Version classiqueVersion mobile

Formaliser les langues avec l’ordinateur : de INTEX à Nooj

 | 
Svetla Koeva
, 
Denis Maurel
, 
Max Silberztein

Premiere partie. Formalisation de ressources linguistiques

10. Portuguese morphology with INTEX 4.33

Cristina Mota

Texte intégral

My many thanks to Elisabete Ranchhod for her suggestions and review of this paper. I am also grateful to Max Silberztein for having presented this paper at Sofia in my behalf, since I was unable to attend the INTEX Workshop. This work was partly supported by FCT, Project ENLEX – Enhancement of Large-scale Lexicons, Ref. POSI/PLP/34729/99 and POSI/Plurianual/CAUTL.

Introduction

1A major part of the Portuguese language resources that have been developed at LabEL (Laboratório de Engenharia da Linguagem), as described in [Ranchhod et al., 2004], comprise DELA type dictionaries of simple and compound words (common lexicon, acronyms and corresponding lexical description, named entity expressions, and insurance terminology), inflectional graphs, lexical graphs (for the recognition of roman, cardinal and ordinal numbers), local grammars (for the analysis of temporal and numerical expressions, and stock exchange characteristic linguistic patterns), and disambiguation grammars (for the unambiguous tagging of complex predicates and noun phrases containing adjectives).

2These resources are integrated, among other systems, in INTEX [Silberztein, 1993], since its earlier versions, keeping up with the system improvements and new functionalities.

3One of such enhancements was the morphological parser introduced with INTEX 4.33 [Silberztein, 2004]. This morphological parser is fully compatible with DELAS and DELAF dictionaries, allowing their combination with either inflectional graphs or the new morphological graphs.

4In the following sections, we discuss the use of these morphological graphs in the analysis of diminutive, augmentative and superlative forms, and show how the new morphological features can be used in the analysis of verb-clitic combinations.

I. Analysis of Diminutive, Augmentative and Superlative Forms

5Nouns and adjectives in Portuguese inflect in gender and number. Moreover, they may also accept diminutive, augmentative and superlative (adjectives) suffixes. These suffixes receive in turn their own gender and number morphemes.

Table 1 : Examples of diminutive, augmentative and superlative suffixing

Table 1 : Examples of diminutive, augmentative and superlative suffixing

6In what concerns diminutive and augmentative suffixing, two main types are possible : (i) the suffix is added to a word stem ; (ii) the suffix is added to an inflected word (the gender and number morphemes of the inflected word must agree with the gender and number morphemes of the suffix). Regarding superlative suffixing, the corresponding morphemes are usually suffixed to the word stems.

1. Representation by inflectional graphs

7Prior to the last version of the morphological parser of INTEX 4.33, nouns and adjectives accepting grade variation were being recognized by introducing a code in the DELAS entries that allowed the generation of the corresponding diminutive, augmentative and superlative forms in the DELAF dictionary [Mota, 2004].

  • 2 Accented vowels in Portuguese may have one of three accents : acute (á, é, í, ó, ú), grave (à), and (...)

8Nonetheless, this solution, as already discussed in [Mota, 2004], was not completely satisfactory due to the suffixing of words with graphic accent2 : the derived (diminutive/superlative) forms loose that accent. For instance :

célula [cell] → celulazinha ; côdea [crust] → codeazinha ;
dúvida [doubt] → duvidazinha ; lágrima [tear] → lagrimazinha ;
pá [shovel] → pazinha ; recaída [relapse] → recaidazinha.

9In order to generate these diminutive forms through inflectional graphs, the base forms must have different inflectional codes. In this case, each graph deletes the adequate number of letters before substituting the accented vowel by the vowel without the accent. For instance, the graph for célula would delete “élula” and then add “elulazinha”, and the graph for dúvida would delete “úvida” and then add “uvidazinha”. This solution clearly increases the number of necessary inflectional graphs. To avoid this situation, all these forms will have the same inflectional code. Thus, the corresponding graph just adds the suffix -zinha to all of them and then an AWK script substitutes the accented vowel by the unaccented vowel, hence obtaining the final DELAF.

