(i) The binyan system is neither perfectly transparent nor completely opaque.
It has a degree of regularity (argument alternations) and irregularity (gaps and multiple, unpredictable interpretations) in it.
(ii) The semantic relatedness of words containing the same radicals and their meanings being too distant to relate derivationally can be explained if we assume that the semantic content of consonantal roots is not fully definable.
Roots are underspecified phonologically, in that 3 or 4 radicals are not pro-‐
nounceable on their own, and semantically, in that they acquire different in-‐
terpretations when they combine with different patterns.
(iii) The dual nature of the binyan system is also predictable if, along the lines of many theories of word building, I postulate two processes of word forma-‐
tion: one for the regular, productive processes (word derivation), another for the irregular, non-‐productive processes (root derivation). Verbs derived from roots may be assigned a variety of interpretations in different morpho-‐
logical environments (binyanim), while verbs formed from existing words retain the semantics (and phonology) of the base word.
In what follows I shall first discuss the contents of Maltese roots, arguing that roots are semantically not fully specified (Sect. 3.1). Following that, Sect. 3.2 is concerned with the form-‐function relation of the binyanim, demonstrating that the characteriza-‐
tion of the binyanim in terms of particular semantic roles (e.g., causative, reflexive) is untenable. Sect. 3.3 develops in detail the idea of regularity and irregularity in the bin-‐
yan system in terms of two processes of word formation, paving the way for the quantitative and qualitative analysis of the database of templatic verbs in Ch. 4.
3.1 The ingredients of roots
“In Arabic as in all Semitic languages the consonants of the word carry the semantic load.” Versteegh’s (1997: 19) claim is more or less the general idea a student of Maltese gets when perusing most grammatical studies on Maltese. Mifsud (1995a: 34) and Borg
& Mifsud (1999: 2) define the root as a “discontinuous morpheme of 3 or 4 consonants in fixed sequence, identifying the basic semantic area”. Looking at it from another angle, others hold that “words containing the same root consonants stem from one concept”
(Akkademja tal-‐Malti 1992: 27, my translation). Under this approach, it is the root that carries the core meaning of the word. I will refer to this view as the traditional ap-‐
proach.
As discussed above, a single root can build more than one word. Let us consider a few examples.
(1) √ktb C1vC2vC3 (v) kiteb ‘to write’
nC1vC2vC3 (v) inkiteb ‘to be written’
C1vC2C3a (n) kitba ‘writing’
C1vC2C2vvC3 (n) kittieb ‘writer’
C1C2vjjvC3 (n) ktejjeb ‘booklet’
(2) √lbs C1vC2vC3 (v) libes ‘to wear, dress’
C1vC2C2vC3 (v) libbes ‘to dress, v.t’
C1vC2C3a (n) libsa ‘dress’
In (1) above, all the words containing the root consonants [k], [t], [b] share a lexical core, having to do with some notion of writing. Similarly, for libes, libbes, and libsa in (2), the traditional claim would be that the root √lbs means something such as dressing or wearing.
However, looking at a larger sample of Maltese data, it becomes difficult to relate certain words with a common root in a systematic fashion. Consider the root √xrb in (3) below. While one may posit that most of the words containing [ʃ], [r], [b] share a common semantic core related to some sort of contact with liquids, they differ consid-‐
erably in meaning. The two verbs xorob and xarrab are derived from the same root, but denote two very different events, drinking and wetting, and cannot be (synchronically) derived from one another in any straightforward way. The same holds true for silet and
issielet in (4), which refer to distinct events, extracting and struggling. They are not eas-‐
ily reducible to a derivational relation, at least synchronically.
