MorphoBr: an open source large-coverage full-form lexicon for morphological analysis of Portuguese

MorphoBr: um léxico de formas plenas de ampla cobertura para a análise morfológica do Português

Leonel Figueiredo de Alencar leonel.de.alencar@ufc.br

Universidade Federal do Ceará, Brasil

Bruno Cuconato bcclaro@gmail.com

EMAp/Fundação Getúlio Vargas, Brasil

Alexandre Rademaker arademaker@gmail.com

EMAp/Fundação Getúlio Vargas and IBM Research, Brasil

Esta obra está bajo una Licencia Creative Commons Atribución 4.0 Internacional.

Morphological analysis is a prerequisite for syntactic parsing with linguistically motivated formalisms like LFG (BUTT et al., 1999; FALK, 2001; DIPPER, 2003), HPSG (POLLARD; SAG, 1994), and GF (RANTA, 2011). Moreover, it is useful in a wide range of applications ranging from dictionary lookup in e-book readers and spell-checkers to sentiment analysis, opinion mining, information extraction, and text classification algorithms. It maps word forms to all possible representations consisting of lemma and grammatical features, which may be filtered by subsequent processing steps (BEESLEY; KARTTUNEN, 2003; DIPPER, 2003; JURAFSKY; MARTIN, 2009).

As Alencar et al. (2014, p. 59-60) observes, "finite-state transducers, due to compact storage and fast processing, have been a preferred implementation of morphological analyzers''. A finite-state transducer (henceforth FST) is a two-tape automaton. In the case of a morphological analyzer, the first tape encodes the analysis (or parse) strings and the second tape the surface strings, i.e. word forms (BEESLEY; KARTTUNEN, 2003).

Figure 1: Example of a single-path FST

Figure 1 exemplifies a simple FST relating the analysis string ver+V+PRF+1+SG to the surface string vi 'I saw', first person singular perfect indicative tense of Portuguese verb ver 'see'. This FST has 8 states connected by 7 arcs, constituting a single path from the initial state 0 to the final state 7. Each path in an FST represents an analysis string with its corresponding surface string. The arcs are labelled by symbol pairs of the form x:y, where x is a symbol of the analysis string and y is a symbol of the surface string. Labels of the form x:x are simplified to x. Label 0 represents the empty string.

Large lexical transducers with millions of paths were compiled for many languages, using a variety of finite-state tools (BEESLEY; KARTTUNEN, 2003; HULDEN, 2009), and used in diverse applications, for example in industrial-scale LFG grammars for deep syntactic parsing with the Xerox Linguistic Environment (XLE) (BUTT et al., 1999; DIPPER, 2003). One additional advantage of FSTs is that they are inherently bidirectional devices, so the same FST can serve both as a generator and an analyzer.

The construction of a wide-coverage morphological analyzer is a continuous task that goes through successive refinements. Ideally one starts from available electronic dictionaries consisting of several hundreds of thousands of word-parse pairs, so-called full-form lexicons, that can be compiled directly into an FST. However, one difficulty in reusing existent resources is that they usually adopt incompatible annotation schemes or fail to provide information that is needed for a full-fledged morphological analyzer. Besides, these resources may have errors and inconsistencies. Although many such deficiencies can be automatically corrected using simple text processing techniques, overcoming some of these problems may be less than trivial.

A more challenging task is expanding the coverage of existent dictionaries in order to deal with new words. Finite-state morphology is the standard approach to tackle this problem (BEESLEY; KARTTUNEN, 2003). It facilitates the computational modeling of grammatical regularities underlying inflection and word formation processes. From these models FSTs can be compiled for analysis and generation of new lemmas and word forms (ALENCAR et al., 2014).

In this paper we present MorphoBr, a full-form lexicon constructed from the combination, revision, and expansion of available free analog resources for Portuguese, mostly derived from Label-Lex (ELEUTÉRIO et al., 1995) and Unitex-PB (MUNIZ, 2004). This effort is part of a research project that aims at developing wide-coverage computational grammars for deep parsing of Portuguese texts. A first version of the resource is already freely available online under an open source, free software license¹.

The main advancement of MorphoBr in relation to previous resources is a finite-state component that almost triples the inventories of nouns and adjectives by computationally modeling the formation of diminutives with -(z)inh-. This is one of the most productive derivational processes in Portuguese. Diminutives mainly function as a means of expressing speakers' emotions and attitudes. As such, they are very common in emotive discourse. Therefore, processing diminutives should be a basic capability of systems dealing with sentiment analysis, opinion mining, etc.

