Inflection of the Portuguese DELAS using FST
p. 35-51
Résumé
The Portuguese DELAF that was initially used in INTEX corresponded to the conversion of an inflected Portuguese dictionary generated by the DIGRAMA system.
In order to generate DELAF under INTEX, we have converted the dictionary of lemmas of DIGRAMA into INTEX formalism, obtaining a DELAS, and then constructed a set of inflectional finite state transducers based on the linguistic information represented in DIGRAMA.
We briefly present the main characteristics of the inflectional formalism in DIGRAMA and INTEX, and we discuss some of the difficulties found in representing in the DELAS some of the linguistic information formalized in the DIGRAMA dictionaries.
Remerciements
I am grateful to Prof. Elisabete Ranchhod for her valuable linguistic advises and careful review of this paper, and to Prof. Samuel Eleutério for having corrected DIGRAS and generated new versions of DIGRAF whenever necessary, making possible the comparison between the DELAF generated by INTEX and the DELAF converted from the DIGRAF dictionary. I also would like to thank Paula Carvalho for her daily support.
This research was partly supported by FCT (Ph.D. scholarship, Ref. SFRH/BD/3237/2000 and Project ENLEX – Enhancement of Large-scale Lexicons, Ref. POSI/PLP/34729/99).
Texte intégral
Introduction
1DIGRAMA (Eleutério et al., 1995) is a system of electronic dictionaries and grammars for Portuguese developed by LabEL.
2The linguistic resources of DIGRAMA, as well as the linguistic resources of INTEX (Silberztein, 1993 and 2001), are composed, among other elements, by large coverage electronic dictionaries that can be applied to texts of several millions of words. These dictionaries are lists of inflected words, simple and compound, generated given dictionaries of canonical forms and a set of inflectional paradigms.
3The first version of the Portuguese DELAF was created for INTEX from the inflected dictionary of DIGRAMA, as depicted in Figure 1.
Figure 1

4In order to inflect the Portuguese DELAF under INTEX, we have, on the one hand, converted the dictionary of lemmas from DIGRAMA into a dictionary of lemmas represented in INTEX formalism, and, on the other hand, constructed a set of inflectional graphs based on the inflectional paradigms described in DIGRAMA, as it can be seen in Figure 2
Figure 2

5In the following sections, we will begin by briefly describing the characteristics of the inflectional systems of DIGRAMA and INTEX (sections 2. and 3., respectively). Then, we will show the several steps involved, first, in the conversion of the DIGRAS dictionary into the DELAS dictionary (section 4.), and, second, in the creation of the inflectional graphs (section 5.). To conclude, we will give some numbers concerning the results we have reached (section 6.).
1- The inflectional system of DIGRAMA
6As previously mentioned, a dictionary of inflected forms, DIGRAF, is generated in DIGRAMA, given a dictionary of canonical forms, DIGRAS, and a formal description of the inflectional paradigms.
1.1. Formalization of the dictionary of lemmas
7In DIGRAMA, each entry of the dictionary of lemmas has the following format:
8DEF category base_form PARAD inflectional_code info
9where category is the syntactic category of the word being described, base_form is a combination of the word’s lemma and some special (or control) characters that give some information on how the word inflects, inflectional_code is an alphanumeric code representing the inflectional paradigm, and info is a list, possibly empty, containing syntactic and semantic information.
