries already described by the most frequent and "populated" inflectional paradigms are the appropriate ones to be firstly included in a lexicon consisting of 20,000 lemmata.

After we had selected the entries following the procedure described above, we have decided to check their appropriateness in accordance to their frequency. Our main aim was to check if our selected entries were appropriate enough to be included in a medium size lexicon as well as if there were still frequent words that were not included in the selected entries. This kind of checking would force us to include in the existing entries those that appear to be the most frequent ones in the available corpus, whether they have an irregular way of inflection or not. In this way, we could balance the population of our lexicon in accordance to the criterion of the frequency of the entries and the criterion of the population of the inflectional paradigms.

In order to check upon the criterion of frequency, we have used an existing sublist of 9,567 lemmata with a minimum frequency range of 10 occurrences in a subset of about 2 million words of the ILSP Reference Corpus.

The 2 million words corpus used was of the following composition :

• Combination of strategies

We have automatically compared the selected lexical entries taken from the ILSP Morphological Lexicon to the list consisting of the frequent lemmata. The output of this comparison was a separate list including the most frequent lemmata appearing in the frequency list but not in the LE-Parole list. The new list consisted of about 6,000 lemmata. This result has satisfied our first hypothesis that the most populated inflectional paradigms describe, after all, the inflection of the most frequent Greek lemmas.

Going through the new list, we have sorted the lemmata according to their morphosyntactic category and, further, according to their inflectional behaviour to the relevant inflectional classes. The lemmata that did not appear in the LE-Parole list are divided according to their morphosyntactic category as follows :

Nouns : 3,000

Verbs : 1,350

Adjectives : 1,210

Other categories : 300

• Exploitation of the above results

We have, therefore, decided to incorporate in the LE-Parole list only those lemmata that can actually be considered frequent - given that our corpus composition is not well balanced and representative.

Summarising, corpora can be deemed as being neither the only nor the basic resource for the Greek LE-Parole Lexicon.

• Validation and complementation of the above results

The validation phase, which followed the construction phase of the morphological layer, had as its aim the quality control of the lemma list / macrostructure of the lexicon, as well as the quality control of the inflectional system described in our lexicon.

The method used for the quality control of the lemmata included two steps :

1. the comparison of the lemma list to a frequency list of about 21,000 words produced from a corpus of approximately 9 million words, and
2. the verification and complementation of the comparison results according to specific lexicographic principles : for instance, nouns which share the same morphological form for the masculine and feminine gender (e.g. "ãéáôñüò” [doctor]) should both appear in the lemma list; despite the fact that one was missing from the frequency list.

Therefore, the final list of entries which are currently included in the LE-Parole lexicon is the output of the above quality checking procedure, that is to say, it includes most of the frequent words found in the 9 million words corpus which have been further checked and completed following certain principles.

During the above procedure the correctness and robustness of the inflectional system were checked against the corpus as well. Thus, the inflectional descriptions assigned to some entries were revised in the light of real data. Such changes mainly affected, for instance, lemmata which had been described by an inflectional paradigm theoretically correct but not fully appropriate to the entries’ distribution in the corpus : for instance, the adjective "äïıñåéïò" - assigned in the previous version of the lexicon a full inflectional paradigm - appears in the corpus (and in general in the Greek language), only in the masculine gender and namely in the fixed phrase "äïıñåéïò ßğğïò” [Trojan horse]; in the new version, it has been assigned an elliptic inflectional paradigm.

On the other hand, certain inflectional paradigms have been enriched with forms which belong to the older Greek system ("katharevoussa"), but which are still used in certain text types - whenever this use has proved still systematic through their frequent occurrence in the corpus.

Our aim in this effort was to use corpus information in order to describe modern Greek morphology "in use" rather than describing a theoretically complete Greek morphological system which, however, does not reflect reality.

1. Conversion Process

As it has been already stated, the Greek LE-Parole Lexicon consists mostly of lexical entries converted from the ILSP Morphological lexicon.