In the next section, we briefly introduce the previous resources that were combined into MorphoBr. In section 3 we first give an overview of the present stage of MorphoBr, describing the format used and explaining the conversion from the source formats. We also detail errors and inconsistencies that we found in the resources and how we solved them. In section 4 we report on the finite-state implementation of diminutive formation in Portuguese. Section 5 presents evaluation experiments, showing that MorphoBr outperforms the resources it integrates. The last section sums up the main results and points out directions for further research.

LABEL-LEX is a collection of resources.² According to Eleutério et al. (1995), the LABEL-LEX dictionaries originate from Costa and Melo (1991). In Ranchhod, Mota, and Baptista (1999), the authors describe a new version of the system, composed of a lexicon and grammars. The lexical data are organized according to the formal complexity of the lexical units. In the current version, the lexicon is organized in three different files: LABEL-LEX-sw (inflected forms), LABEL-LEX-mw (multi-word forms), and LABEL-LEX-gr (grammars)³. Only the first one is freely available for download. The LABEL-LEX-sw version 4.1 contains 938,445 word-parse pairs, e.g. gato,gatar.V:P1s. In this entry, the lemma gatar 'to fail' and the verb category label V are assigned to the word form gato, classifying it as first person singular of the indicative present.

LABEL-LEX resources were expanded (with adaptations) and integrated in many different systems over time. To mention some of them UNITEX⁴, INTEX⁵, and FreeLing (PADRÓ; STANILOVSKY, 2012). In Garcia and Gamallo (2010), the authors describe the integration of Label-Lex dictionary into FreeLing and its conversion to the EAGLES tag schema (LEACH; WILSON, 1996). Besides the tags conversion, many other adaptations were done to mitigate conflicts between morphological decisions of the dictionaries and of the data collected from annotated corpora. The current Portuguese dictionary in FreeLing version 4.0 is composed of 1,214,093 word forms.

Later, Garcia et al. (2014) presents different dictionaries of the new orthography as well as a new freely available testing corpus, containing different varieties and textual typologies. The combined European and Brazilian Portuguese dictionary expanded the FreeLing dictionary with more 492,896 word-parse pairs. Considering all the additions and improvements to the Label-Lex distribution made by Garcia and Gamalo, we opted to use the Garcia and FreeLing files (henceforth GFL) instead of the Label-Lex distribution.

Following the European projects, the Brazilian NILC group developed their own lexical resource in the context of the Unitex-PB project (MUNIZ, 2004)⁶. It is divided into four files: DELAS-PB (single-word lemmas), DELAF-PB (inflected forms), inflectional graphs, and DELACF-PB (multi-word forms). Version 2 from May 2015 of DELAF-PB contains 9,072,338 word-parse pairs, of which more than 8 million are verbs, 80,000 are nouns, and 90,000 are adjectives.

Both FreeLing and DELAF-PB are freely distributed under free software, open source licenses. FreeLing dictionaries are obtained from different open-source external projects, so they have different copyright holders and different licenses than the rest of FreeLing packages. Nevertheless, both FreeLing's Portuguese dictionary and DELAF-PB are freely distributed under free software, open source licenses. The former is available under the General Public License 3, the same license of the original LABEL-LEX distribution⁷. The latter is distributed under the GNU Lesser General Public License 2.1⁸, which is approved by both the Free Software Foundation⁹ and the Open Source Initiative¹⁰. This means that end-users can freely use these libraries and developers can modify them and share the modified versions.

These two resources differ in how they handle derived words. While Unitex-PB lists all diminutives with -(z)inh- in DELAF-PB, FreeLing 4.0 applies affixation rules to analyze formations like dorzinha at runtime (i.e. during text analysis), correctly classifying them as diminutive forms of the corresponding lemma (dor 'pain' in the case at hand). However, FreeLing's dictionary contains more than 5,000 -(z)inh- diminutives derived from nouns and adjectives, including hundreds of completely regular formations (e.g. abelhinha, abertinho, and amorzinho, derived from abelha 'bee', aberto 'open', and amor 'love', respectively) and cases with predictable stem alterations (e.g. amiguinho from amigo 'friend' and aneizinhos from anel 'ring' in plural), which could be analyzed by means of such rules. This shows that FreeLing does not consistently handle regular diminutive formation in Portuguese by means of rules as a means to spare storage space.

The present version of MorphoBr comprises the most numerous word classes, namely, nouns, adjectives, adverbs, and verbs. The other classes will be incrementally added in the next development stages. The major part of the data derives from the corresponding entries from DELAF-PB, which were not only converted to the format described below, but also corrected, enriched, or supplemented. The rest of the data consists of the following sets: (i) the set difference between GFL and DELAF-PB and (ii) the expansions described in section 4.