2. Representation by morphological graphs

10The morphological parser of INTEX 4.33 makes the representation of the suffixing process more uniform. Figure 1 illustrates a simplified version of the graph that analyses diminutive forms.

  • 3 The grade variation attributes produced by the morphological graphs are being included as syntactic (...)

11As it can be seen, the keyword <N> is being used to prevent INTEX of matching accented letters, and whenever a vowel is found it is substituted by the same vowel with an accent. For instance, if celulazinha is being analyzed, variables $BeforeAccent, $Vowel and $AfterAccent will be assigned the values “c”, “e”, “lula”, respectively, and the form célula ($BeforeAccent#é#$AfterAccent) is created so INTEX can lookup the dictionary and verify it is a singular feminine noun (N : fs). Then, the dictionary entry celulazinha, célula. N +Dim:fs3 is generated.

Figure 1 : Simplification of a morphological graph for the analysis of - zinha diminutives

Figure 1 : Simplification of a morphological graph for the analysis of - zinha diminutives

12After applying the morphological graph in conjunction with the DELAF dictionary, the morphological parser recognizes both the diminutives created from non-accented words :

aldeiazinha, aldeia. N +Dim : fs
aventurazinha, aventura. N +Dim : fs

13as well as the diminutives created from accented words :

celulazinha, célula. N +Dim : fs
codeazinha, côdea. N +Dim : fs
duvidazinha, dúvida. N +Dim : fs
lagrimazinha, lágrima. N +Dim : fs
pazinha, pá. N +Dim : fs
recaidazinha, recaída. N+Dim : fs

14Besides recovering the accented vowel, it may be necessary to recover the original form of the stem and also the gender and number morphemes. For instance, the diminutive of the word marca (mark) is marquinha. Using the graph of Figure 2 variables $NoAccents and $Part will be assigned “mar” and “qu”, respectively, and the word marca ($NoAccents#ca") will be hence generated.

Figure 2 : Simplification of a morphological graph for the analysis of - inha diminutives

Figure 2 : Simplification of a morphological graph for the analysis of - inha diminutives

15The suffixes considered in the morphological graphs were : (i) the diminutive suffixes -inho, -ito, -ico, -eco, -ote, -zinho, -zito and - zeco ; (ii) the augmentative suffixes -ão, -alhão, -arrão and -aço ; (iii) the superlative suffixes -íssimo -ílimo and -érrimo, all allowing the variation in gender and number.

16Three distinct structured graphs were created to represent each suffixing process. After minimization and determinization, the diminutive suffixing transducer has 128 states and 355 transitions, the augmentative transducer has 79 states and 196 transitions, and the superlative transducer has 79 states and 192 transitions.

  • 4 Only about 50 forms (from approximately 11 000) are not being recognized, mainly due to the fact th (...)

17Almost all the diminutive, augmentative and superlative forms included in the DELAF4 are recognized by the morphological graphs. Nevertheless, there are misleading results. For instance, the word batota (cheat) is being analyzed as the diminutive of bata (smock). Even though this analysis is formally correct it is not acceptable, since the diminutive forms of bata are batinha or even batazinha.

18The morphological graphs were also applied to a list of words not recognized by DELAF in the 80 000 first excerpts of the CETEM Publico corpus [Santos & Rocha, 2001]. This unknown word list mainly includes misspelled words, acronyms, proper names, neologisms, and other words resulting from productive affixing processes. After application of the morphological graphs, we observed that the main source of error was proper names that can be formally associated with diminutive or augmentative forms of common words (e.g., Milhaços/milho [corn], Espadinha/espada [sword]). Other examples of errors are : (i) misspelled words that unexpectedly are equivalent to unusual diminutive or augmentative words (e.g., praticos/prato [plate], instead of prático [pratical]) ; (ii) incorrect recovery of the word stem (e.g., rocinha is being analyzed as the diminutive of roça [plantation], which is correct, but also as a diminutive of roca [distaff], which appropriate diminutive is roquinha due to phonetic reasons).