The main difference between the morphological model used for the codification of the ILSP Morphological Lexicon and the LE-Parole model lies in the way they handle stress. The ILSP morphological model handles stress through dynamic processes, while the LE-Parole model forces for a flat treatment of stress. To clarify this difference, let's examine certain examples.

The neuter nouns ending in -o and carrying a stress either on the ultimate syllable or on the penultimate syllable are classified under the same inflectional paradigm in the ILSP M.L.

e.g. the inflectional paradigms for âïõíü [mountain] and âéâëßï [book]

Singular Number	Singular Number
âïõ í-ü	âé âëß-ï
âïõ í-ïı	âé âëß-ïõ
âïõ í-ü	âé âëß-ï
âïõ í-ü	âé âëß-ï
Plural Number	Plural Number
âïõ í-Ü	âé âëß-á
âïõ í-şí	âé âëß-ùí
âïõ í-Ü	âé âëß-á
âïõ í-Ü	âé âëß-á

As we can see, the behaviour of the stress is the same in the inflectional types of both columns : the stress remains at the same syllable, irrespective of whether this concerns the ultimate or the penultimate syllable.

While in the ILSP Morphological Lexicon we could exploit this kind of the stress behaviour, classifying both kinds of words under the same inflectional paradigm, for the conversion to the LE-Parole Model we had to distinguish between these two kinds of words, constructing two inflectional paradigms (in the LE-Parole phrasing, "systems of written inflection"): one for the "âïõíü" [mountain] type words that have stressed endings (and one unstressed stem) and one for the "âéâëßï" [book] type words that have one stressed stem (and unstressed endings) (cf. general information on the notion of the system of written inflection given in paragraph 5 [System of Written Inflection] of the User's Manual, included in the "Report on the Morphological Layer", Version 3.3, November 2, 1994).

The conversion process of the ILSP Morphological lexicon data to the LE-Parole model focused on the conversion of the ILSP Morphological Lexicon inflectional paradigms to the LE-Parole model ones (see also PP-PAROLE report "The Greek Lexicon - Inflectional Morphology").

In general lines, the correspondence between the existing inflectional paradigms and the ones of the LE-Parole model is not a one-to-one, but usually of the following schema:

1 inflectional paradigm in the ILSP M.L.à at least 2 inflectional paradigms in the LE-Parole Lexicon

Conversion of the lexical entries

The difference between the two morphological models has forced us to check one by one the lemmata per inflectional paradigm to be included in the LE-Parole Lexicon, in order to assign them the correct reference to the LE-Parole inflectional paradigms as well as additional information concerning, mainly, the different stems ("spellings" in the Parole phrasing) to be included in the new entries of the SGML form. We have further developed a specialised software for the conversion of the ILSP data to the LE-Parole SGML form.

Construction of the inflectional paradigms

The LE-Parole inflectional paradigms have been constructed manually, that is to say without using the tool. This task has proved quite difficult as regards the construction of the inflectional paradigms for verbs, which have an average of 9 different stems to combine with a variety of different endings.

 

2. Construction and conversion of the syntactic layer

The syntactic layer is built from scratch since the only existing computerised lexicon is the one developed during the EUROTRA project; it was considered more time-consuming to convert this data into the new model - taking into account the limited number of entries that exists for the syntactic layer in Greek - than to build something from scratch. The theoretical background, however, of this project has been valuable in the decision-taking process.

The development of the syntactic layer is made outside the tool. Thus, special software routines have been created in order to create the syntactic units and link them to the morphological ones, while the descriptions are created in an SGML file; for checking purposes, the English version of the Parole-e.dtd is being used.