Tokenization decisions affect what counts as candidate for entry in a full-form lexicon. This is specially the case with verb-clitic clusters in Portuguese, which in principle can be treated as a single unit or tokenized into two different units, each one with its own entry in the dictionary. In Portuguese, clitic pronouns can occur in three contexts: (i) left-adjacent to the verb in proclitic position, (ii) right-adjacent to the verb in enclitic position, or (iii) between infinitive base form¹¹ and inflectional endings in mesoclitic position, as exemplified in (1), (2), and (3), respectively. In these interlinear glosses, as in analogous examples in this paper, we use the Leipzig Glossing Rules¹², the most widespread conventions for glossing linguistic examples. While the clitic is separated from the verb by a blank in proclitic position, in enclitic and mesoclitic position it is separated by one and two hyphens, respectively.

(1) nos visitávamos

nos=visitávamos

REFL;1PL;ACC=visit:IMPF;1PL

'we visited ourselves'

(2) visitávamo-nos

visitávamo=nos

visit:IMPF;1PL=REFL;1PL;ACC

'we visited ourselves'

(3) visitar-nos-emos

visitar=nos=emos

visit=REFL;1PL;ACC=FUT;1PL

'we will visit ourselves'

In cases where the pronominal clitic is not separated from the verb by a blank, we adopted DELAF-PB's strategy of handling the verb-clitic cluster as a single token with its own entry in the dictionary, as exemplified in (4) and (5), glossed in (2) and (3). GFL adopts a different strategy in this regard, assigning clitic and verb form to different entries. Thus, when converting data from GFL to MorphoBr, verb entries like (6) were discarded, because this verb form is only used with an enclitic pronoun, as in (2). The canonical form is the one in example (1).

(4) visitávamo-nos,visitar.V+PRO:I1p

(5) visitar-nos-emos,visitar.V+PRO:F1p

(6) visitávamo visitar VMII1P0

The dictionary entry format used in MorphoBr corresponds to the standard format in finite-state morphology (BEESLEY; KARTTUNEN, 2003). It consists of pairs of the form (w, p) separated by a NEWLINE, where w is a word form and p is an analysis string, separated by a TAB. The analysis string consists of a series of tags, the first one being the lemma. The other tags represent morphosyntactic features¹³. Tags are separated by a plus sign (a dot is used to separate the features of clitics, see below). In (7), the lemma comprar 'to buy' is assigned to the verb form comprei, classified as first person singular of the indicative perfect tense:

(7) comprei comprar+V+PRF+1+SG

The tagset used in MorphoBr was designed to be more mnemonic than the formats of the two resources it draws upon. For this reason, tags are mnemonic abbreviations, generally following conventions common in the linguistic literature. Almost all tags were drawn from Fradin (s.d.), whose tagset represents an improvement in relation to the one of the Leipzig Glossing Rules¹⁴. By contrast, many tags in DELAF-PB, as shown in Table 1, are completely arbitrary single-letter abbreviations which are difficult to decode.

Table 1: Comparison between tags from DELAF-PB and MorphoBr.

Tags in FreeLing are less arbitrary compared to DELAF-PB. However, FreeLing departs from widely accepted assumptions in linguistics. First, it represents mood and tense by two different tags. However, in the morphological analysis of Portuguese, these two types of information are usually represented by one single inflectional suffix (MONTEIRO, 1987; ROCHA, 2008; VILLALVA; SILVESTRE, 2014).¹⁵ Second, the left-to-right order of the tags in FreeLing does not always reflect the concatenation of the morphemes representing the individual morphosyntactic features. This is the case with the diminutive forms of nouns, as in example (8) from FreeLing. In these forms, gender and number tags F and P (feminine and plural, respectively) precede diminutive tag D. However, in the surface form abelhinhas (diminutive of abelha 'bee'), glossed in (9), the diminutive suffix -inh- precedes the inflectional endings representing feminine plural.

(8) abelhinhas abelha NCFP00D

(9) abelh-inh-a-s

bee-DIM-F-PL

In entries for diminutive forms of adjectives, however, the tags in FreeLing do reflect morpheme concatenation order, as can be seen in (10). In this example, tag C, which represents the diminutive morpheme of adjectives, precedes the gender and number tags.

(10) abertinhas aberto AQCFP00

(11) abert-inh-a-s

open-DIM-F-PL

Differently than FreeLing, MorphoBr handles the diminutive of nouns and adjectives uniformly, since there is no morphological difference between the formation of diminutives in these two word classes. In MorphoBr, the order of tags encoding morphosyntactic features of nouns, adjectives, and verbs directly reflect the order of the corresponding morphemes, as exemplified in (12) and (13)¹⁶:

(12) abelhinhas abelha+N+DIM+F+PL

(13) abertinhas aberto+A+DIM+F+PL

(14) comprá-va-mos comprar+V+IMPF+1+PL

buy-IMPF-1PL

One of the main goals of MorphoBr is morphological analysis with FSTs in the context of syntactic parsing. The entry format adopted allows for straightforward conversion to the so called spaced-text format exemplified in (15), where word characters and tags are separated by a blank. This format, in turn, can be compiled into an FST with the proprietary Xerox Finite-State Tools (XFST) (BEESLEY; KARTTUNEN, 2003) or with Foma (HULDEN, 2009), its free-software, open source clone, using the command read spaced-text. In (15), the first line maps to the upper (analysis string or lexical) language of the transducer, while the second line maps to the lower (or surface string) language (BEESLEY; KARTTUNEN, 2003).

