Prefix allomorphy in complex verbs of Brazilian Portuguese Alomorfia prefixal em verbos complexos do português brasileiro

This paper investigates the empirical properties of morphologically complex verbs of change in Brazilian Portuguese. Its main goal is to find out whether there are (not) correlations between prefix morphophonological form, root semantics and verb argument structure. All verbs analyzed denote events in which the internal argument is the undergoer of a change, and a classification in four semantic subclasses according to their root meaning is proposed: state, place, transfer of possession (either abstract or concrete) and reconfiguration. Contrary to what previous literature has suggested, finegrained investigation within each of these classes has shown that it is not possible to exclusively associate morphological forms to specific root semantics or morphosyntactic behavior, although it is possible to observe general tendencies. To provide an analysis to account for the robust homogenous structural properties and general semantics behavior across the class, in face of the morphological variation in what regards prefix choice among the forms a-, enand es-, it is proposed that prefixes are the phonological realizations of a functional head in the lower domain of the verb (labeled R, Relational) whose choice is given in terms of contextual allomorphy. It is shown that vocabulary insertion at R is guided by locality with the root. The theory of allomorphy proposed in Embick (2010), which is based on locality and linearity, was efficient in accounting for the selection of allomorphs of R, v and Th (Theme Vowel) heads.


Submitted on September 9 th , 2018
Accepted on November 26 th , 2018

Introduction
Verb formation in Brazilian Portuguese (hereafter, BP) is a process that may involve phonologically realized morphemes: in its morphological form it is possible to recognize a prefix, a base (root, noun or adjective), and an optionally overt suffix, to a lesser degree, and a theme vowel, besides of the person/number and tense/mood affixes in the inflectional domain. In what regards argument structure, verbs containing these prefixes always require a complement, being: i) unaccusative verbs, that can possibly be provided with external arguments, 1 or ii) transitive verbs, but they will never behave as unergative verbs: However, the assumption that the verb general behavior can be derived solely from intrinsic properties of these prefixes does not hold upon closer inspection. Detailed investigation will make clear that the behavior of these affixes is quite complex, challenging superficial classifications, especially the ones that assume they carry specific semantic properties, such as directionality. We hope to make clear the fact that the final behavior of the verb must be derived structurally. Thus, the descriptive questions to be answered in this paper are the following: i.
What is the contribution (if any) of these prefixes to the final verb?
ii. What is the base contribution to the final verb?
iii. When (and if) prefixes are allomorphs, what guides the choice of their phonological form?
iv. Is there any morphophonological dependency between prefix and suffix forms when they co-occur in a verb?
The paper is organized as follows: in the remaining of section 1 we briefly present the methodology of data selection. In section 2, there is a general proposal for the treatment of the lower structure of all verbs of change and the empirical evidences for the suggested analysis. In section 3, verb classes are fully investigated in what regards their root semantics and in special the morphophonology of prefixes. Section 4 provides a morphophonological analysis for the distribution of prefixes based on locality conditions. Section 5 concludes the paper.

Methods
The descriptive results of this research are based on an initial corpus of 380 verbs. The methodology for collecting these data followed several steps, summarized below: a. Initial Selection of verbs started by a-, ad-, en-, e-, em-, in-, es-, ex-from the 2007 edition of the digital database of Dicionário Houaiss da Língua Portuguesa (HOUAISS;VILLAR, 2007), resulting in 5.173 occurrences; b. Frequency extraction: verbs were ranked according to their Google frequency (in 2010) and only those with more than 10.000 hits were conducted to the next step. The frequency was automatically extracted and the Language filter for Brazilian Portuguese was used. This lead to 1.471 occurrences; c. Data cleaning: was composed of the following substeps, and lead to a total of 380 verbs: • Exclusion of verbs for which these sound sequences were mere initial parts of roots or prefixes denoting negation (in-) (e.g. amar 'to love', alterar 'to alter', entrar 'enter', invalidar 'to invalidate'); • Exclusion of adjectives ending in -ar (e.g. escolar 'academic'); • Exclusion of verbs that, even with a number of frequency greater than 10.000, seemed extremely old fashioned or were completely unfamiliar 4 (e.g. aquinhoar 'to distribute', arregimentar 'to regiment').
A first major classification was proposed in order to separate one group of semantically compositional and morphologically transparent verbs from two other groups: morphologically opaque and semantically non-compositional verbs, 5 leading to 276 verbs from which 259 are verbs 4 The exclusion of the following verbs was based on my intuition as a researcher and native speaker of BP: amuar,amontar,anuviar,apainelar,apiedar,apoucar,aprovisionar,aquinhoar,arregimentar,arregalar,acantoar,embeiçar,encabar,abancar,encarniçar,atucanar,emparceirar,encapelar,abaratar,arruar,avultar,atraiçoar,arrimar,entroncar,encordoar,abaciar,encampar,aventar,acautelar,arruar,enjeitar. 5 We refer the reader to Bassani (2015) for this first classification. of change and result. In this paper, we will explore these 259 data. A list of these verbs organized by frequency is available in the appendix.

