2.1 Theoretical Assumptions

In DM, an acategorial root is adjoined to categorising heads such as a, n, and v (Wood & Marantz 2017). The category assigning v head verbalises the root and merges the internal argument(s) as complement(s). The external argument is introduced by the higher functional head, Voice (Kratzer 1996; Pylkkänen 2008; Marantz 2013). Following Alexiadou & Anagnostopoulou & Schäfer (2006) and Pylkkänen (2008), we take v and Voice to be separate nodes along the verbal spine with the external argument being located in Spec, VoiceP. T is the locus of tense with only the exponent T, since Malayalam does not express phi-agreement.

Since Malayalam, like many other languages, has overt intransitive morphology, we assume a uniform Voice head (following Alexiadou & Anagnostopoulou & Schäfer (2015) and others (Kastner 2019; Nie 2020)) for all verb types including unaccusatives where the Voice head, while present, nonetheless prohibits a specifier. To generate the different verbs, we propose the following. We assume that v in Malayalam is covert/null and that Voice can bear the features [+transitive], [–transitive] or be unspecified.² While the feature [–transitive] is realised by the vocabulary item -N, [+transitive] is phonologically null, but always induces phonological changes on the verb stem. Unspecified Voice has no phonological content. Phonologically null elements of the morphosyntax like v and unspecified Voice are assumed to be removed by pruning (Embcik 2010).

We also adopt the Voice theory (Alexiadou & Anagnostopoulou & Schäfer 2006; 2015; Kastner 2019) which takes causative alternation to be a Voice alternation without the introduction of an additional causing event into the syntax. In general, valency changing properties are attributed to v and Voice, and the morphemes indicating transitivity are assumed to be variant expressions of Voice as has been shown for German (Alexiadou & Anagnostopoulou & Schäfer 2015), Greek (Alexiadou & Doron 2012), Hebrew (Doron 2003; Kastner 2019) and other languages (Schäfer 2009). Both the lexical (formed via phonological change) and the morphological (formed via affixation) causatives of Malayalam are low-causatives which are cross-linguistically shown to be transitives, and to possess idiosyncratic properties such as differences in productivity, semantic opacity etc. (Marantz 1997; Travis 2000; Embick 2004). While the phonological changes that the verb undergoes to signal an increase in valency is assumed to be induced by the [+transitive] feature on Voice, the exponent -kk that marks such valence change is not an exponent of Voice, but an expression of √agent. Our treatment of -kk as an agentive morpheme, and not as either a causative or a transitive affix, follows from the fact that the exponent occurs with unergatives, transitives, and causatives (see Section 2.2). We will also show in detail in Section 3 that √agent may adjoin at v or Voice, or at both.

This approach is similar to that of Kastner’s for Hebrew. Kastner (2017) introduces an element √action which may be adjoined to Voice and which imposes a [+human] requirement on the DP in Spec, VoiceP. He sees √action as an ‘agentive modifier with predictable spell-out and consistent semantics’. Since this element does not have any syntactic requirements but has phonological and semantic content, Kastner proposes that √action is root-like. We believe that this is true of the √agent that we are proposing for Malayalam as well. While √agent has semantic and phonological content, it does not have an impact on the syntax and, hence, it is also root-like. However, our proposed element √agent differs from Kastner’s √action in the following ways. √agent may adjoin to either v or to Voice, to even both or to neither. It does not impose a [+human] (or other) requirement on the external argument in Spec, VoiceP. Our proposed structure also differs from the Voice-over-Voice structure proposed by Nie (2020) for morphological causatives in languages like Halkomelem, Tagalog, Japanese, and Kinande. Unlike these languages, a single Voice head is enough to account for causative alternations in Malayalam, since morphological causatives are not formed via the affixation of transitive or causative morphemes but via √agent adjunction.

There is much cross-linguistic evidence to differentiate agentivity and causation (Sichel 2010; Alexiadou et al. 2013). In Malayalam, the locus of adjunction of √agent is directly correlated with the interpretation of the external argument as an agent or as a causer. The addition of √agent at v signals the direct performance of the action by the subject while its addition at Voice always signals causation, i.e., the external argument is the causer. However, both external arguments are agents, one that acts and one that causes. By default, Malayalam roots may have inherent agentive semantics which in turn will generate unergative and transitive structures without an overt agentive affix at v. However, there are certain roots that can only receive added agentive semantics through √agent adjunction at v. There are phonological properties of roots that distinguish the former type of verbs from the latter type. However, both kinds of roots can undergo √agent adjunction at Voice to produce low causatives. We see √agent on v as modifying the external argument of the root’s event structure (root semantics) marking it as an agent, while √agent on the Voice modifies an added external argument, associated with intransitive/causative alternations.