10Some examples of entries of the DIGRAMA dictionary of lemmas are:
(1a) | DEF NOME afilhad<o PARAD n001 (nivel (1), traco (hum)) |
(1b) | DEF NOME coelh<o] PARAD n001_dh001_dt001 (nivel (1)) |
(1c) | DEF NOME palha<ç] o PARAD n003 dh005 dt005 (nivel (1)) |
(1d) | DEF NOME in [certeza< PARAD n300 (nivel (1)) |
(2) | DEF ADJ mono.s [silábic<o PARAD adj001 |
(3a) | DEF VERBO aconselh<a.r PARAD v1f01t (nivel (1)) |
(3b) | DEF VERBO aque<c] e.r PARAD v2f03t (nivel (1)) |
11From the syntactic point of view, the dictionary entries (1a)- (1d) associate the syntactic category NOME (noun) to the lemmas they describe: afilhado (godson), coelho (rabbit), palhaço (clown) and incerteza (uncertainty), entry (2) associates ADJ (adjective) to monossilábico (monosyllabic) and entries (3a) and (3b) associate the category VERBO (verb) to aconselhar (to advise) and aquecer (to heat).
12At the morphological level, all examples include the character "<" in the lemma, which separates the subsequence of the lemma that will remain unchanged from the one that will be substituted during the inflectional process. In a general way, this character identifies the stem (entries (1a), (1b), (1d), (2) and (3a)); when such is not the case, mainly due to phonological reasons, the character "]" identifies the stem and "<" the subsequence to be substituted (examples (1c) and (3b)). Moreover, the character " [" is used to separate a prefix from the stem (entries (1d) and (2)) and the character "." has a double meaning: before the stem indicates the presence of a connecting morpheme between the prefix and the base form of a derivation (entry (2)), whereas after the stem is used only with verbs and isolates the thematic vowel (entries (3a) and (3b)).
13In addition, all entries but (2) are considered as belonging to the common lexicon (nivel (1) represents the level of the word) and entry afilhado is subcategorized as human (hum is a syntactic-semantic mark).
1.2 - Representation of the inflectional information
14The inflectional paradigms are described by structures with the following format:
DEF PARAD paradigm category (inflection type1(inflection11,..., inflection 1n), -
..., -
inflection typem(inflectionm1,..., inflectionmk))
15where paradigm is the name of the inflectional paradigm as referred in the description of the dictionary entry shown above, category is the syntactic category of the forms inflected by the paradigm being described, inflection_typei (i=1... m) depends on the syntactic category and inflectionij ... inflectionis (i=1... m, j=1... m and and s=n... k) are the subsequences that will be added to the lemma form, after removing the sequence marked on the lemma to the right of the character "<", thus generating the inflected forms. One of these subsequences begins by the character @ indicating that the form inflected using this subsequence is the canonical form. Moreover, these subsequences may also begin by the character # to indicate that, if the canonical form has an acute or circumflex accent, the inflected form will not have the accent.
16For instance, the above paradigm n001 is thus represented by the following structure:
DEF PARAD n001 NOME (GER (@o, a, o.s, a.s))
17which in combination with the information in the dictionary entry generates the inflected forms:
afilhado: - [afilhado] N001. N: ms{@}
afilhada: - [afilhado] N001. N: fs
afilhados: - [afilhado] N001. N: mp
afilhadas: - [afilhado] N001. N: fp
18Each inflected form was associated with its syntactic category, in the example N (noun), its lemma, afilhado, and its inflectional attributes: afilhado, afilhados and afilhadas are, respectively, the feminine singular (fs), the masculine plural (mp) and the feminine plural (fp) of the canonical form afilhado, which is masculine singular (ms).
19The dictionary of inflected forms will then be comprised of these entries.
2 - The inflectional system of INTEX
20As previously mentioned, in INTEX, the dictionary of inflected forms DELAF is generated given a dictionary of lemmas, named DELAS, and a set of inflectional graphs.
2.1 - Formalization of the dictionary of lemmas
21The entries of a DELAS dictionary have the following format:
lemma, inflectional_code + info1 +... + infon
22where lemma is the canonical form of the entry being described, inflectional_code is an alphanumeric code that indicates the syntactic category of the entry and its inflectional behavior, and info1 ... infonare various types of linguistic information that characterize the entry. The alphanumeric code is the name of an inflectional graph that describes how the word inflects.