(3) √xrb C1vC2vC3 (v) xorob ‘to drink’
C1vC2C2vC3 (v) xarrab ‘to wet’
nC1(t)vC2vC3 (v) inxtorob ‘to be drunk/shrink’
C1vC2C3a (n) xarba ‘drink’
tvC1C2iiC3 (n) tixrib ‘wetting’
(4) √slt C1vC2vC3 (v) silet ‘to extract’
tC1vvC2vC3 (v) issielet ‘to struggle’
C1vC2C3a (n) silta ‘extract’
mvC1C2vvC3a (n) misluta ‘earring’
In examining these examples, we are confronted with contrasting evidence. Some word families with the same three or four consonants, such as those in (1) and (2), share a common semantic core and are semantically related. This semantic relation, however, is hard to pin down in other groups of words containing the same consonantal root, such as those in (3) and (4). Moreover, in the latter families, some members ap-‐
pear to be semantically related (e.g., xorob ‘drink’ and xarba ‘a drink’; silet ‘extract’ and silta ‘an extract’), but the meaning of some members is far apart from the meaning of other words with the same root (e.g., xorob ‘drink’, xarrab ‘wet’; silet ‘extract’, issielet
‘struggle’, and misluta ‘an earring’). One particular verb, inxtorob, seems to be a hybrid case of semantic closeness in the passive sense ‘be drunk’ and semantic distance in the sense ‘shrink’. This curious case is discussed in Sect. 3.3.2.1
3.1.1 Problems with the traditional approach
As the data in (5) and (6) below confirm, the main problem regarding the contents of the root is the following. It is very often the case that the meanings of words sharing the same root consonants are closely related on the one hand (e.g., daħal ‘enter’, daħla ‘en-‐
trance’, daħħal ‘insert’; ħarreġ ‘train’, taħriġ ‘training’), and too distant or unrelated on the other (e.g., indaħal ‘interfere’, dħuli ‘friendly’; ħareġ ‘take/go out’, stħarreġ ‘investi-‐
gate’).
(5) √dħl C1vC2vC3 (v) daħal ‘to enter’
C1vC2C2vC3 (v) daħħal ‘to enter, insert’
nC1vC2vC3 (v) indaħal ‘to interfere’
C1vC2C3a (n) daħla ‘entrance’
C1C2vvC3i (a) dħuli ‘friendly’
1 Further research is needed in this area to determine, in a systematic way, the semantic relatedness of words sharing the same root consonants, using Latent Semantic Analysis (cf. Landauer & Dumais 1997;
Landauer et al. 1998; Moscoso et al. 2005; inter alia), and such psycholinguistic techniques as the cross-‐
modal priming task (cf., e.g., Marslen-‐Wilson et al. 1994).
(6) √ħrġ C1vC2vC3 (v) ħareġ ‘to take/go out’
C1vC2C2vC3 (v) ħarreġ ‘to train’
stC1vC2C2vC3 (v) stħarreġ ‘to investigate’
C1vC2C3a (n) ħarġa ‘outing’
tvC1C2iiC3 (n) taħriġ ‘training’
The semantic relations between members of one family are too loose for a deriva-‐
tional account where one word is derived from the other, and at the same time too tight to admit an account where words are simply unrelated.2
Now, in most grammars of Maltese, the root generally comes with a meaning asso-‐
ciated to it. Roots are like dictionary entries with fully specified meanings. The first verbal pattern is presented as the mamma ‘roughly, matrix’, the source from which the other verbal (and sometimes nominal) patterns are derived. In some sense, it is consid-‐
ered as phonologically and semantically more elementary than the rest (cf. Sect. 3.2).
Accordingly, it is common practice to take the (primary) meaning of binyan I as the meaning associated with the root. All other forms are then based on the meaning of the root in that binyan. In this way, we often think of the meaning of √ktb as ‘write’, √xrb as
‘drink’, and √dħl as ‘enter’. However, as the data in (3) to (6) show, this approach runs into a number of problems, as roots acquire numerous nominal and verbal interpreta-‐
tions, which may or may not be related to the supposed lexical core.