The differences discussed above also underlie the two different valence changing processes in Malayalam that we outlined earlier, that is, phonological change on the verb (lexical causatives) versus -kk affixation (morphological causatives). These two processes are in complementary distribution. Intransitive to transitive alternations (e.g. aaʈ- ‘swing_intr’ – aaʈʈ- ‘swing_tr’) are effected on the verb stem by a Voice that bears the [+transitive] feature. Affixal transitives are derived, in our account, by the adjunction of √agent at v, and morphologically instantiated as kk-affixation. Thus, combinations of Voice_[+tr] and √agent-v are in complementary distribution, and their co-occurrence is ruled out.

Contrary to the view offered by Amritavalli & Jayaseelan (2008) which argues that Malayalam lacks Tense, we assume here that Malayalam has the grammatical category Tense and a TP layer, as has also been argued for by Babu & Punnappurath (2003), Menon (2011), and Swenson (2017; 2019). These latter accounts all agree that Malayalam marks the past tense overtly, and this agreement is adequate for the purposes of this paper. However, Babu & Punnappurath (2003) and Swenson (2017) posit -u as a past tense affix, but we believe that this identification is incorrect. In concurrence with Rajaraja Varma (1917) and Krishnamurti (2003), we take the two past tense exponents to be -Tu and -i.

We propose (10) as the basic morphosyntactic structure for Malayalam, a head-final language. Since this paper focuses on the generation of different verb structures stemming from the combination of various morphosyntactic features on v and Voice together with the subsequent selection of the past tense allomorph, we will use the minimal structure given in (11) in the discussions that follow.

2.2 Roots, affixes, and verbal forms in Malayalam

We can identify two primary types of verb structures in Malayalam: i) structures with no intervening affixes at v/Voice between root and T,³ and ii) structures with phonologically expressed features/affixes at one or the other of v and Voice or at both. In the former type, the semantics of the root is fixed when it combines with v and this structure can yield unaccusatives, unergatives, and transitives. For example, minn- ‘strobe’ (12) is an unaccusative verb with an internal argument which must raise to Spec, VoiceP, while paaɖ- (13), an action, is a transitive verb with an internal argument (an object) as well as an external argument. The roots which generate unergative verbs and transitive verbs have inherent agentive semantics.

These verbs directly condition the allomorphy on T since, at Spell-Out, the categorised roots and T will come to be linearly adjacent to each other (assuming pruning) as shown in (14).

Any intervening exponents at the heads v and Voice have a crucial role to play both in the semantic interpretation of the verbs and in determining the alternation, intransitive or transitive. In these structures, the roots do not and, in fact, cannot condition the allomorphy at T since there are intervening affixes at v and/or at Voice, and the root cannot come to be linearly adjacent to T during Spell-Out. For example, absent affixes at Voice, v will come to be in a linearly adjacent position to T at Spell-Out and will provide the conditioning context for the past tense selection at T as in (15a). If there are affixes at Voice, then Voice and not v will provide the context for conditioning the past tense allomorphy at T as in (15b). We use dotted lines with and without a cross to visually represent possible/impossible adjacency relations. In short, a root can enter into different kinds of derivations by combining with v and Voice with varying feature specifications and derive verb spines with varying syntax and, consequently, semantics. We describe each structure in detail in Section 3.

When the categorised roots themselves directly condition the allomorphy at T, the past tense exponent is realised as -i. In all the other cases, where the intervening affixes (at v and/or at Voice) condition the allomorphy at T, the past tense exponent that is selected is always -Tu, whose initial consonant /T/ is phonologically place-underspecified.

Malayalam has a number of transitive-intransitive pairs of verbs created from the same root.⁴ Diachronically, the two affixes -kk and -N marked the transitive and intransitive (typically, unaccusative) verbs, respectively. Evidence for this is readily available in the history of Dravidian languages. For example, the root kuɻa- forms the transitive (16) and the unaccusative (17) counterparts through such affixation of -kk and -N, respectively.