General description and proposal
The major class of semantically compositional and morphologically transparent verbs treated in this paper can be classified as Change or Result verbs LEVIN, 1998LEVIN, , 2010ALEXIADOU, 2014). These verbs denote an event in which the internal argument is the undergoer of a change, and the result of this change is somehow related to the root semantics. Although all these verbs can ultimately be classified as verbs of change, the idiosyncratic meaning of the root specifies the kind of change affecting the internal argument: some roots denote/are related to states, places, entities and by-products of the event. Then, structurally, the unifying characteristic of these verb subclasses is the obligatoriness of an internal argument, which is interpreted as the affected object of the change expressed by the event (theme or experiencer, to a lesser degree). In relation to the morphophonological properties of the data, it is evident that most of the verbs present overt prefixes (a-, en-, es-) but most do not present overt suffixes (when they appear, the forms are -iz and -ec). There are also verbs of change and result that do not present affixes. Below we propose a general syntactic lower structure for this class and the empirical evidences for it, and in the next section we describe four (root) semantic subclasses to be examined in detail in what regards their argument structure, lexical semantics and the morphophonological form of affixes.
We propose that morphologically complex verbs of change present morphological evidence for a syntactic decomposition of the event (along the lines of Marantz (1997); Alexiadou et al. (2006)). In broad terms, prefixes can be analyzed as the realization of a functional head in the result state clause and suffixes are the realization of v, which take that result phrase as complement. We propose that there is a formal requirement in Portuguese in order for the internal argument to enter the derivation in a position where it is able to be interpreted as the affected argument of the change: for the argument to relate to the root semantics there must be the mediation of a functional head that is prepositional/ relational in nature. Then, we assume that the prefix is the phonological realization of a head that introduces the internal argument in the structure and relates it to the root, as in the representation in (5) below. Our analysis go against approaches in which the root can directly take a complement, such as Marantz (1997), Embick (2004), Harley (2014), since the evidence to be presented shows that the root fails to take a complement without the mediation of a functional category. 6 We assign the label R for Relational head, and IA stands for Internal Argument. In the structure in (5), prefixes are the phonological realization of an inner and active functional head attaching outside the Root domain, but not above the first Categorizer/Cyclic Head (EMBICK; EMBICK, 2010), which is v. In what regards R status, we assume, based on empirical evidence to be presented below, that it is a kind of predicational head, which is part of the verbal domain, however it is a non-cyclic head and is therefore a subword level morpheme. It can be phonologically realized as a-, en-or es-. It is in fact the category v that is responsible for the categorization of the Root and the formation of a M-word (EMBICK; NOYER, 2006). The head v in these verbs can be phonologically realized by -e-, -ec-[es] and -iz-or can be null.

Empirical evidences for the proposed structure
Following several empirical evidences presented in Bassani (2012), we assume that R is a functional head in the verbal domain attached to the root and complement of v. First, if we compare the bases to which these prefixes attach to the verbs they form, we find that they seem to be interfering with, or, in fact, being responsible for, the introduction of an extra argument. For example, a noun like garrafa 'bottle' in (6), an adjective like vazio 'empty' (7) cannot take arguments by themselves. Although it is clear that an adjective like vazio is predicative and needs an entity to be composed with, it cannot do this by itself in all languages, as Hale and Keyser (2002) have already observed. In addition, the examples in b. below show that it is impossible to derive a verbal structure for these Roots without the phonological realization of a prefix.
'to bottle the wine'.
'to empty the box'.
c. Esvaziar (v) a caixa. 'to empty the box' Finally, their presence is restricted to verbs: they only occur with nouns and adjectives in case they are derived from verbs, like in derived eventive nominalizations with suffixes -ção and -mento (i), to a lesser degree in back derivations (ii), and widely in resultative participles (iii): 'to take advantage' c. *Aproveito (n) 'advantage' d. Aproveitamento (n) 'utilization' iii. Resultative Participles The ungrammaticality of the data in c. from (8) through (15) above and the exclusive occurrence within verb domains is also evidence that the prefixes are not able to categorize a structure alone, and this is due to the fact that they are non-cyclic heads depending on v. In addition, these prefixes can co-occur with (open) verbal categorizers ((16) to (18)), which shows that they cannot be the phonological realization of v, as was suggested by Oltra-Massuet (2000) for similar Catalan verbs like enrikir and aprofundir: (16) a-terror-iz-a-r pref-√terror-suf-tv-inf 'to terrorize' (17) en-riqu-ec-e-r pref-√weak-suf-tv-inf 'to enrich' (18) es-faqu-e-a-r pref-√knife-suf-tv-inf 'to stab' Having presented evidence for the structure of the lower part of this general verb class, we turn now to specific details of subclasses according to the root semantics and the morphophonological form of the affixes. Our challenge is to account for its consistency in relation to the general structural properties and semantics of change but its variety when it comes to the choice of morphemes.

Verb Classes and root types
In what follows we investigate possible correlations between the pieces of morphology, root semantics and argument structure.