1. Current Lexicon Contents
1. Morphological layer

Number of simple morphological units	20149
Number of agglutinated morphological units	1
Number of graphical morphological units	20397
Number of simple inflection modes	818

 

Category	Subcategory	Number of Units
Noun	common	11803
Noun	proper	239
Verb	impersonal	88
Verb	main	2926
Adjective		3405
Numeral	ordinal	18
Numeral	cardinal	59
Numeral	multiplicative	6
Numeral	analog	5
Numeral	collective	18
Pronoun	personal	1
Pronoun	possessive	1
Pronoun	demonstrative	8
Pronoun	indefinite	20
Pronoun	interrogative	4
Pronoun	relative	2
Pronoun	relindef	7
Pronoun	definite	2
Article	definite	1
Article	indefinite	1
Adverb		1396
Adposition	preposition	48
Conjunction	coordinative	17
Conjunction	subordinative	34
Interjection		21
Unique	negation	2
Unique	future (sscfuture)	1
Unique	subjunction	1
Unique	other	15

 

1. Splitting criteria for identifying Morphological Units

1. The grammatical category : the same lemma which corresponds to two distinct syntactic categories gives rise to two distinct lemmata : e.g.

1. the lemma "ùò" [till] (which falls under the grammatical category "ADPOSITION")
2. the lemma "ùò" [like] (which falls under the grammatical category "ADVERB")

are two distinct lemmata, since they have a different syntactic behaviour : the adposition "ùò" can never appear with a NP complement in the nominative case, while the Adverb "ùò" can.

2. The grammatical subcategory (referred to as Type in the LE-Parole model) : two lemmata of the same grammatical category are split into two Morphological Units. This is required by the model, that is to say it is a technical requirement, as well as a lexicographic requirement forced by the fact that usually morphological units with a different type (grammatical subcategory) have a different morphological and syntactic behaviour. This is the case for the Greek verbs which are split into different Morphological Units when taking different values for the Type (MAIN/IMPERSONAL). We have adopted this splitting criterion at the Morphological layer, because we have observed that the distinction between main and impersonal verbs is triggered not only by syntactic reasons but by morphological reasons too.

For instance, the verb "ìğïñş" [can] with the value "MAIN" is a personal verb which has a regular inflection and a certain syntactic behaviour (its Subject is realised as a NP), while the verb "ìğïñåß" [can] with the value "IMPERSONAL" is an impersonal verb which has a distinct morphological behaviour (e.g. it can form only the third singular person of the imperfective aspect) and a distinct syntactic behaviour (e.g. its subject is realised only as a Clause).

3. Meaning plus was also considered to be a criterion for splitting Morphological Units. When saying "meaning plus", we mean that meaning was not considered the only criterion for splitting, since it is usually accompanied by phonetic or/and morphological differentiation as well.

For instance, the verb "âéÜæù" ["violate"] and the verb "âéÜæù" ["hurry"] are two distinct units, not only because they have a different meaning, but also because they have a distinct phonetic and morphological behaviour as well. In the verb "âéÜæù" [violate] the two vowels "éá" are spelt as two distinct syllables, while in the verb "âéÜæù" [hurry] the same vowels are spelt as one syllable. Since the model cannot offer us the possibility to codify this phonetic distinction, we thought it better to split the two words in two distinct MuS. Additionally, the phonetic differentiation between these two words leads to distinct morphological realisations as well. The verb [violate] forms the past tenses "âßáæá, âßáóá " while the verb [hurry] forms the past tenses "İâéáæá, İâéáóá", although they coincide in some other morphological instantiations.

4. Gender was used as a splitting criterion as regards nouns. We have coded in two distinct Morphological Units the nouns that can appear both with a masculine and a feminine gender (in context, they are distinguished by the genders of the accompanying determiners):

1. Splitting criteria for identifying lemmata

We have coded as lemma the standard spelling/type of the word.

For instance, between the two spelling variants "áâãü" and "áõãü" [egg], we have opted for the variant "áõãü" to be the lemma of the Morphological Unit, following the standard Greek Grammar.

When there were no standards to distinguish between those cases, we have opted for the most common and frequent type of the word, using the data of contemporary Greek dictionaries as well as the data found in the corpus.

1. Syntactic layer

Number of Syntactic descriptions	483
Number of framesets	52
Number of Syntactic Units	25092

Category	Number of Units
Noun	14548
Verb	5397
Adjective	3558
Adverb	1410
Adposition	73
Numeral	106

All work on the syntactic layer is currently being curried out outside the AlethGD tool.