3. Which representation should be chosen ?

19The main advantages and disadvantages related to each approach are pinpointed in Table 2.

Table 2 : Inflectional graphs versus morphological graphs in the analysis of diminutive, augmentative and superlative forms

Table 2 : Inflectional graphs versus morphological graphs in the analysis of diminutive, augmentative and superlative forms
  • 5 The graphs must have normal priority because there are homographs of these forms, e.g., caminha (ca (...)

20Two solutions seem possible : (a) remove the diminutives, augmentatives and superlative forms from the DELAF and analyze them with normal priority morphological graphs5 ; (b) keep the diminutives, augmentatives and superlatives in the DELAF and use the morphological graphs with low priority, so they would only be applied to unknown words.

21The first solution entails very restrictive morphological graphs, otherwise one looses precision. In such case, the graphs become very complex and it is necessary to add accurate information (similar to the inflectional codes) to the DELAS entries.

22The second approach is more flexible. The morphological graphs can be conceived in such a way that would help easily en-larging the DELAS. We have opted for this latter solution.

II. Analysis of Verb-Clitic Combinations

  • 6 Future and conditional verbal forms are subject to a tmesis process : Eles comprarão os livros [The (...)

23In declarative sentences the clitic pronouns immediately follow the verbal form and are bound to it by a hyphen6 (e.g., Ele comprou alguns livros [He bought some books] / Ele comprou-os [He bought them]). Due to historical phonetic reasons, both the clitic and the verb may undergo formal modifications.

24The modification of the clitic pronouns o (3ms), a (3fs), os (3mp), as (3fp) depends on the verbal form termination. Thus, if the termination is :

  • a vowel or an oral diphthong, the clitic forms do not undergo any modifications : o, a, os, as (e.g., Ele comprou um livro ontem [He bought a book yesterday] / Ele comprou-o ontem [He bought it yesterday]) ;

  • a nasal diphthong, the clitic forms change to : no, na, nos, nas (e.g., Eles compraram um livro ontem [They bought a book yesterday] / Eles compraram-no ontem [They bougth it yesterday]) ;

  • -r, -s or -z, the clitic forms change to : lo, la, los, las. In this context, the verbal forms are also modified, loosing the final consonant. The vowel preceding the -r forms, will receive an accent (acute or circumflex depending on the thematic vowel of the verb). For instance, Nós comprámos um livro ontem [We bought a book yesterday] / Nós comprámo(s)-lo ontem [We bought it yesterday]

25In the presence of the reflexive and dative pronoun nos (1p) the first and second plural verbal forms ending in -s are also modified ; the clitics do not suffer modifications (e.g., Nós lavámo(s)-nos [We washed ourselves]).

1. Verb and clitic formalization

26In order to be able to recognize the modified verbal forms, they are also described in the verb inflectional graphs and conesquently are generated simultaneously with the non-modified forms when the DELAF is created.

Table 3 : Simple Present of the Verb comprar (to buy)*

Table 3 : Simple Present of the Verb comprar (to buy)*

* The entries containing underlined inflectional information correspond to verbal forms that are modified by the presence of clitics

  • 7 This type of information was already present in DIGRAMA dictionaries [Eleutério et al., 1995].

27However, the INTEX 4.2x DELAF version did not contain information7 that allowed (i) distinguishing the two forms and (ii) guaranteeing the correct combination of the verbal form with the clitic.

28Similarly to DIGRAMA dictionary entries, the 11413 DELAS verbal entries have now been assigned with such information. The possible clitic codes are :

  • i, the verbal form may occur without clitics ;

  • c, neither the verb nor the clitic are modified ;

  • o, verbal form co-occurs with clitic forms o, a, os, as ;

  • l, verbal form co-occurs with clitic forms lo, la, los, las ;

  • n, verbal form co-occurs with clitic forms no, na, nos, nas ;

  • q, verbal form co-occurs with reflexive and dative forms nos.