In this traditional view, if words that contain the same root consonants have dif-‐
ferent meanings, this is because of polysemy or homonymy. Under the polysemy per-‐
spective, different meanings are derived from the root in some rather metaphorical or figurative way. Departing from the view that the root √dħl means something such as
‘enter’ (cf. binyan I, daħal), to account for the ‘interfere’ meaning it acquires in the envi-‐
ronment of binyan VII, indaħal, some linguists would argue that interfering means to enter in the middle of something that is not one’s concern. Likewise, because √ħrġ re-‐
fers to a core lexical meaning of taking/going out (cf. binyan I, ħareġ), then to train someone, the interpretation it acquires in binyan II, ħarreġ, would mean causing some skill to come out of the trained entity. And again, the meaning ‘investigate’ in binyan X, stħarreġ, would be understood as the “s of facts or information.
2 It is important to note that here I am not taking into consideration cases that involve homophonous roots, such as √sfr1 sfar ‘turn yellow’, √sfr2 saffar ‘whistle’, √sfr3 siefer ‘travel’. Such roots are distinguished, fol-‐
lowing Aquilina’s (1987-‐1990) dictionary, which points out different etymological sources for homo-‐
phonous roots. They are generally the result of historical merges of different phonemes (cf. Aquilina 1970;
Mifsud 2008). See also the brief discussion in Sect. 4.1.
The homonymy approach is adopted by Borg (1981, 1988). Positing that root meaning can be homonymous, he explains the semantic distance between ħareġ
‘take/go out’, ħarreġ ‘train’ and stħarreġ ‘investigate’ as forms derived from two homo-‐
nymic roots √ħrġ1 (ħareġ ‘take/go out’, inħareġ ‘be taken out’) and √ħrġ2 (ħarreġ ‘train’, tħarreġ ‘train oneself’, stħarreġ ‘investigate’). He takes the meaning ‘to investigate’ to be in some sense derived from the ‘train’ meaning, that is “to exercise (oneself) in relation to a matter” (Borg 1988: 241-‐242, my translation). Taking a similar approach, Camilleri (1990: 11) argues that ‘to investigate’ is an independent lexical meaning of √ħrġ3, as “it bears no semantic relation to any other verb derived from the [same] root” (my transla-‐
tion). Presumably, we would need a fourth instance of the root, √ħrġ4, to account for the interpretation it takes on in the nominal pattern C1vC2C2vvC3a, ħarrieġa ‘a stone pro-‐
jecting from the wall’.
It follows from this analysis that to explain the various interpretations of words containing the root consonants [d], [ħ], [l], we would need at least three homonymic roots: roughly, one meaning ‘enter’, a second one meaning ‘interfere’, and another one meaning ‘being friendly’. The homonymy approach leads inevitably to circularity, as a root is considered homonymous by virtue of the different meanings the words it ap-‐
pears in acquire.
I believe both treatments are influenced by a deterministic approach of the binyan system, where roots, specified with a full dictionary meaning (questioned in Sect. 3.1), interleave with patterns, which serve a number of fixed functions, such as causative, reflexive, passive (questioned in Sect. 3.2), in a semantically transparent way. While dictionaries and grammars present all word meanings as derived from the meaning of more basic elements leading down to the root, a large body of studies on other Semitic languages agrees that root-‐and-‐pattern morphology contains a significant element of opacity (cf., e.g., Ravid 1990). In what he terms the “root fallacy”, Barr (1961: 100) ar-‐
gues against the idea of transparency in the verbal system of Hebrew; that is, the view that “there is a “root meaning” which is effective throughout all the variations given to the root by affixes and formative elements, and that therefore the “root meaning” can confidently be taken to be part of the actual semantic value of any word or form which can be assigned to an identifiable root.” This theory is untenable and the claim that there is a degree of opacity is by now no longer controversial, at least in the treatment of such Semitic languages as Hebrew (cf. Arad 2005: 61).
As a consequence of this, some linguists opt for a diametrically opposed view, where roots are devoid of any semantic content. It is only words that have meaning.