(15)

c o m p r a r +V +PRF +1 +SG

c o m p r e i

The format used in MorphoBr also allows for the enrichment of the annotation of verb-clitic clusters such as (2) and (3), since DELAF-PB provides no lemma or morphosyntactic information for clitics in enclitic or mesoclitic position, as evidenced by (4) and (5), where the +PRO tag only specifies that the verb is conjoined with a clitic. Clitic properties, however, are crucial for deep syntactic parsing and semantic analysis in formalisms like LFG, because the clitic realizes an argument of the verb (FALK, 2001). Therefore, in MorphoBr, the +PRO tag from DELAF-PB is substituted by one of our conversion tools for a sequence of tags representing the grammatical properties of the clitic.

Table 2: Comparison between DELAF-PB'S and MorphoBr's annotation of verb-clitic clusters (tags in italics and explanations in quotes).

This conversion is not always a simple string replacement operation, because clitic nos can be ambiguous in relation to lemma depending on the verb form. While this clitic unambiguously represents accusative or dative of nós 'we' in entries like (4) and (5), it can be lemmatized in a two-fold way in entries like like (16), namely, either as nós 'we' or as eles 'they'. Our conversion tool successfully handles all these verb-clitic clusters, assigning the clitic in each case the corresponding lemma and inflectional informations. Accordingly, entry (16) is converted to the two entries (17) and (18), where the lemma ambiguity is resolved: in (17), the lemma of the clitic is ele 'he', while it is nós 'we' in (18).¹⁷

(16) compravam-nos, comprar.V+PRO:I3p

(17) compravam-nos comprar+V.ele.ACC.3.M.PL+IMPF+3+PL

(18) compravam-nos comprar+V.nós.AD.1.PL+IMPF+3+PL

Table 2 shows the correlation between the different components of the annotation of entries (16)-(18). Tags describing properties of clitics are separated by a dot instead of the plus sign used for the other tags. This distinction is relevant for syntactic processing. For example, .PL and +PL both encode plural number, but the former specifies the number of an object of the verb, while the latter specifies the number of the subject.¹⁸

Two aspects of the annotation scheme adopted for verb-clitic clusters in cases like (17) and (18) should be highlighted. First, the relative order of the individual tags describing properties of clitics is somewhat arbitrary, because these elements consist of portmanteau morphemes, conflating different properties in one single morphological unit. For example, clitic nos conflates case, person, and number in one single morpheme. A clitic such as las, however, can be segmented into different gender and number morphemes, motivating the tag sequence F.PL., as exemplified in (19).

The second aspect refers to the the position of clitic tags as a whole in relation to inflectional tags of the verb. Due to direct translation of tag sequence .V+PRO from DELAF-PB in the way specified in Table 2, clitic tags precede verb inflectional tags. In this case, annotation does not mirror morpheme concatenation, since the enclitic pronoun follows the verb. In mesoclitic clusters, however, the clitic precedes the inflectional suffixes of the verb, see (20), so that morpheme concatenation is obeyed. We leave for a future version of MorphoBr a solution to this discrepancy.

(19) compramo-las comprar+V.ele.ACC.3.F.PL+PRF+1+PL

(20) comprar-lhe-emos comprar+V.ele.DAT.3.SG+FUT+1+PL

For the conversion between the different formats, two tools were developed and are made available in the project repository. The first tool is an ad hoc Python script that converts from DELAF-PB's format to MorphoBr's. The second tool adopts a more systematic approach. It performs conversion between DELAF-PB's, GFL's format and MorphoBr's using a declarative approach. However, it can not enrich the annotation of clitics yet.

Using the GF programming language, we wrote a set of grammars that specify how abstract trees (encoding the dictionaries entries data) are mapped to strings in each dictionary's concrete format. GF grammars specify both parsers and printers, so we are able to use them to translate between different formats by parsing a string in a certain format to an abstract tree, and then printing it in another format (with some preprocessing to handle the spacing).

In the example below, we can see a GFL-formatted entry being parsed to an abstract syntax tree in the first output line. The following lines show the linearization of this abstract syntax tree in MorphoBr's format, which in this case corresponds to two entries, one for each gender (forms that admit two genders are collapsed into one "common gender'' in the GFL representation).