29This information is enclosed between square brackets in the inflectional features field, as it can be seen in Table 3.

30Furthermore, the accusative, dative and reflexive clitic dictionary entries were assigned with the corresponding clitic codes c, o, l, n, q (e.g., nos, eu. PRO : 1p[q] : 2’mp[n] : 3mp[n]).

2. Disambiguation of verb-clitic pairs

31The introduction of the clitic codes in the dictionary entries allows disambiguating verb-clitic pairs, since these codes can be used in disambiguation grammars to specify the adequate verb-clitic combination (cf. Figure 3).

Figure 3 : Simplified grammar for disambiguating verb-clitic pairs

Figure 3 : Simplified grammar for disambiguating verb-clitic pairs

32Consider, for instance, the pair dá-os (give them). As graph in Figure 4 shows, this sequence is ambiguous : the verbal form can be a form which co-occurs with either l clitics (e.g., second person singular of simple present, P2s[l]) or s, c, q, and o clitics (e.g., third person singular of simple present, P3s[scoq]) ; the o form can be either a determiner, a demonstrative pronoun or a clitic.

Figure 4 : Verb-clitic combination before and after the application of the graph of Figure 3. The strikethrough analyses correspond to analyses that are eliminated.

Figure 4 : Verb-clitic combination before and after the application of the graph of Figure 3. The strikethrough analyses correspond to analyses that are eliminated.

33After applying the disambiguation grammar, the verb inflectional fields that do not contain the feature o are eliminated, given that the personal pronoun is an o clitic ; in such case, the analyses of the form o as a demonstrative pronoun and determiner are also eliminated.

3. Verb-clitic morphology in syntactic transformations

34Exploiting the new features of the morphological parser, the clitic codes can also be used in syntactic transformations to obtain the correct forms of the verbs and clitics. For instance, the graph in Figure 5 describes that when transforming the negative context into a declarative context, the second person singular of the verb form, which is preceded by the clitic, is replaced by the form which occurs with l clitics ($V_P2s[l]).

Figure 5 : Analysis of the second person singular of simple present with clitic in a negative context. Substitution by declarative context.

Figure 5 : Analysis of the second person singular of simple present with clitic in a negative context. Substitution by declarative context.

35So, if this graph is applied to the sentence O jornal, tu não o compras todos os dias (The newspaper, you don’t buy it everyday), the declarative sentence O jornal, tu compra-lo todos os dias (The newspaper, you buy it everyday) is generated.

Conclusion

36We have shown that the INTEX 4.33 morphological graphs can be used to analyse Portuguese diminutive, augmentative and superlative forms, in particular those having accented base forms. However, since the inflectional graphs give more accurate results, we have opted to keep the DELAS entries unchanged, generating those forms in the DELAF. The morphological graphs will then be used to help in the formalization of diminutive, augmentative and superlative forms that are not already in the DELAS.

37We have also described the clitic codes that were added to the inflectional field of verb and clitic forms, and how this information can be used to disambiguate verb-clitic combinations, as well as to generate the adequate verb-clitic pair in the context of a syntactic transformation, taking advantage of the enhanced features of the new morphological parser.

Bibliographie

Bibliography

ELEUTÉRIO Samuel, RANCHHOD E., FREIRE H., BAPTISTA J. (1995), “A System of Electronic Dictionaries of Portuguese”, in Lingvisticae Investigationes, XIX:1, Amsterdam / Philadel-phia, John Benjamins, p. 57-82.

Mota, Cristina (2004), “Inflection of the Portuguese DELAS using FST”, in C. MULLER, J. ROYAUTE, M. SILBERZTEIN, INTEX Pour la Linguistique et le Traitement Automatique des Langues, Série Archive, Bases, Corpus, nº1, Besançon, Presses Universitaires de Franche-Comté, France, p. 35-51.

RANCHHOD Elisabete, CARVALHO P., MOTA C., BARREIRO, A. (2004), “Portuguese Large-scale Language Resources for NLP Applications”, in Proceedings of the 4th LREC, Lisbon, p. 1755-1758.