Root Semantics and argument structure
These verbs denote a change of state undergone by the internal argument, which can be a physical state (alisar X 'to smoothen X') or a psychological state (as in acalmar X 'calm down X'). From (19) to Closely observing this class, it is clear that the (supposed) categorical status of the base is not relevant to the formation of a change of state verb: both nouns and adjectives could be part of these verbs, provided that they or their root can denote state properties/characteristics acquired by the internal argument as a result of the event of change. Thus, the split into deadjectival and denominal verbs in this class is totally unnecessary. Formally, there are two kinds of bases to derive the verbs: bare roots and derived adjectives and nouns denoting states. A simple entailment test can split the two kinds of bases: while the result state from verbs formed from bare roots can be expressed by simple adjectives (often scale adjectives) (22), the result state from verbs with derived nouns can only be expressed by adjectival participles or prepositional phrases (23): Although adjectival participles can be also used as entailments in change of state verbs with stative bare roots (ficar adoçado / enfraquecido / esfriado 'to get sweetened/weakened/cooled down'), it reveals, in this case, the outcome of the event (resulting state) as a whole and not the result of the change in the innermost layer of the inchoate subevent (one in the RP projection). Moreover, the interpretation of the auxiliary ficar ('to get') is closer to the meaning of changing and remaining in the state, more than just denoting the punctual change of state. So the most natural paraphrase for this type is indeed the one made with the simple adjective.
Derived state data leads to an important theoretical discussion, the one of structural licensing. It seems clear that it is not the state root that projects a change of state verb (or adjectives, as in many lexicalist proposals, including Hale and Keyser (2002) and even in Distributed Morphology approaches (such as HARLEY, 2014)), but this type of root is licensed in the event structure if it is compatible with it (cf. MARANTZ, 2013). An extra piece of evidence that corroborates this fact was found among the data: roots that prototypically denote entities, with no possibility of lexical projection, or "assignment of thematic roles" can nevertheless form change of state verbs as long as it has some specific property that can be highlighted as a proper licensing. For example, for the root that forms pedra ('stone') to become part of the verb empedrar ('to pave'), what is relevant is its property of being solid, hard, and not any other characteristics of the noun pedra ('stone').
What to do when the milk pref-√stone-prs.3sg 'what to do when the milk becomes like stone'? 7 so whenever a root or an XP (X Phrase) denoting a state is licensed, (by intrinsic properties or coercion, as in the example above), as the complement of an R that introduces an internal argument, the whole structure will be interpreted as a change of state structure. As an illustration, we present in (25) the analysis for the lower part of a change of state monoargumental structure resulting in the formation of the verb amadurecer ('to ripen'): (25) Amadurecer a fruta (to ripen the fruit).
In this structure, the head R allows the introduction of an internal argument and relates it to the Root. Then a small clause denoting a result state is formed, which in combination with a v head generates the interpretation of an event of change and result.

Morphology
The first question to be answered regarding the morphologysemantics relation is whether there is any correlation between the form of the prefix and this subclass. Acedo-Matellán (p.c.) points out that, in Catalan, the prefix en-/em-conveys the meaning of change of place while the prefix a-conveys change of state. 8 For BP such a division of labor between a-and en-is not empirically attested. In BP, the prefix en-has wide occurrence within the class of change of state verbs: 53 out of 136, co-occurring or not with suffixes. The prefix a-is still the most numerous, with 75 occurrences. The prefix es-shows up in 8 verbs only. Regarding the morphology of suffixes, the most evident general characteristic in this subclass is the exclusive presence of the suffix -ec (cf. (17)). The suffix -iz is also present: of the three occurrences in the whole corpus, two belong to this class. However, the suffix -iz occurs very rarely concomitantly with the presence of prefixes, a defining feature of our corpus. So, the most certain conclusion is that the presence of -ec suffix is correlated with this subclass. However, two observations are in order. First, -ec must be the phonological realization of a verbal categorizer since it is in complementary distribution with other verb suffixes. Second, such a correlation is not two-way: it is not true that in order to derive a change of state verb the suffix -ec must be present. Thus, we must conclude that -ec could be one of the phonological realizations of v. Within this class, the suffixes are distributed as follows (table 3): 136 verbs, 39 with -ec, two with -iz, two with -e and most of the verbs do not have overt suffixes (93 verbs). 9 The next question to be answered is whether there is such a strong correlation between suffixes and prefixes in these verbs that makes them circumfixes rather than individual affixes, as assumed by Rio-Torto (2004) and Pereira (2007). According to this analysis, the realization of the prefix en-in a change of state verb would be the result of its dependence on the suffix -ec. Again, our data show that this is not empirically attested: -ec suffix co-occurs with the three prefix forms, as shown in table 4 below. Although it does co-occur more frequently with en-, we must address this fact as a numerical trend and not as empirical evidence for the postulation of a circumfix. Moreover, this would be the only case of circumfixation in the whole language system. A critical fact in this subclass refers to the total lack of semantics of directionality that is supposedly related to the presence of a-, en-and es-prefixes. More importantly, as this is the numerically larger class among the verb classes with the presence of prefixes, the classification of such morphemes as directional elements (see table 1) seems highly problematic. By observing the behavior of the next (minor) classes of verbs, we can return to this topic and draw more robust conclusions based on the data. In order to do this, we turn now to the description of general patterns of Change of Location verbs.

Root Semantics and argument structure
Another pattern found among the general class of verbs of change is what we call verbs of change of location. In sum, they result from structures where there is a change of the internal argument for the place denoted by the root. Some examples are provided below: 'to send out of country' Three facts discriminate this subclass from the previous: the root denotes a place rather than a state, there is an apparent directional semantic contribution of the prefix, and they present distinct behavior in face of syntactic tests that reveal inchoative events: they are transitive and do not present alternation patterns. This happens because such verbs denote a complex event in which the external argument must cause the transfer of what is denoted by the internal argument towards what is denoted by the root. In other words, they cannot configure spontaneous internally caused events like some change of state verbs can (e.g. amadurecer 'to ripen').