1. Splitting criteria for syntactic units

We have decided to use mainly syntactic criteria for splitting syntactic units - i.e. a different syntactic behaviour gives generation to a syntactic unit. Semantic criteria are used only when this has implications on syntax.

Therefore, the following criteria are used for syntactic unit splitting :

• obviously, when different meanings are associated with different syntactic behaviours (e.g. áñèñşíù ôï ëüãï [articulate a word] vs. áñèñşíù êÜôé óå ïìÜäåò [articulate something into groups]).
• when the presence of a complement of a verb affects its semantics : for instance, the verb "ìéëÜù" [speak] without an object means [to acquire the faculty of speech] and is coded as two syntactic units rather than one with an optional object.
• when optionality is constrained by conditions. For example, if a verb subcategorises for a direct and an indirect object and both complements are optional, but the presence of the second one presupposes the presence of the first one, we create two different descriptions: the first one has no complements while the second one subcategorises for an obligatory direct object and an optional indirect object. e.g. ğñïäßäù (ğñïäßäù / ğñïäßäù + object / ğñïäßäù + object + indobject / *ğñïäßäù + indobject) [to betray (I betray / I betray + object / I betray + object + indobject / *I betray + indobject)]
• when alternative realisations of the same position (e.g. a Noun Phrase or a Clause functioning as object of the same verb) are constrained through specific features. For instance, the verb "ğåßèù" [to persuade] subcategorises for a noun phrase (functioning as indirect object) and a clausal complement. The clausal complement can be realised either as a that-clause or a na-clause. Control information needs to be coded only with na-clause complement: the indirect object controls the subject of the na-clause; this is not true when a that-clause complements the verb. Two syntactic units are, therefore, encoded.

On the other hand, we decided not to split syntactic units when a verb has more than one (quite similar) meanings, but all meanings have the same syntactic behaviour. The distinction between different meanings will be made at the semantic level.

1. Choices in relation with optionality

As discussed in the previous paragraph, optionality is not a splitting criterion unless it is linked to meaning change or constrained by conditions of other complements' presence.

We decided to encode the subject as being always obligatory, even though the pronoun-subject in Greek is optional, given the inflectional nature of the language. Our conclusion is that pro-drop and pronominalisation are grammatical phenomena and, therefore, should be treated in the grammar.

In order to decide for the optionality of a complement, we take into account unmarked "complete" sentences out of any context : e.g. Ï ÃéÜííçò ãñÜöåé [John writes] vs. *Ï ÃéÜííçò âñŞêå [*John found].

2. Choices in relation with alternatives of realisations

We have decided that different realisations of the same position (e.g. a Noun Phrase or a Clause functioning as object of the same verb), i.e. that have the same function and no further constraints, are coded in the same frame and position. On the other hand, alternatives of realisations constrained through specific features are split into two different syntactic units.

Thus, for example,

• the verb "ëİù" [to say] will be given one syntactic unit, where the same description is used for the realisation of the object as a noun phrase in the accusative case or a clause and the indirect object as a prepositional phrase or a noun phrase in the genitive case, and
• the verb "ğåßèù" [to persuade] will be given two syntactic units : it can take a noun phrase (indirect object) and a clausal complement (object), but with one of these control information must be stated : if it is followed by a na-clause, its indirect object controls the subject of the na-clause; this is not true when a that-clause complements the verb.

1. Frameset construction strategy

A frameset is used when a verb can be associated to more than one constructions and the positions of these constructions are semantically related between them. It should be noted that passivisation is considered a grammatical phenomenon and not encoded as a frameset.

To encode cases related to each other, we have used the mechanism of frameset as proposed by the model (i.e. using the "Related" facility to link together the subcategorised elements). Among others, the following cases are treated by using a frameset:

• verbs that occur in a causative and an ergative construction:

• verbs that have a reciprocal meaning, when they are in the plural. A frameset will link all relevant descriptions: (a) for verbs with three complements (e.g. äİíù [tie]): object and oblique in one frame is related to one object (NP plural / coordinated NP) in the other frame, or (b) for verbs with two complements (e.g. óõíáíôéİìáé [meet]) : subject and oblique in one frame is related to one subject (NP plural) in the other frame. Information on the number of the verb is given in the SELF.