SANTOS Diana & ROCHA Paulo (2001), “Evaluating CETEM Público, a free resource for Portuguese”, in Proceedings of the 39th Annual Meeting of the Association for Computational Linguistics, Toulouse, 9-11 July 2001, p. 442-449.

SILBERZTEIN Max (1993), Dictionnaires électroniques et analyse automatique de textes : le système INTEX, Masson Ed., Paris.

SILBERZTEIN Max, INTEX Manual (English version), http://msh.univ-fcomte.fr/intex/downloads/Manual.pdf [last visited : February, 2005].

Notes

2 Accented vowels in Portuguese may have one of three accents : acute (á, é, í, ó, ú), grave (à), and circumflex (â, ê, ô); a tilde can also appear to mark nasalization (ã, õ). There are a few words with an acute accent and a tilde (e.g., órfão [orphan]).

3 The grade variation attributes produced by the morphological graphs are being included as syntactic-semantic attributes (+Dim, +Aum, +Sup) to distinguish them from the attributes produced by inflectional graphs, which are integrated in the inflectional field ( : D, : A, : S, respectively).

4 Only about 50 forms (from approximately 11 000) are not being recognized, mainly due to the fact that they result from less productive suffixing processes, which are not yet represented in the graphs.

5 The graphs must have normal priority because there are homographs of these forms, e.g., caminha (cama, N [bed]/caminhar, V [walk]).

6 Future and conditional verbal forms are subject to a tmesis process : Eles comprarão os livros [They will buy the books] / Eles comprá-los-ão [They will buy them]. However we will not discuss such process.

7 This type of information was already present in DIGRAMA dictionaries [Eleutério et al., 1995].

Table des illustrations

Titre Table 1 : Examples of diminutive, augmentative and superlative suffixing
URL http://books.openedition.org/pufc/docannexe/image/27069/img-1.jpg
Fichier image/jpeg, 24k
Titre Figure 1 : Simplification of a morphological graph for the analysis of - zinha diminutives
URL http://books.openedition.org/pufc/docannexe/image/27069/img-2.jpg
Fichier image/jpeg, 27k
Titre Figure 2 : Simplification of a morphological graph for the analysis of - inha diminutives
URL http://books.openedition.org/pufc/docannexe/image/27069/img-3.jpg
Fichier image/jpeg, 12k
Titre Table 2 : Inflectional graphs versus morphological graphs in the analysis of diminutive, augmentative and superlative forms
URL http://books.openedition.org/pufc/docannexe/image/27069/img-4.jpg
Fichier image/jpeg, 37k
Titre Table 3 : Simple Present of the Verb comprar (to buy)*
Légende * The entries containing underlined inflectional information correspond to verbal forms that are modified by the presence of clitics
URL http://books.openedition.org/pufc/docannexe/image/27069/img-5.jpg
Fichier image/jpeg, 50k
Titre Figure 3 : Simplified grammar for disambiguating verb-clitic pairs
URL http://books.openedition.org/pufc/docannexe/image/27069/img-6.jpg
Fichier image/jpeg, 20k
Titre Figure 4 : Verb-clitic combination before and after the application of the graph of Figure 3. The strikethrough analyses correspond to analyses that are eliminated.
URL http://books.openedition.org/pufc/docannexe/image/27069/img-7.jpg
Fichier image/jpeg, 13k
Titre Figure 5 : Analysis of the second person singular of simple present with clitic in a negative context. Substitution by declarative context.
URL http://books.openedition.org/pufc/docannexe/image/27069/img-8.jpg
Fichier image/jpeg, 5,5k

Auteur

Instituto Superior Técnico (Lisboa – Portugal)
Courriel : cmota@ist.utl.pt.

Le texte et les autres éléments (illustrations, fichiers annexes importés) sont sous Licence OpenEdition Books, sauf mention contraire.

Rechercher dans OpenEdition Search

Vous allez être redirigé vers OpenEdition Search