Morphology
As just said, one important factor that potentially contrasts change of state and change of place verbs is a possible directional contribution of the prefixes in the latter class. Pereira (2007) states that in European Portuguese, depending on the denotation of the base, the change of location may be of the type i) 'get close to Noun' and is realized by a-(as in aterrar 'land', alunar 'land on the moon'); ii) 'get in(side) Noun' and is realized by en-(as in engarrafar 'to bottle', ensacar 'to bag') or iii) 'take of Noun' and is realized by es-/ex-(as in espipar 'extract', expatriar 'expatriate'). Looking closely at the BP data, the situation seems more complex, especially when we consider a-and en-.
In order to situate the discussion, we present the following summarizing tables, showing that the most numerous form in this class is the prefix en-, being present in 25 out of 36 verbs, followed by awith 9 occurrences. The fact that the form es-in only present in 2 verbs with some directional meaning, added to the fact that this piece has no independent status as preposition in the language (different from the other two) and that it is present in most verbs considered as reanalyzed as simple forms by Bassani (2015), lead us to concluded that its status as a vocabulary item in the BP Vocabulary inventory is not stable. First, in descriptive terms, there is a semantic correlation between the root and the prefix. In general, a-combines with roots that denote places of the type space and en-combines with places of the type container. We observed that the prefix en-has a more homogeneous and specific behavior; its occurrence is associated with roots that denote container places, and its often interpreted as a directional piece meaning inside. What seems is that a-is associated with an underspecified semantics of directionality, negotiating with the roots its final interpretation. In this sense, formations with a-seem to have semantically lighter directionality than en-.
A key question that arises now is whether, in fact, the prefixes originally carry specific directional semantic features (we can hypothesize ['inside'] for en-, ['outside'] for es-, ['near'] for a-) and are licensed in compatibility with the roots or have their directional features completely underspecified and its interpretation is contextually given when they are concatenated with the roots.
As we saw in the previous subsection, directional meanings are completely absent when the roots / XPs complements of R denote states, indicating that the locality of attachment and the denotation of the root play a fundamental role in the interpretation of (non)directionality in the result phrase. Before deciding on what is the best analysis for the features of R and its vocabulary insertion process, we shall examine two other semantic classes found in the data.

Root semantics and argument structure
The third type does not express neither a change of state or a change of the location of the internal argument. In this verb class there is a change of possession or localization of an entity denoted by the root that targets the object denoted by the internal argument. Some examples of prefix verbs from BP belonging to this class are provided below. It can be seen that the entity denoted by the root can be a concrete, as in (29) and (31), or an abstract one, as in (30) and (32).
(32) en-feitiç-a-r pref-√spell-tv-inf 'to bewitch' Cançado and Godoy (2012) classify these verbs as verbs of change of possession and divide them into two types, according to the denotation of the root: when there are roots that denote concrete objects, such as manteiga and carpete ('butter' and 'carpet') verbs are called Locatum verbs, and for roots that denote abstract entities to be transferred, such as benção e conselho ('blessing' and 'advice'), verbs are called Benefactive verbs. However, as verbs like amaldiçoar ('to curse'), enfeitiçar ('to bewitch'), castigar ('to punish'), with malefactive internal arguments rather than benefactives, fit in the classification of this last subclass, we prefer the terms change of concrete and abstract possession verbs.
Comparing change of concrete possession and change of place verbs, we observe that the transfer of the entity denoted by the root to the internal argument does not imply the expression of a specific directionality (inside, near, out). In change of concrete possession structures, the head R appears to mediate a directional abstract relation between the root and the internal argument since there is a physical transfer. It seems incorrect to assume that directional features of the type ['inside'] or ['out'] are marked for some distinction within this class. Going further, if we compare change of place verbs and change of possession verbs on one hand and change of state verbs in another hand, we see that in the latter case, not even an unmarked directionality is found. Again, in this last class, the idea of transfer and direction is completely absent.

Morphology
Below we can see the tables regarding prefixes in change of concrete and abstract possession. As happens with verbs of change of place, in the first subgroup there is a predominance of occurrence of en-prefix, followed by a-. What we have observed so far with respect to morphology is that the lower the expression of directionality (what implies that there is no physical transfer and movement), the greater the occurrence of a-. In change of state verbs, for example, a-occurs more frequently (76 out of 139 cases). Change of possession data seems to support this idea. In cases of concrete change, where the entities denoted by the root have physical nature, the transfer implies some directionality (but a not specific one), but when the entity to be transferred has no physical nature, there is no need for expression of directionality and most of them contain the prefix a-, as we see in the following table: We now turn to the last subgroup identified among the data.

Root semantics and argument structure
We name Change of Configuration Verbs the last subclass proposed in our study. A verb belongs to this class if its root denotes an entity that is created in the event. Additionally, these entities must be created from the reorganization or reconfiguration of the material denoted by the internal argument. In (34), for example, the entity denoted by the root √pile is a product of the reconfiguration of the material in the internal argument, the books.

Morphology
In this subclass, 15 verbs have prefix en-, 6 verbs have prefix aand 6 verbs are formed by prefix -es. We first notice that the presence of -es more representative here than in the other three previous subclasses. The presence of the prefix es-appears to be related to a semantic compatibility with the type of entity denoted by the root. Everytime es-is present in this type of verb, the reconfiguration of the internal argument implies a division in parts of an entity (36) or multiple entities (37), which is not expressed by verbs with en-and a-, where only one entity (whole, of concrete or abstract nature) is created (ex. list, group, pile). Some studies consider -es as an aspectual prefix of iterative value in these verbs. Again, we believe that the idea of producing parts may be correlated to the presence of this piece, but it is not exclusive or inherent in it, since there are verbs which behave exactly like the verbs above, such as fatiar ('to slice'), and do not have affixes. On the other hand and in the same line of arguments, there are verbs with -es without the expression of iterative semantics (e.g. esverdear 'to turn green' and estripar 'to disembowel'). Once more we argue for the final semantics as a result of the root semantics and the locality of attachment of the prefix with it and not as a property of the morphological piece by itself.