1. List of syntactic functions used

The functions used in Greek are HEAD, SUBJECT, OBJECT, INDIRECTOBJECT, SUBJPRED, OBJPRED, OBLIQUE, ADVERBIAL, CLAUSCOMP, ACOMP, NCOMP, ADVCOMP, PREPDEPENDENT. A set of morphosyntactic and semantic criteria, specific to the Greek language, is used to identify the proper function for each subcategorised for element.

A first discussion of these criteria was given in the PP-PAROLE reports on syntax, namely "Greek instantiation of the PAROLE specifications for verb subcategorisation" (Athens 1996) and "Greek instantiation of the PAROLE guidelines for non-verb subcategorisation" (Athens 1996). However, amendments have been, in order to take account of the decisions taken following the syntax workshop held in Barcelona and the discussions made throughout the project.

We present here only significant differences from these PP-PAROLE reports :

• the classification of a set of complements, namely nominal phrases in the accusative case subcategorised for by measure, weight, etc. denoting verbs, under the function ADVERBIAL instead of the previously suggested OBLIQUE - given that criteria for the assignment of these two have been further refined
• the use of the functions ACOMP, NCOMP, ADVCOMP and PREPDEPENDENT for complements of simple adjectives, simple nouns, adverbs and prepositions respectively
• the use of CLAUSECOMP only in the case of psychological/perception verbs e.g. the verb "÷áßñïìáé" [to be glad/happy] subcategorises for a clause complement introduced by "ğïõ" (÷áßñïìáé ğïõ Şñèåò [I am glad that you came])
• the encoding of an internal argument for all predicative adjectives.

Specific decisions for the instantiation of the general syntax model into Greek lie in the following points :

• use of the function ADVERBIAL in the following cases :

1. for prepositional phrases, weakly bound by the head, which can be substituted by adverbs.
1. for complements denoting: source, goal, manner, time, location, direction, measure etc.
1. for prepositional phrases which confer to the verb phrase a special meaning

• the functions NCOMP, ACOMP, VCOMP and PREPDEPENDENT are used instead of the functions NCLAUSCOMP, NPREPCOMP, NOFCOMP, as grammatical realisation of the complement should not affect the labelling of the function. This strategy allows to preserve coherence and economy in the lexicon, as well as a better organisation, since the relevant information is kept at each entity of the model separately
• modifiers and determiners are not encoded, as it is the grammar that handles determination and modification
• idiomatic expressions are not encoded
• appositions are not coded, as this phenomenon is deemed regular and, therefore, handled by the grammar.

1. Strategy to describe control information

Control in Greek is the phenomenon that describes the referential identity between the subject of a subordinate clause - more specifically, na-clauses - and the subject or the indirect object of the main clause verb. Raising is encoded as a type of control, following the specifications. Thus, control and raising information is encoded in accordance to the model, i.e. using the feature COREF to be coded on the controller and on the controllee, distinguishing the type (control or raising) in the CONTROLT feature.

1. Bibliography

1. GENELEX report on the morphological layer - Version 3.3 (Genelex consortium, 1994)
1. GENELEX report on the syntactic layer - Version 4.0 (Genelex Consortium, 1993)
1. PP-PAROLE report : Blueprint of PAROLE guidelines to the encoding of syntactic information in the lexicon - Verb subcategorisation (Pisa 1996)
1. PP-PAROLE report : Blueprint of PAROLE guidelines to the encoding of syntactic information in the lexicon for nouns, adjectives, adverbs and prepositions (Paris-GSI 1996)
1. PP-PAROLE report : Greek instantiation of the PAROLE guidelines for non-verb subcategorisation (Athens 1996)
1. PP-PAROLE report : Greek instantiation of the PAROLE specifications for verb subcategorisation (Athens 1996)
1. PP-PAROLE report : Lexicon - Morphosyntactic specifications : Language Specific Instantiations (Parole consortium, 1996)
1. PP-PAROLE report : The Greek lexicon - Inflectional morphology (Athens 1995)

Constructions for the category VERB