(21)

Morpho> import MorphoMBR.gf MorphoFL.gf

Morpho> parse -tr -lang=FL "dentistas dentista N C C P 0 0 0"

| linearize -all -lang=MBR

mkN "dentistas" "dentista" (mkNF Common Pl ZDegree)

dentistas dentista +N +M +PL

dentistas dentista +N +F +PL

Table 3: Examples of incongruences in the encoding of homonymous verb forms in DELAF-PB (incomplete paradigms lack some or all forms).

While converting DELAF-PB's entries to our format, we found numerous errors and inconsistencies in the original data. Just a few problems of this sort were detected when converting GFL's data. All detected problems were corrected either by the format conversion tools or by shell scripts. Due to space limitations, we limit ourselves here to the more significant examples.

DELAF-PB has 318,683 repeated entries. It contains 176 simple formatting errors due to spurious colons in inappropriate places, compare the incorrectly formatted entries in (22a) and (23a) with the corrected counterparts in (22b) and (23b), respectively.

(22) a. abstinhas:-lhe,abster.V+PRO:I2s

b. abstinhas-lhe,abster.V+PRO:I2s

(23) a. mantinhas:,manter.V:I2s

b. mantinhas,manter.V:I2s

There are 3070 verb-clitic clusters without the obligatory separator hyphen, e.g. pruirlhes,pruir.V+PRO:W3s instead of pruir-lhes,pruir.V+PRO:W3s. In 714 entries, the last character is number 1 instead of letter s (singular), e.g. abstrair,abstrair.V:W31. Another problem are 26,151 systematically missing verb forms: 1st conjugation verbs lack 3rd person singular forms of imperfect indicative (all forms missing) and pluperfect (all but 5 missing), while most 2nd conjugation verbs lack the 3rd person singular of future subjunctive (728 forms missing from a total of 800). In all these cases, the missing 3rd person singular forms are identical to the corresponding 1st person singular forms, as exemplified in Table 3.

One could argue that these forms were omitted following a general design decision to spare storage space. However, the evidence counters this assumption, suggesting that the omissions are unexpected side-effects of the lexicon compilation process. In fact, DELAF-PB's documentation does not refer to this lexicon-size reduction strategy. If it were implemented, there should be a special tag for collapsing the person information of the homonymous forms. Instead, the usual tag combination 1s for 1st person singular is used. Moreover, as Table 3 shows, DELAF-PB does explicitly encode both members of similar pairs of homonymous forms, such as 1st and 3rd person singular forms of both imperfect indicative and pluperfect of 2nd conjugation verbs.

Diminutive formation is one of the most productive derivational processes in Portuguese (ROCHA, 2008, p. 122-123). In this language, there are more than 20 diminutive suffixes (CUNHA; CINTRA, 1985). These suffixes underlie lexicalized words from different lexical categories, ranging from nouns and adjectives to adverbs, numerals, personal pronouns, and verbs (RIO-TORTO, 2016, p. 359). They are continually used for the creation of new words. In fact, every noun and every adjective can have a diminutive form (LAPA, 1982, p. 77-82; ROCHA, 2008, p. 123; LIMA, 2011, p. 137). Among the diminutive suffixes, -inh- and -zinh- are the most productive (LAPA, 1982, p. 79; CUNHA; CINTRA, 1985, p. 91-92; RIO-TORTO, 2016, p. 373). This section describes a finite-state implementation of the formation of diminutives from these two suffixes in Portuguese. The goal of this implementation was expanding the coverage of MorphoBr, since the integrated resources have tens of thousands of gaps in this regard.

Denotatively, diminutive suffixes form hyponyms from nouns or modulate the intensity of adjectives, but more often they just convey speaker's emotions or attitudes, e.g. appreciation, dislike, empathy, politeness, etc. (LAPA, 1982; CUNHA; CINTRA, 1985; ROCHA, 2008; BAZENGA, 2012; VILLALVA; SILVESTRE, 2014; RIO-TORTO, 2016), being very common in emotive discourse, opinion or persuasive texts, etc. Therefore, analyzing diminutives should be a basic capability of NLP systems targeted at sentiment analysis, opinion mining, text classification, etc.