Deriving the morphophonological form
In the previous section, we have offered a description of prefixes occurrence in transparent and compositional complex verbs, and we also have suggested a general analysis of its lower structure. In Distributed Morphology, each affix is potentially the phonological realization of a head. In technical terms, an affix form is potentially a vocabulary item inserted in a terminal node, which is composed of (a set of) features. In this section, we will investigate the principles that guide the specification and manipulation of features and the insertion of vocabulary items at the terminal node R mainly. As this paper focuses on the prefix realization, we will not discuss in detail vocabulary insertion at v and Th (theme vowel) heads, we leave it aside for future work. The analysis to be outlined below relies on the localist theory of allomorphy outlined in Embick (2010), which presents a version of Distributed Morphology theory called C1-LIN. This theory proposes that contextual allomorphy patterns are constrained by notions of locality based on the ideas of linearity (linear adjacency) and phase-cyclicality. It is a serialist view of morphological and syntactic derivation, where each "step" of the derivation serves as input to the next and therefore can influence it, since they are in the same domain. Contextual allomorphy (or lexically conditioned allomorphy) for Distributed Morphology is determined during the process of vocabulary insertion, i.e., when terminal nodes receive phonological content. Compatible vocabulary items compete for insertion at a terminal node and the most specific is inserted. The allomorphy determined by this type of operation is always a form of suppletion, since it is not possible to derive one form from the other only by phonological operations, for example. Furthermore, the theory assumes a syntactic derivation by phases (CHOMSKY, 2001) with cyclic spell-outs, which restrict the amount of information that is available in a given cycle of the derivation and the possible allomorphic interactions. Objects can interact only if they are active in the same computational cycle. The heads that trigger spell out are phase heads within the word, i.e., categorizers, always represented by lowercase letters.
The notion that categorizers define cyclic phases is very important for the analysis we will outline to the verbs in study. As pointed out in , and also in Embick and Marantz (2008), there is a difference in the expected results in terms of semantic and phonological terms depending on the place of attachment of the affix and the head type. The first syntactic (categorial) heads that merge with bare roots define categories of words. The first categorizer head is in the inner domain of word formation, domain that seems to allow a special relation in terms of sound and meaning. The main generalizations that follow from the cyclic part of the theory are: Considering that categorizers (a,n,v) are cyclic heads: However, there is evidence to show that such a combination is very restrictive because elements in the external domain can suffer root-conditioned allomorphy, provided that they are in certain structural configurations outlined below. These two generalizations are based on two main hypotheses: (H1) Contextual allomorphy is possible only with elements that are concatenated.
(H2) Cyclic Spell-Out domains define which nodes are present in a given cycle of PF computation and thus potentially ''active'' (capable of being referred to) for the purposes of contextual allomorphy. In some cases, superficially adjacent nodes cannot influence each other allomorphically because in terms of cyclic Spell-Out, they are not active in the same PF cycle.
Returning to the analysis of BP data, we suggest that the lower syntactic structure associated with a complex prefix verb necessarily consists of four elements: the root, functional head R, v and Th. These structures may be provided latter in the derivation with Voice and other inflectional heads, which are not under investigation at this point (BASSANI, 2013(BASSANI, , 2015. The following structure, which is central to all verbs of change, represents a complex head which proceeds to linearization / morphological reorganization after the syntax output. The head R is specified to precede the root, taking place superficially as a prefix, and Th is inserted after syntax (OLTRA- MASSUET, 2000). c. v cannot have its form contextually determined by R because the root will never be phonologically null.
If proved that R insertion is contextually determined by the root, we will be contributing to an important theoretical discussion of Distributed Morphology theory about the nature of roots. Insertion at R occurs before a possible vocabulary insertion at the root node, since linearization is a morphological operation that must apply before vocabulary insertion, and which, in this case, places R linearly before the root. If it is proved that the insertion of R depends on the root, we must assume that the root has to be provided with phonological material before the first vocabulary insertion of the whole structure, at R. If there were late insertion for roots, there would be no material to determine the insertion of R. In the following sections we test and discuss the relevant predictions.

Vocabulary insertion at R: one or many prefixes?
We propose that the prefixes a-, en-and es-are vocabulary items that realize the functional head R, which minimally contains a purely functional relational feature [+ r]. We have seen that change of state, abstract possession and configuration verbs do not imply any physical transfer (PT below) and there is no directionality encoded in the formation (DIR), we have also seen that in change of concrete possession verbs there is some notion of physical transfer without specification of directionality and that in change of place verbs, there is apparently a directionally specified by the nature of the root (source of directionality -SDIR). The table below shows that in every subclass of verbs of change there is variation in the phonological form that realizes R, with some numerical trends: Being attested that the same prefix form can occur in classes where there is no expression of physical transfer or directionality and that it can also occur on a smaller scale correlated to these interpretations, the questions that arise are: how are the features of R specified? Does R carry directional features? There are two logical possibilities, that we will call analysis by total specification (A1) and analysis by underspecification (A2).
(A1) Total specification: for every node R there is a specific set of syntacticsemantic features that are realized by specific vocabulary items.
(A2) Underspecification: for every node R there are multiple syntacticsemantic features configurations depending on the element to which it attaches and there are multiple vocabulary items may be inserted in R.
The analysis by total specification is that in which R has a specific and predetermined set of features since List 1, and there are specific vocabulary items matching these nodes. The implication of this analysis is the assumption that there are three types of Rs, different but similar, and each set of features in each terminal node must have a corresponding vocabulary item in perfect compatibility as represented in sequence: This alternative is not empirically supported. The observation of the data shows us that the feature [+ dir], which represents general directionality, and its possible specifications {near}, {inside} and {out}, are rarely interpreted and depend crucially on the semantic type of the root.