23The dictionary entries for the words represented in DIGRAMA (section 1.1) correspond in INTEX to:
(1a) | afilhado, N001+Hum+z1 |
(1b) | coelho, N001_dh001_dt001+z1 |
(1c) | palhaço, N001 dh005 dt005 +z1 |
(1d) | incerteza, N300+z1 |
(2) | monossilábico, A001 |
(3a) | aconselhar, V101t+z1 |
(3b) | aquecer, V203t+z1 |
24As it can be seen by the examples, none of the entries contain explicit information about the morphological properties of the words, as it was the case in DIGRAMA. In INTEX, the inflectional behavior is only represented in the graphs. Another observable difference is that the inflectional codes of words that inflect in gender and in number in the same way, but have different diminutive forms due, for instance, to phonological reasons, only differ in what concerns the diminutive. Such is the case of the entries coelho and palhaço: in DIGRAMA they have respectively the codes n001 and n003 whereas in INTEX they both have the code N001 (refer to section 04.2.).
2.2. Inflectional description by graphs
25The nodes of an inflectional graph have as an input a sequence that will perform the necessary transformations on the lemma to generate the inflected forms, and as an output the corresponding inflectional information. The graph in Figure 3 shows the description of the inflectional paradigm N001.

Figure 3
26The input of the first node is empty, which means that no operation will be performed on the lemma afilhado. So, the inflected form is equal to the lemma and it is associated to the morphological information ms (masculine and singular). The input of the second node is a sequence beginning by the number 1 which will delete one character from the end of the lemma, obtaining pat; the remaining input sequence is then a, which will be added to the modified lemma, generating the inflected form afilhada. This form will be associated with the nodes output fs (feminine and singular). The same procedure applies to generate the other two inflected forms.
27Given the entry afilhado and this graph four inflected forms are generated:
afilhado, afilhado. N+Hum+z1: ms
afilhada, afilhado. N+Hum+z1: fs
afilhados, afilhado. N+Hum+z1: mp
afilhadas, afilhado. N+Hum+z1: fp
28which make part of the DELAF dictionary.
29In the inflectional graphs it is also possible to name other graphs, allowing the construction of more complex inflectional paradigms in a modular way. For instance, a graph such as:

Figure 4
30describes the inflection of an adjective (graph A001 inflects the adjective in gender and number) which has diminutives (generated using graphs dh001 and dt001) and superlative (inflected with graph ss001).
3 - Conversion of dictionaries
31In order to automatically convert the dictionaries of lemmas from one system into the other, two scripts were written in AWK.
32The first script converts each entry on the DIGRAMA dictionary:
DEF category base form PARAD inflectional code (info1 (val1), ..., infon (valn))
33into INTEX formalism:
lemma, inflectional code +val1 +... + valn
34- lemma is equal to base_form without the DIGRAMA control characters;
35- the inflectional_code remains unchanged;
36- if infoi is equal to "nivel", it will be added +zvali to the info list; if infoi is "traco" then +vali will be added; any other information is ignored.
37Applying this first script to the following DIGRAMA entries:
DEF VERBO ata<c] a.r PARAD v1f06t (nivel (1))
DEF NOME diafragma< PARAD n300 (nivel (1), traco (Anat))
DEF ADJ fléx<il] PARAD adj129 (equiv (flexível))
DEF PRON DEM aquel<e PARAD pn dem2 (nivel (1))
DEF DET ART DEFINIDO <o PARAD dt artd1 (nivel (1))
38results in the following INTEX dictionary:
atacar, v1f06t+z1
diafragma, n300+Anat+z1
fléxil, adj129
aquele, pn dem2+z1
o, dt artd1+z1
39As it can be observed, there is information that can not be represented in DELAS. In particular, it is not possible to express that flexível is an alternative spelling of the adjective fléxil (flexible).