Out of the total of 361,485 nouns and adjectives converted from DELAF-PB and GFL (henceforth DGFL-ADJN), only 15,938 are diminutives, 10,338 of which are formed with -(z)inh-. Consequently, there are many gaps in these resources, because, for tens of thousands of words, they do not provide the corresponding diminutive. In DELAF-PB, for example, there are diminutives for cobra 'snake', jacaré 'alligator', zebra 'zebra', gavião 'hawk', and cheiro 'smell', but not for elefante 'elephant', javali 'boar', jumento 'donkey', amor 'love', odor 'smell', and dor 'pain'. Another deficiency of DELAF-PB and GFL is the lack of -zinh- for corresponding -inh- diminutives. For example, both resources include cobrinha (diminutive of cobra 'snake'), but not the equally grammatical parallel form cobrazinha. All these gaps seem completely arbitrary. In fact, the corresponding diminutives can easily be found in texts on the Internet, e.g.:

(24) cobrazinha, amorzinho, dorzinha, elefantinho, elefantezinho, jumentinho, jumentozinho

To fill these gaps, we took the standard assumption that the lexicon of a natural language consists not only of existent words, but also includes potential words, i.e. words that can be created by applying word-formation rules to existent words (ROCHA, 2008; VILLALVA; SILVESTRE, 2014). Finite-state morphology is the standard paradigm for the construction of rule-based computational models of inflectional and word-formation processes. This approach has two strengths. First, morphological processes can be formalized in a way that closely mirrors linguistic descriptions. Thus, one does not have to reinvent the wheel when implementing a certain morphological phenomenon already described in detail in the linguistic literature. All one needs to do is to translate the description from a natural language into a formal specification. Second, this formal specification can be compiled into an FST using free, open source software, e.g. Foma (HULDEN, 2009). The resulting FST, in turn, can be used in a compact and efficient way during text processing.

Our implementation of diminutive formation with -inh- and -zinh- generally follows the analysis by Villalva and Silvestre (2014) and Rio-Torto (2016), according to which there are two types of diminutive suffixes in Portuguese: evaluative suffixes (-inh-, -it-, etc.) and z-evaluative suffixes (-zinh-, -zit-, etc).

As Villalva and Silvestre (2014, p. 119-120) points out, there is some variation in the distribution of these two types of suffixes across different dialects of Portuguese. A typical case of regional divergence are -inh- diminutives like anelinho (from anel 'ring'), derived from stems ending in l in singular (with a null thematic marker surfacing as e in plural), which are restricted to European Portuguese. Besides, there are preference differences between speakers or depending on word length or frequency. Diminutives from longer, less frequent words like parlamento 'parliament' are preferably constructed with z-evaluative suffixes. Accordingly, while both parlamentozinho and parlamentinho are grammatical, the former is considered more acceptable.

Abstracting away from these factors, the generalization holds that evaluative suffixes are restricted to stems of words ending with one of the thematic unstressed vowels -o, -a, and -e (e.g. cheiro, zebra, and elefante), while z-evaluative suffixes, as exemplified in Table 4, attach to inflected words. Therefore, for the first group of words, both types of suffixes are licensed (cf. cobrinha and cobrazinha from cobra ['kɔbrɐ] 'snake'), while all other words only license z-evaluative suffixes (compare cafezinho and *cafeinho from café [kɐ'fɛ] 'foot' or motorzinho and *motorinho from motor 'motor').

Table 4: Examples of diminutives with -zinh- derived from plural forms.

Apparent exceptions to this generalization are due to adjustments in the orthographic shape of the concatenated elements, due to general orthographic or phonological constraints in the language, e.g. lapisinho from lápis 'pencil' can be explained by deletion of z from -zinh-. Other exceptions are lexicalized words like colherinha (diminutive of colher 'spoon'). In deviation from the productive pattern in the standard language, the -inh- suffix in this example applies to a base ending in a phoneme other than one of the thematic vowels -o, -a, and -e. Examples like *reporterinho (from repórter 'reporter') evidence that this pattern is not productive.

In the two-level approach in finite-state morphology, morphological regularities are factorized into two modules (BEESLEY; KARTTUNEN, 2003). The morphotactics component describes the possible combinations of morphemes. In the second module, morphographemic alternations handle allomorphy, i.e. changes in the orthographical form of morphemes when combined to build new words or word forms (ŠEVČÍKOVÁ, 2018)¹⁹. While some alternations reflect phonological changes affecting pronunciation, as in lapisinho referred to above, other alternations are purely orthographical, as exemplified in Table 5.

Table 5: Examples of orthographical changes not affecting pronunciation.

Following the two-level approach, we modeled the morphotactical phenomena in the formation of -(z)inh- diminutives by means of a grammar in the LEXC formalism. This a high-level declarative language that enables the specification of finite-state automata and transducers in a linguistically intuitive way (BEESLEY; KARTTUNEN, 2003). Both XFST and Foma, its free software, open source counterpart, provide compilers for LEXC.