(A1) Analysis by total specification
Moreover, in such an analysis, the fact that we have the presence of prefixes a-, en-and es-in all semantic classes (except for es-in change of abstract possession) ends up as a simple case of homophony and, more generally, the analysis resembles a classic lexicalist one, resulting in a list of several homophonous prefixes organized by representativity, as exemplified below: These lists are widely used in traditional grammars, but have no explanatory power. Moreover, they do not capture the semantic, morphological and structural similarities between the "many" homophones prefixes: it becomes a mere coincidence. Such complete homophony analysis applies well to cases such as the phonological form /a/ from Portuguese, for example, which is the phonological realization of a negative prefix (anormal 'abnormal'), can form verbs of change (amaciar 'to smooth') and is a definite feminine determiner (a menina 'the.FEM girl'). In this case, the set of features, the morphological and categorical environment of each of the forms are completely unrelated.
An analysis that works purely by locality of attachment is very attractive and compatible with DM and could also preserve the full specificity of R. However, it must also be discarded for the BP data based on empirical evidence. Svenonius (2004) argues that the division between lexical and superlexical prefixes should be analyzed in terms of the different positions of the prefixes in a syntactic decomposition of the clausal structure. More precisely, he proposes that 'lexical' prefixes (with resultative, spatial and idiosyncratic meanings) are attached below the VP (Verb Phrase) and superlexical prefixes (with aspectual and quantificational meanings) are attached above VP. Let's look at the proposed structure for a Serbian word of that contains an inner prefix with resultative meaning (iz-) and an outer prefix with distributive meaning (po-): (43) Po-iz-bac-i-va-ti DSTR-out-throw-v-IMP-INF 'Throw out one by one' (SVENONIUS, 2004, p. 239) We cannot pursue this type of analysis for our data because we have no empirical evidence to claim that possible different semantic contributions of prefixes a-, en-and es-are the result of different attachment loci in syntactic structure. As we have seen, evidence shows that they behave identically when it comes to argument structure and distributional properties and must be attached below vP.