40Since in INTEX (i) the syntactic categories are written in uppercase, (ii) that some DIGRAMA codes are more appropriately represented in INTEX by a code followed by a syntactic property and (iii) in order to make some minor adjustments due to differences between the two formalisms, a second script was written. This script transforms the first obtained INTEX dictionary in a final DELAS. Hence, the previous dictionary entries become the final entries:
atacar, V106t +z1
diafragma, N300+Anat+z1
pequinense, A101
aquele, PRO04+Dem+z1
o, DET03+Art+Def+z1
41The final DELAS has about 110 000 entries1.
4 - Representation of the inflectional information
42The inflectional information represented in DIGRAMA was converted in inflectional graphs in a semi-automatic way by means of Word and Excel macros and several AWK scripts.
4.1- General procedure
43A simple paradigm, containing only one type of inflection, such as:
DEF PARAD scode scat (inflection type (inflection1,..., inflectionn))
44leads to a simple inflectional graph such as:

45For instance, the paradigm structure:
DEF PARAD n001 NOME (GER (@o, a, o.s, a.s))
46is converted in inflectional graph represented in Figure 5.

Figure 5
47If the paradigm is composed by different types of inflectional information, such as:
DEF PARAD scode scat (inflection type1 (inflection11,..., inflection 1n), -
..., -
inflection_typem (inflectionm1,..., inflectionmk))
48a main graph (see Figure 6, first graph on the left) composed by as many sub-graphs as the different types of inflection is created, as it is illustrated in Figure 6.

Figure 6 - scode.grf, inflection_type1.grf and inflection_typem.grf
49Therefore, the inflectional description:
DEF PARAD n001 dh001 dt001 NOME -
(GER (@o, a, o.s, a.s), -
DIM inho (#inh.o, #inh.a, #inh.o.s, #inh.a.s), -
DIM ito (#it.o, #it.a, #it.o.s, #it.a.s))
50is equivalent to the inflectional graphs:

Figure 7 - N001_dh001_dt001.grf, N001.grf, dh001.grf and dh001.grf
4.2 - Particularities due to diminutives, augmentatives and superlatives
51As previously mentioned, in DIGRAMA, the character # is used to indicate that an inflected form will not have the acute or circumflex accent of the canonical form. This feature of DIGRAMA has particular interest to formalize the inflectional paradigms of words that have, for instance, diminutives. So, accented words, such as tábua (board), amêndoa (almond) or pêra (pear), can have the same inflectional code of an unaccented word such as mesa (table): n301 dh301 dt301. In what concerns the diminutives, they can be generated by removing the morpheme -a and by adding the morphemes -inha/ita (singular diminutive) and -inhas/itas (plural diminutive), as it is represented in the following structure:
DEF PARAD n301 dh301 dt301 NOME -
(GER (, @a], , a] s), -
DIM inho (, #inh.a, , #inh.a.s), -
DIM ito (, #it.a, , #it.a.s))
52If the diminutive morphemes were not preceded by the character #, the incorrect forms:
tábu (inha, inhas, ita, itas)
amêndo (inha, inhas, ita, itas)
pêr (inha, inhas, ita, itas)
53would be generated, instead of the correct forms:
tabu (inha, inhas, ita, itas)
amendo (inha, inhas, ita, itas)
per (inha, inhas, ita, itas)
54In INTEX, however, the only way to represent this behavior is by creating different inflectional graphs for each different accent and the possible position where it will be found. Since this is not very reasonable to do, because it would unnecessarily increase the number of different graphs, the solution that we adopted was to generate the inflected forms with the accents and then remove the accents from the diminutive, augmentative and superlative forms by applying an AWK script.