The root is no more than a phonological tool for the formation of words, with no mean-‐
ing whatsoever (cf., e.g., Ravid 1990 for Hebrew). In this proposal, which we may term as the opacity or lexicalization view, semantic relatedness of different words sharing the same root is understood as a mere tendency, which can possibly be accounted for on historical grounds. There is no systematic relation between form and meaning, and any lexical item has a meaning which is idiosyncratic or lexicalized and which cannot be related to its form. It goes without saying that the theory of semantically empty roots is also at odds with the evidence presented above. There is, to a certain extent, a correla-‐
tion between the semantic relatedness of words and shared phonological material. As is evident from the data in (1)-‐(2), but also, in part, from the data in (3)-‐(6), there is meaning similarity across forms sharing the same root consonants.
To sum up, it is equally undesirable to have roots with no semantic content what-‐
soever and to have roots with clear semantic values, which are then inherited composi-‐
tionally by the words they form. Both transparency and opacity are theoretical ex-‐
tremes. In the discussion that follows, I argue that roots are neither semantically naked nor well-‐dressed. In order to explain the contrast between the semantic relatedness of words formed from a single root and their being semantically too far apart from other members of the same word family, I take an in-‐between path, which, to my mind, is the most direct analysis of the phenomenon.
3.1.2 Roots are underspecified
Underpinning this analysis are two assumptions which help unravel the puzzle of words containing one root being, at the same time, semantically close and semantically far apart. First, I consider, with Arad (2003a, 2003b, 2005), roots to be underspecified lexical and phonological cores. Second, in step with recent models of word formation, I assume that some words are derived from roots while others are derived from previ-‐
ously formed words. Let us now look more closely at the semantic content of roots. In Sect. 3.3, I then deal with the second assumption.
Taken on its own, the root is incomplete from a phonological point of view: [k], [t], [b] form a consonantal skeleton that is not pronounceable on its own. It is only when the root consonants are embedded in one or more patterns that they become a con-‐
tinuous string. Similarly, the root is semantically not fully specified, its meaning is not well-‐defined. As its phonological shape is not yet determined and as it does not yet have a fixed semantic interpretation attached to it, the root may be incarnated in many dif-‐
ferent ways, i.e. it may be cast into numerous nominal and verbal patterns, creating words with a specific phonological and semantic content.
This is not to say that roots are semantically empty, but rather that, paraphrasing Arad, they are “potentialities”, which may be realized in a number of ways.3 And, in-‐
deed, many Maltese tri-‐ and quadri-‐consonantal roots acquire several interpretations when they fuse with different nominal and verbal patterns. One of the conclusions that come out of my examination of the Maltese data is that this ability of consonantal roots to create multiple verbs when appearing in several binyanim is a chief characteristic of Maltese consonantal roots (cf. Ch. 4).4
In this respect, some linguists (e.g., Arad 2005: 101; Harley 2009, 2011) draw a parallelism between Semitic roots and Latinate bound roots in English, such as √ceive,
√pose or √fer, which create different verbs (conceive, perceive, receive or compose, pro-
pose, suppose or confer, infer, refer) that are semantically far apart from one another, and are by no means predictable from the combination of the root and word-‐forming morphology. See, among others, the discussion in Aronoff (1976) and Baeskow (2006) on the status of bound roots of Latin origin in English.
Consonantal roots are underspecified enough to take up meanings that differ con-‐
siderably from one another, such as ‘drink’ (xorob) and ‘wet’ (xarrab), or as ‘take/go out’ (ħareġ), ‘train’ (ħarreġ) and ‘investigate’ (stħarreġ). However, there may be de-‐
grees of underspecification, which are arbitrary, just like the linguistic sign itself. Al-‐
though all roots are semantically underspecified, some roots can be more underspeci-‐
fied than others. There is thus a difference between the degree to which roots such as
√ktb and √xrb are underspecified. While the words containing the former root are all related to the activity of writing (kiteb ‘write’, inkiteb ‘be written’, kitba ‘writing’), the agent who performs it (kittieb ‘writer’), and its end result (ktieb ‘book’, ktejjeb ‘book-‐
let’), the words that share the latter root differ greatly in meaning (cf. the examples in (3) above). The more underspecified the lexical core of the root is, the more likely it is to be assigned different interpretations in the environment of different patterns (cf.