(A2) Analysis by Underspecification
The facts that make us disregard (A1) lead us to conclude that the supposed prefix directionality is residual and structurally defined. It is residual because the heads realized by -a, en-and es-must have been intrinsic directional at some point in time, but it is not true for the majority of the cases now, and when it is present (in a minority of cases) it is structurally defined in the sense that it shows up only in combination with certain root types.
For these reasons, we propose here an analysis in which the directional features in the terminal node are deleted (or become opaque) from the syntax when the root is not able to "activate" them, and then it makes competition for vocabulary items possible at the moment of vocabulary insertion, causing contextual allomorphy. We have to assume a deleting operation other than classical Impoverishment, as proposed in Bonet (1991) and Halle and Marantz (1993), since when Impoverishment applies it causes post-syntactic deletion of features and do not prevent its interpretation at LF. As directional features are not interpreted when relevant roots are not present, we propose their (non) activation in the syntax before spell-out and hence before vocabulary insertion.
In the derivation of a change of state verb, a change of abstract possession verb, or a change of configuration verb the possible directional features of R are irrelevant and are thus not interpreted.
(44) Change of state / abstract possession / configuration RP clause We propose the derivation of these subclasses happen in the following way: in the syntax, the R head merges with roots of the type state or entity. As these roots do not make any directional feature of the head R to be relevant, there is only one relational feature in R, represented by [+r]. After the complex head is linearized, the result is a complex verb head like the one below in the structure (45), and in (46) we represent the concatenation patterns derived from the structure: (45) Complex verb after linearization (46) R√ROOTvTh During vocabulary insertion, the vocabulary items competing for insertion at R must be underspecified as proposed in (47). If they are fully specified as drafted in hypothesis A1, insertion will not be possible since there will be conflicting features between the VI and the terminal node in cases like this, where directional features are not present. Besides, the information that the VI to be inserted has affix nature has to be added so that prepositions, non-affix VIs, are not part of the competition for insertion in R as they possibly also have the feature [+r].
According to the subset principle as defined below (HALLE, 1997, p. 128), and considering the feature present in R and the eligible VIs, all items above have the same chances of insertion at R: Subset principle: The phonological exponent of a vocabulary item is inserted into a morpheme in the terminal string if the item matches all or a subset of the grammatical features specified in the terminal morpheme. Insertion does not take place if the Vocabulary item contains features not present in the morpheme. Where several Vocabulary items meet the conditions for insertion, the item matching the greatest number of features specified in the terminal morpheme must be chosen.
The first hypothesis to investigate is whether the insertion of R could be guided by purely phonological conditioning of the root, since they end up directly concatenated after syntax. That is, could the first segment of root or its prosodic form determine the choice among items a, b and c above? In fact there are two general phonological restrictions on prefix attachment: 1) only Roots started by consonants can be prefixed; when the Root is started by a vowel, only suffixation is available (with very few exceptions), as in (48), or the formation of a verb with no affixes, as in (49); 2) there are no cases of attachment of prefix es-to roots started by strident fricatives ([s], [ʒ], [ʃ]). In additional to the discussion made above, in a theoretical level, this phonological sensitivity to the root first segment form is strong evidence that late insertion does not occur for Roots since it proceeds first in the prefix node.
(48) a. ameno 'mild' b. *a/en/es.amen.ar 'to ease' c. amen.iz.ar 'to ease' (49) a. azul 'blue' b. *a/en/es.azul.ar 'to turn blue' c. azular 'to turn blue' In other cases, the forms /a/, /en/ and /es/ can be attached before all consonants. Furthermore, although our research on this topic is brief, it does not seem that there are prosodic constraints for prefixation, since it occurs with bases of one, two and three syllables, and the position of the stress on the base also appears to have no influence in this respect. In this sense, we can disregard the hypothesis that phonology guides the choice among a, b and c. However, the phonological rules applying between the prefix and the Root serve as evidence for the proposed structure, with R attaching inside vP.
The second hypothesis to be investigated is whether the choice of the form of R may be guided by the form of v. Although this possibility is not predicted by C1-LIN theory, since they are not directly concatenated, it is worth empirical scrutiny, since positive results in this regard would be evidence to prove the existence of circumfixes (as assumed by Pereira (2007) and Rio-Torto (2004)). In the case of the verb emagrecer ("to lose weight"), for example, we would have to assume that the insertion in R and v occurs at the same time and that en-x-ec is a circumfix. There are two empirical arguments to discard this hypothesis, one of them has already have advanced in more general terms, and there is one extra theoretical argument.
The first empirical argument is based on the independent occurrence of such prefixes and suffixes. The data in table 11, below, show contexts in which prefixes occur either alone (column 2) or with different suffixes (columns 1 and 3). Furthermore, the same suffix occurs concomitantly with the three prefixes (column 1) and also with no prefixes (column 4) and the same prefix occurs in different contexts (rows 1, 2 and 3). Also, consider the data in column 1 in comparison with 2 and 3: the same prefix form co-occurs with different theme vowels. 12 The second empirical argument is based on the existence of doublets or morphological pairs like the following: (50) Doublets a. en-ric-Ø-ar vs. en-riqu-ec-er 'to enrich' b. en-doid-Ø-ar vs. en-doid-ec-er 'to freak out' If the prefix choice did in fact influence the suffix choice, or vice versa, these pairs should not have the same meaning, and they do. If the meanings were different, we could think of different (semantic) functional heads R and v and the contexts of insertion could be different, but this is not the case. In addition to the empirical arguments, we believe that the fact that there is no other case of circumfixation in Portuguese is also a valid theoretical argument to rule out this analysis.
After discarding phonological conditioning and contextual conditioning by v, the alternative left then is that the choice of the prefix form is purely idiosyncratic and depends on contextual allomorphy determined by the Root. This alternative is covered by C1-LIN theory and attests its predictions in the following way: it partially attests 12 First column shows vowel e for second conjugation verbs and the other cells vowel a for first conjugation class. Third conjugation class vowel i is restricted to non-transparent complex verbs (BASSANI, 2015). We won't be able to fully explore it in this paper, but theme vowel data shows that insertion at R is not influenced by Th and vice versa, as predicted by C1-LIN. prediction a. and fully attests prediction b. reproduced below and, at the same time, presents itself as empirical evidence against the idea of late insertion for roots. a. The root can contextually determine the form of R and v; b. R cannot have its form determined by v and Th since the insertion occurs from the inner to the outer node; The vocabulary items in (47) a, b and c above, which are equally eligible for insertion at R in all the structures below could generate emagrecer, amagrecer, esmagrecer as much as aquartejar, enquartear, esquartejar and afeitiçar, enfeitiçar and esfeitiçar. Therefore, we must assume that there is contextual information specified in vocabulary items in the form of lists of roots, as shown above, in vocabulary insertion rules to be read as follows: a. the feature [+r] is phonologically realized as /a/ when concatenated with Roots from a list of Roots, say List A; the feature [+r] is phonologically realized as /en/ when concatenated with Roots from a list of Roots, say List B; the feature [+r] is realized phonologically as /es/ when concatenated with Roots from a list of Roots, say List C. The multiple realizations of the same feature by different phonological forms is a case of contextual allomorphy.
Our analysis of this kind of choice as idiosyncratic leads us to the predictions that this specification must be learned by native speakers and also during second language acquisition. We won't be able to investigate such predictions now, but it is expected that speakers produce mismatches during the acquisition of complex verbs, so that a different prefix is used instead of the expected one.
At this time, we can clearly distinguish two types of allomorphy defined by the components in which they are derived. In the case of contextual root conditioning, the allomorphy occurs at the moment of vocabulary insertion, different from phonologically conditioned allomorphy that is given in PF. Therefore, the allomorphy present in change of state verbs amaciar (to smooth), emagrecer (to lose weight) and esvaziar (to empty) is contextual while the allomorphy in engordar (to fatten) and emagrecer is phonological, since it is the alternation between /en/~/em/~/e/ caused by application of assimilation followed by deletion (dissimilation): emagr- The derivation of change of place and change of concrete possession have to be different from change of state, possession and configuration verbs since the former shows directionality influenced by the nature of the root, which is a place, and the latter have non-specified directionality also influenced by the specific concrete nature of the root. In the second case, there is only one physical transfer without specification of directionality of movement. In other words, the feature [+dir] needs to be interpreted at LF in these structures. We propose the next two structures to be derived for these cases along with examples resumed for clarity: (53) Change of place verbs c. a-prision-a-r Data suggests that in change of place verbs the piece en-is correlated to the feature [+dir {in}] while a-is less specified and es-is very residual. It seems like that en-is the only piece that still carries some kind of compatibility to be written in its insertion rule. However, it is a fact that this compatibility is not relevant when roots are properties of states, for example, and en-is still widely inserted (as in enfraquecer 'to weaken', empobrecer 'to impoverish', and encurtar 'to shorten').
For this reason, we are going to propose that there are two homophonous vocabulary items: the one described in (47) As es-is very residual, we don't have enough consistent evidence to propose a vocabulary item related to [+dir {out}]. Evidence for this comes from the fact that es-is not able to form new words. The feature out is only productively encoded in roots that form verbs like tirar ('take out'), remover ('take away, remove') and in fora ('out'). As the directionality encoded in verbs with a-is light and seems to be inserted in the general case, we can still use the same vocabulary item inserted in change of state, change of abstract possession and change of configuration verbs. For verbs of change of concrete possession, we assume that the nature of the root makes possible for the directional interpretation to remain visible, but do not further specify the directional feature. Then, we propose the following structure to be derived. In relation to insertion rules, we have seen that there are no requirements of compatibilities between the types of root and the pieces inserted. For this reason, we can use the same vocabulary items proposed in (47) for change of concrete and change of abstract possession verbs, and we will assume that the choice of the allomorph is idiosyncratic. The fact that en-is numerically greater than in this subclass a-remains as a coincidence.
We conclude that this lack of directional information encoded in the vocabulary items is due to the fact that encoded directionality in the verb in BP has decreased, and is perhaps fossilized information in the data analyzed. Evidence for this come from data like the ones from (58) through (63), which are apparent cases of redundancy in which the directionality previously encoded in the prefix within the verb happens to be encoded out of it, in a PP element. Bassani (2015, p. 124) presents data that shows that speakers do not interpret directionality in the fossilized prefixes ((58) to (61)): As this paper focuses on the study of prefixes it is not our primary concern to detail the analysis for the vocabulary insertion at v. However, we have evidence to conclude that the insertion at v can be contextually determined by the root or by a categorizer merged before v (n or a), so it is also defined by locality conditions. Within this scenario, the alleged prefix-suffix dependence in change verbs is an epiphenomenon derived from the fact that both vocabularies insertions at R and v occur in a context in which both are directly concatenated with the root, as in the representation in (46) resumed in (64).