55Words such as anjo and barco, which are masculine nouns, inflect in number by adding the morpheme -s and both have diminutive forms created by adding the morpheme -inho; in the case of barco, the diminutive will have to be created by changing the "c" to "qu". As expected, both in DIGRAMA and in INTEX the inflectional codes to create the diminutive forms are different. However, in DIGRAMA, the code to inflect in number is also different:
DEF NOME anj<o] PARAD n201 dh201 dt201
DEF NOME bar<co] PARAD n202 dh204 dt204
56This happens because, in order to account for the creation of the diminutive, the control character that indicates which part of the lemma will remain unchanged has to be placed immediately after bar. Therefore, the inflection in number will have to be GER (@co], , co] s, ) instead of GER (@o], , o] s, ) as it is the case for anjo.
57On the contrary, in INTEX, all the inflectional process is represented in the graphs and the inflectional in number can be independent from the generation of the diminutive forms. Consequently both entries can have the same inflectional code in gender and only the diminutive codes are different:
anjo, n201 dh201 dt201
barco, n201 dh204 dt204
58This behavior is common to many other entries, not only because of the diminutives but also due to augmentatives and superlatives. As a result, it was possible to decrease the number of inflectional codes after adjusting, mainly manually, the inflectional codes in INTEX.
4.3 - Verbal forms and clitics
59When inflecting a verbal form in Portuguese, one has to account for the possibility of having a clitic bound to the verbal form. For instance, the first person plural of the simple past tense of a verb such as comprar (to buy) is comprámos (bought). Yet, when bound to an accusative clitic form such as o (it), the verbal form is comprámo, which is a truncated form of comprámos. Moreover, the clitic also has to be the variant lo of the clitic o. This type of information is coded in DIGRAMA by the letters A, L, N, O, Q, S, Y and Z, which indicate the form of the pronouns which co-occur with the corresponding verbal forms, as it can be seen in the description of the inflection of the simple past tense (ipps) of a verb such as comprar:
ipps (SAQO: ei, SAQO: a.ste, SAQO: ou, -
SA: á.mos=QL: á.mo, SAQ: a.stes=L: a.ste, SAQN: a.ra.m)-
60The equal sign states that both forms comprámos and comprámo are the first person plural. One, comprámos, is used with the clitic forms S or A, and the other, comprámo, is used with the pronouns marked Q or L.
61For now, in INTEX, only the possible truncated forms of the verbs and the variants of the clitics were coded. The relation between them was not yet integrated since there is not an adequate way of representing and using this information in the system. Hence, the two inflected entries:
comprámos, comprar. V: J1s
comprámo, comprar. V: J1s
62do not contain information that distinguishes them regarding the use of the correct clitic pronoun.
Final Remarks
63The aim of this work was the construction of a dictionary of lemmas that could be inflected in INTEX using inflectional graphs. The source information used to accomplish this goal was the dictionary of lemmas and the inflectional paradigms coded in the DIGRAMA system.
64In a first stage, the dictionary of lemmas was converted from one system to the other using two scripts written in AWK. The first script makes the conversion of the formalisms, and the other makes some necessary adjustments. The DELAS obtained has about 110 000 entries.
65The second stage of our work consisted in representing, using INTEX inflectional graphs, the inflectional information described in the DIGRAMA system. This process lead to the construction of 3467 graphs: 1095 are main graphs, used to code some dictionary entry, and the remaining are only sub-graphs.
66Finally, given both the DELAS and the inflectional graphs we were able to inflect the DELAF dictionary in INTEX (versions 4.22-4.23e) obtaining about one million entries.
Bibliographie
Des DOI sont automatiquement ajoutés aux références bibliographiques par Bilbo, l’outil d’annotation bibliographique d’OpenEdition. Ces références bibliographiques peuvent être téléchargées dans les formats APA, Chicago et MLA.
Format
- APA
- Chicago
- MLA
References:
10.1075/li.19.1.04ele :Eleutério S.; Ranchhod E.; Freire H.; Baptista J. (1995). “A System of Electronic Dictionaries of Portuguese”, Lingvisticae Investigationes, XIX: 1, pp.57-82, Amsterdam/Philadelphia: John Benjamins.