A LEXC grammar specifies different classes of elements, modeled as LEXICONS, and the possible combinations between them, as in the oversimplified example in (25). This example generates the -(z)inh- diminutive forms of alegre 'happy'. The first line declares the symbols that constitute multicharacter arc labels in the FST the grammar compiles into. The following lines specify five LEXICONS, each consisting of one or more entries. The first element in each entry is a pair of the form x:y, where x is a fragment of the analysis string and y is a fragment of the surface string. Typically, this pair represents a morpheme, i.e. an atomic unit of meaning and form. The identity relation x:x can be encoded as x. The empty string is represented by 0. The second element in an entry is a continuation class, which defines the LEXICON the x:y pair can be concatenated with. The number sign "#" represents the end of a word. This grammar compiles into an FST relating analysis strings like alegre+A+DIM+F+PL to their corresponding surface forms, alegre^inh^a^s and alegre^zinh^a^s in this case, where the caret "^" represents a morpheme boundary. These are intermediate forms which are converted to the actual forms alegrinhas and alegrezinhas by conditional replacement rules in the module encoding morphographemic alternations, see Beesley and Karttunen (2003) for a detailed explanation.

(25)

Multichar_Symbols +A +DIM +M +F +SG +PL

LEXICON Root

alegre Lemma ;

LEXICON Lemma

+A:0 Suffix ;

LEXICON Suffix

+DIM:^inh Gender ;

+DIM:^zinh Gender ;

LEXICON Gender

+M:^o Number ;

+F:^a Number ;

LEXICON Number

+SG:0 # ;

+PL:^s # ;

Long-distance dependencies between elements of non-adjacent classes can be elegantly modeled by means of unification-based constraints named flag diacritics (BEESLEY; KARTTUNEN, 2003). These constraints function as filters at runtime during parsing and generation, blocking ungrammatical paths. Obtaining the same effects in a LEXC grammar without resorting to flag diacritics makes the code less intuitive. Alternatively, as Beesley and Karttunen (2003, p. 339) points out, excluding ungrammatical paths by composition of transducers can make the size of the resulting transducer explode.

The first class in our grammar contains 334,284 noun and adjective entries from the total of 361,485 in DGFL-ADJN. Seemingly improbable bases were filtered out, e. g. diminutives (cf. *lapisinhozinho) and superlatives (cf. *rapidissimozinho from rapidíssimo 'very fast'). Augmentatives, however, were included, since diminutives derived from these formations are considered grammatical (RIO-TORTO, 2016, p. 364) and attested in texts on the Internet (e.g. casarãozinho from casarão, augmentative of casa 'house'). The other classes of the LEXC grammar contain the diminutive suffixes, the gender morphemes, and the number morphemes.

Diminutive formation in Portuguese involves a dependency relation between the gender of the base and the gender marker, as exemplified in (26)-(29). This dependency is non-local because the diminutive suffix intervenes between these two elements.

(26) a. problem-inh-a

problem(M)-DIM-M.SG

b. problem-a-zinh-o

problem-M.SG-DIM-M.SG

(27) a. trib-inh-o

tribe(F)-DIM-F.SG

b. trib-o-zinh-a

tribe-F.SG-DIM-F.SG

(28) a. pont-inh-a

bridge(F)-DIM-F.SG

b. pont-e-zinh-a

bridge-F.SG-DIM-F.SG

(29) a. dent-inh-o

tooth(M)-DIM-M.SG

b. dent-e-zinh-o

tooth-M.SG-DIM-M.SG

Examples (26)-(29) show that gender marking in Portuguese diminutives constitute a complex phenomenon, because masculine gender can also be marked by a and feminine gender by o, which are the canonical markers for feminine gender and masculine gender, respectively, compare (26a) and (27a) with (26b), (27b), (28), and (29). Another difficulty for the formalization is the opposing behaviour of -inh- and -zinh- in cases such as (26) and (27): while the former selects the non-canonical markers, the latter selects the canonical markers. In our LEXC grammar, these facts are handled by means of flag diacritics in a linguistically intuitive way.

Morphographemic alternations were modeled by a cascade of 11 conditional replacement rules. These rules follow the general template A --> B C _D, which informally reads "substitute A for B in the context of C and D", where C is the left-hand and D the right-hand context. This module includes, besides the orthographical changes from Table 5, plural -s deletion before z (azuis+zinhos => azuizinhos, plural diminutive of azul 'blue'), thematic vowel deletion (casa+inha => casinha, diminutive of casa 'house'), optional thematic e deletion in plurals like luzezinhas and florzinhas (producing luzinhas and florzinhas) and stem i deletion before another i (e.g. cheiinho=>cheinho and saiinha=>sainha, derived from cheio 'full' and saia 'skirt', respectively).

The current implementation is biased towards contemporary Brazilian Portuguese, where the forms generated by the last two rules are attested in standard language texts and considered grammatical²⁰. Another consequence of the present limitation is that the formation of diminutives like anelinho, which is only productive in European Portuguese, was not implemented yet.