Concluding remarks
The main contributions of this paper are situated within the search of a better understanding about the internal structure associated with complex verbal heads, as part of the ongoing research program of Distributed Morphology, but which has been theme of discussion of many approaches, such as Lexical Semantics and Nanosyntax. For this, we have focused on prefixes in Brazilian Portuguese verbs of change.
We have proposed that these prefixes are the phonological realization of an inner and active functional head labeled R attaching outside the Root domain, but not above the first Categorizer/Cyclic Head v, which minimally contain the feature [+r]. We have also offered a detailed description of four semantic subclasses regarding the root meaning in order to investigate the correlations between the pieces of morphology, root semantics and argument structure. We came to the conclusion that the feature [+dir], which represents general directionality, and the possible specifications {near}, {inside} and {out} as dependent and specific features of [+dir], is rarely interpreted and depend crucially on the semantic type of the root. With this, we hope to have made clear the relevance of the root semantics in locality with the head R to derive the interpretation of the structures.
Concerning the prefix form, we have not found strong evidence associating a unique type of prefix to a particular Root or structure. We have evaluated two working hypothesis for the vocabulary insertion at R: total specification and underspecification. We have argued for an analysis by underspecification of the vocabulary items and made use of contextual conditioning of the root to explain the prefix choice, what characterizes a case of contextual allomorphy. We also have called attention to the fact that the apparent prefix-suffix dependence in change verbs is an epiphenomenon derived from the fact that both vocabularies insertions at R and v occur in a context in which both are directly concatenated with the root.
It may seem that our analysis for insertion at v and R is too idiosyncratic. However, it confirms in BP data the known fact that the choice of prefixes in Romance languages is widely idiosyncratic. This fact alone leads us to confirm that the choice of prefix and suffixes forms is, in fact, guided by idiosyncrasy and therefore is specific to each language. Observe the differences in the table below where we compare a small amount of verbs in Portuguese, Catalan, Spanish, French, Italian and Galician 13 . Even in Latin, in inchoative verbs, there are big prefix differences. Take for example the difference between duresco and vesparesco corresponding to endurecer (to harden) and entardecer (to become evening) respectively. 13 An anonymous reviewer has pointed out that some translations and uses among the Romance verbs are not completely comparable and that English translations are not always exact. The comment is relevant and could lead to a fine-grained inspection, but the main objective of this table is to show in general terms that the realization of the prefixes may vary in form and sometimes are non-existent in somehow related verbs. We conclude then that despite the fact that the prefix phonological form is defined in idiosyncratic ways, its presence is deeply related to a certain syntactic configuration, giving support to constructionist approaches to word formation.