Ranchhod E.; Eleutério S. (1996). Construção de Dicionários Electrónicos do Português. Problemas Teóricos e Metodológicos. In Actas do Congresso Internacional sobre o Português, pp. 265-282, Lisboa: Colibri.
Ranchhod E. (1998), “Dicionários Electrónicos e Análise Lexical Automática”, Linguística Computacional. Investigação Fundamental e Aplicações (Orgs. Palmira Marrafa e Maria Antónia Mota), pp. 221-233, APL, Lisboa: Colibri.
Ranchhod E.; Mota C.; Baptista J. (1999), "A Computational Lexicon of Portuguese for Automatic Text Parsing", SIGLEX99: Standardizing Lexical Resources, 37th Annual Meeting of the ACL (College Park, Maryland, USA, June 20-26, 1999), pp. 74-80.
Ranchhod E. (1998-1999), “Ressources linguistiques du portugais implémentées sous INTEX” Lingvisticae Investigationes, XXII, Amsterdam/Philadelphia: John Benjamins Publishing Company, pp. 263-278
Ranchhod E. (2001), “O uso de dicionários e de autómatos finitos na representação lexical das línguas naturais”, in Ranchhod, E. (ed.): Tratamento das Línguas por Computador, Caminho, Lisboa, pp. 13-47.
Silberztein M. (1993), Dictionnaires électroniques et analyse automatique de textes: le système INTEX, Masson Ed.: Paris.
Silberztein M. (1996), INTEX 3.4 reference manual, LADL, Université Paris 7.
Silberztein M. (1998), "Finite State Transducers and the processing of natural languages", distributed with INTEX version 4.12.
Silberztein M. (2001), INTEX manual (available from the WEB site: www.nyu.edu/pages/linguistics/intex), unpublished.
Notes de bas de page
1 It should be stressed that DIGRAS contains about ten thousand more entries than DELAS, mainly due to the fact that the words with Brazilian-Portuguese spelling contained in DIGRAS were not included in the DELAS dictionary.
Auteur
-
Cristina Mota
LabEL/Instituto Superior Técnico Av. Rovisco Pais, 1 - 1049-001 Lisboa, Portugal
Cristina@label.ist.utl.pt
Le texte seul est utilisable sous licence Licence OpenEdition Books. Les autres éléments (illustrations, fichiers annexes importés) sont « Tous droits réservés », sauf mention contraire.
La formation d’une opinion démocratique
Le cas du Jura, de la révolution de 1848 à la « république triomphante » (vers 1895)
Pierre Merlin
2017
Les mutations récentes du foncier et des agricultures en Europe
Gérard Chouquer et Marie-Claude Maurel (dir.)
2018
Deux frontières aux destins croisés ?
Étude interdisciplinaire et comparative des délimitations territoriales entre la France et la Suisse, entre la Bourgogne et la Franche-Comté (xive-xxie siècle)
Benjamin Castets Fontaine, Maxime Kaci, Jérôme Loiseau et al. (dir.)
2019
Un mousquetaire du journalisme : Alexandre Dumas
Sarah Mombert et Corinne Saminadayar-Perrin (dir.)
2019
Libertaire ! Essais sur l’écriture, la pensée et la vie de Joseph Déjacque (1821-1865)
Thomas Bouchet et Patrick Samzun (dir.)
2019
Les encyclopédismes en France à l'ère des révolutions (1789-1850)
Vincent Bourdeau, Jean-Luc Chappey et Julien Vincent (dir.)
2020
Le patrimoine industriel au prisme de nouveaux défis
Usages économiques et enjeux environnementaux
Marina Gasnier
2018
La petite entreprise au péril de la famille ?
L’exemple de l’Arc jurassien franco-suisse
Laurent Amiotte-Suchet, Yvan Droz et Fenneke Reysoo
2017
Une imagination républicaine, François-Vincent Raspail (1794-1878)
Jonathan Barbier et Ludovic Frobert (dir.)
2017