Both the LEXC grammar and the morphographemic alternations were compiled into FSTs, which, in turn, were composed into one single FST, which we call DIM1²¹. This FST has almost 2 millions paths, but just a fraction is licensed by the unification constraints that operate during analysis or generation. In fact, extracting the grammatical word-parse pairs from DIM1 reveals that their number amounts to 625,716 pairs (see Table 6). In order to reduce complexity, as proposed by (ALENCAR et al., 2014), these word-parse pairs where compiled into a second FST using the read spaced-text command referred to in section 3. This derived FST we designate by DIM2. Among other reasons, DIM2 is less complex than DIM1 because there are no unification constraints to be solved.

Table 6: Comparison of the complexity of different FSTs compiled with Foma²². The second column specifies space in disk of the word-parse pairs and the third, FST size in memory.

Table 6 compares the complexity of DIM2 to two other FSTs, which we label ADJN and ALL. ADJN was compiled by applying the read spaced-text command to all word-parse pairs from DGFL-ADJN (i.e. the set of all nouns and adjectives that were converted from DELAF-PB and GFL). ALL is the FST resulting from union of DIM2 and ADJN. The last line of Table 6 allows us to assess coverage gain as well as the cost of uniting DIM2 with ADJN. The numbers in brackets show the percentage increases in relation to the numbers in the penultimate line. ALL has 170,3% more paths than ADJN, but the complexity cost in terms of FST size and number of arcs and states ranges from 16% to 28,6%. On the other hand, the word-parse pairs from ALL occupy 220% more space in disk than those from ADJN.

In order to assess coverage and accuracy of our resource, experimental evaluation was carried out in two different phases of the project development. In the first phase, evaluation was restricted to MorphoBr's entries from DELAF-PB and GFL. Three experiments were performed in this stage. First, these entries were used to lemmatize the Universal Dependencies (henceforth UD) Portuguese GSD corpus. In this way, we could have a measure of the resource's coverage on real-world data. In the next experiment, an accuracy test was carried on Bosque (another Portuguese UD corpus), by comparing the lemmas MorphoBr assigned to the words in this corpus to the ones Bosque actually had. In the third experiment, we measured FreeLing's coverage of verb clitic-clusters by means of suffix rules.

In the second project development phase, evaluation was restricted to the diminutive entries created by the finite-state approach and encoded in DIM2. Coverage comparisons were performed against the two previous relevant resources: (i) the set of -(z)inh- diminutives from DGFL-ADJN, i.e. the set of all nouns and adjectives converted from DELAF-PB and GFL, and (ii) FreeLing's suffix rules.

We have presented MorphoBr, a new wide-coverage full-form lexicon for Portuguese, released under a free, open source software license. It represents a two-fold contribution. First, previous freely-available resources were consolidated, removing several thousands of errors and gaps. Entries were converted to a uniform format using more mnemonic and linguistics-oriented tagging conventions. This format not only is more human-readable but also allows for straightforward compilation of finite-state morphological analysers.

MorphoBr, however, is not just a combination and correction of previous resources. Its main contribution is the systematic treatment of word-formation by computationally modeling the underlying linguistic regularities. As a test case of this approach, the formation of -(z)inh- diminutives was implemented in the paradigm of finite-state morphology. Previous resources either provide very incomplete lists of diminutives or formulate ad hoc rules that cover only a small part of the cases. By contrast, our systematic treatment of diminutive formation rules resulted in 170% more nouns and adjectives than DELAF-PB's and FreeLing's dictionaries combined. As regards the total amount of -(z)inh- diminutives, the finite-state implementation generated 60 times more such items than listed in these previous resources.

MorphoBr is still work in progress, but the experimental evaluation results seem promising. It clearly outperformed FreeLing's suffix rules in the coverage of -(z)inh- diminutives. This makes it more adequate for tasks dealing with texts where diminutives are very common

Regarding the FreeLing coverage test, we are aware of the fact that many missing items could be avoided with improvements in FreeLing's affixation rules for Portuguese. Nevertheless, we leave as a future work testing whether this type of rules can deal as efficiently with all forms derived by our FST approach, which involves the formalization of intricate phenomena both at the morphotactic and morphographemic levels.

Other future related work includes: expanding our FSTs with other productive word-formation rules, also taking into account particularities of European Portuguese; reviewing the verb-clitic clusters and their annotation; dealing with MWEs; and implementing the grammatical word classes, e.g. determinants, conjunctions, etc.

The latter topic requires, however, a clarification about the lexical representations of these words, which, in turn, depends on the implementation of syntactic rules in a concrete grammatical formalism, GF and LFG in the case of our project. This means that the exact form of these entries can only be defined after developing grammar fragments in these formalisms covering the relevant grammatical phenomena.

We are indebted to Fabricio Chalub for his contributions to a previous version of Section 5.4. We also thank an anonymous reviewer for valuable comments and suggestions.