Arad 2005: Ch. 3). But what does it exactly mean to be semantically underspecified?
3 “Because roots are underspecified potentialities, the interpretations assigned to them in the environment of patterns are to a certain extent arbitrary. They are not predictable from the semantics of the root, nor are they compositionally computed by combining the denotation of the root with that of the pattern” (Arad 2005: 17).
4 While multiple verbs formed from a single consonantal root are typical of Hebrew and similar languages (cf. Arad 2003, 2005), they are uncommon in a language like English. In Maltese, this is typical of templatic verbs but not of concatenative verbs, where roots generally combine with only one verbalizing suffix (Cf.
Ch. 5). The issue of language-‐specific properties of roots is briefly referred to in Sect. 3.2 & Sect. 3.3.
Semantic underspecification
The concept of underspecification, that is, the notion that certain features are not ex-‐
pressed in a representation, turns up in different linguistic domains. It emerged in pho-‐
nology but was adopted in semantics in the 1980’s (cf., e.g., Bierwisch 1982, 1988;
Bierwisch & Lang 1989). In lexical phonology, underspecified representations have been proposed to account for, among others, variation of phonetic realizations in con-‐
nected speech (cf. Kiparsky 1982; Steriade 1995; inter alia). For instance, the phoneme /n/ is realized in many different ways, depending on the onset of the following word: as a coronal in on duty, a dorsal in on call, or a labial in on board. According to the idea of underspecification, in English the nasal is underspecified in terms of place of articula-‐
tion and voicing. That is, while /n/ can be realized in many different ways depending on the context it is in, the underlying representation of the phoneme is thought to be the same in all cases.
At the lexical-‐semantic level, the notion of underspecification has been particularly fruitful in the discussion on ambiguity and the representation of lexical items with dif-‐
ferent senses (polysemy) and different meanings (homonymy), cf. Bierwisch 1996, 1997), Blutner (1998a, b), inter alia. It is also used in the lexical semantic analysis of affixes. For instance, Lieber (2004) considers derivational suffixes like -er to be seman-‐
tically underspecified (she uses the term “underdetermined”) in English in order to ac-‐
count for the different interpretations, such as agent (writer), patient (loaner), instru-‐
ment (opener), etc., associated with it. Similarly, von Heusinger & Schwarze (2002, 2006) avail themselves of the notion of semantic underspecification to explain the am-‐
bivalent nature of the prefix s- in Italian denominal verbs of removal. The prefix may in fact refer both to the removal of the base noun in verbs like s-bucciare ‘peel’ (cf. buccia
‘peel’) as well as to the removal of another entity from the nominal base, as in s-tanare
‘cause to come out of the burrow’ (cf. tana ‘burrow’). In line with the model of two-‐level semantics, which posits a distinction between a level of meaning defined by grammar and a level of interpretation based upon conceptual knowledge, von Heusinger &
Schwarze argue that the two subtypes of denominal verbs of removal have one under-‐
specified semantic representation, which is then specified by conceptual categorization.
Much work in this area has been carried out in computational linguistics and psy-‐
cholinguistics, and a number of proposals have been made of what an underspecified lexical representation might look like (cf. Rayner & Duffy 1986; Copestake & Briscoe 1995; Pustejovsky 1995; Poesio 1996; Frisson & Pickering 2001; Bos 2004; Sereno et al.
2006; Frisson 2009; Egg 2010a; inter alia). In the realm of psycholinguistics, the idea of underspecification was put forward to account for experimental data, especially eye-‐
tracking research in homonymous and polysemous words (cf., e.g., Frisson 2009 for an overview). In Natural Language Processing, semantic underspecification is seen as an important technique to keep semantic representations tractable in spite of the prob-‐
tracking research in homonymous and polysemous words (cf., e.g., Frisson 2009 for an overview). In Natural Language Processing, semantic underspecification is seen as an important technique to keep semantic representations tractable in spite of the prob-‐