Krishnamurti (2003) identifies -kk as the transitivity affix and -ntu as the past tense marker for Proto-Dravidian intransitive verbs. Rajaraja Varma (1917) and Sadanandan (1999) also consider the nasal to be part of the past tense suffix. However, we differ from these accounts in arguing that the feature specification for -kk is not [+transitive] but [+agentive] and, further, that the past tense marker for intransitive verbs is not -ntu or -Nu, but -Tu, the actual past tense suffix, seen in combination with the [–transitive] affix -N. If -kk were a transitivising affix as suggested by Krishnamurti (2003), a feature clash between -N characterised as [–transitive] and -kk characterised as [+transitive] is inevitable while forming the past tense. We note that all verb forms that are generated through the affixation of -kk, whether transitive or intransitive, are all agentive. This becomes evident when we consider unergative verbs (agentive and intransitive) which have -kk and -N affixed at v and Voice, respectively, and select the past tense marker -Tu as shown in (18). Rajaraja Varma (1917) rightly treated -kk as an agentive suffix, but he considered these unergatives exceptional, since he could not explain the presence of the nasal. A closer analysis of these unergative verbs shows us that they are generated through the morphosyntactic combination in (19) with the corresponding morphophonological combination in (20).

Acquisition data provide us additional evidence for the affixal nature of -N and -kk. Omission errors in early productions are seen to target affixal morphemes and children often omit one or the other of these morphemes (among others) in their utterances, establishing their affixal status (Krishnan 2019).

Extant accounts of verbal morphology in Malayalam analyse the agentive -kk and intransitive -N markers quite differently. While Krishnamurti (2003) identifies it as a transitivity affix, Sadanandan (1999) treats -kk as a causative affix, and Killimangalam & Michaels (2006) argue that it plays three different roles, namely, a verbaliser, a transitiviser, and a causativiser. These characterisations are all inaccurate.

Killimangalam & Michaels (2006) and Michaels (2009) propose that -kk affixation applies across the board to all roots. In roots where -kk affixation is not seen overtly, they argue that it has coalesced with the root. However, they then concede that their coalesced roots’ behaviour with respect to past tense selection, compounding, and OCP (Obligatory Contour Principle) effects is inexplicably different from the roots that bear the affix overtly. We demonstrate that valence change is signalled in more than one way and that there is no need to suppose any coalescence of root with some underlying affix, because such an account fails to explain the morphosyntactic behaviour of these verbs.

We have already demonstrated that -kk is an agentive suffix (seen on intransitive and transitive verbs) as has also been pointed out by Marantz (2007). The unaccusative poʈʈ- ‘break’, for example, can be changed to the transitive poʈʈikk- ‘break’ by adding the agentive affix -kk. This is not causativisation. In fact, a second change can be effected by stacking the -kk suffix, i.e., adding the agent (a causer) with concurrent OCP (Sadanandan 1999; Killimangalam & Michaels 2006) induced dissimilation of the first -kk to -pp, as in (23). In this causative construction, the causee can appear as an (optional) oblique instrumental argument or may remain implicit as shown in (23c). The OCP effect that changes -kk-kk to -pp-kk as in poʈʈippikk ‘break_cause’ only occurs when there is a stacking of two -kk affixes. All sequences of -kk-kk are not similarly treated by the phonology, and OCP application requires this recursive morphological context. We must also point out that the visible -cc- sequence in the past tense forms as in (23c) poʈʈippiccu is not the product of OCP. The -cc- sequence is the result of assimilation of the underspecified /T/ of the past tense suffix -Tu.⁵

The feature [+transitive] on Voice has no exponent with phonological content, but induces alternations which are evidenced as direct phonological changes on v. Such changes are possible since root-v-Voice are thought to be part of the same phase (Marantz 2007). The feature [–transitive] is realised by the exponent -N which can only be phonetically interpreted if it is followed by a consonant-initial suffix with which it can share a phonological timing slot and acquire place of articulation features. Except for the past tense exponent -Tu, other verb inflections in Malayalam are either vocalic or vowel-initial suffixes. Hence, the nasal that marks [–transitive] becomes visible only in the past tense forms. The exponent -Tu at T can look back and anchor the nasal at Voice. Since both the nasal /N/ and the consonant /T/ of the past tense affix lack place features, they may assimilate a place feature from the root; absent a suitable root consonant, they may be assigned the default place value (which in Malayalam is dental). In the other tense forms, the nasal, lacking a supporting consonant, will fail to be phonologically represented and receive a phonetic interpretation and must delete, as shown in (24).

This is one of the main ways in which the past tense morphology helps us unlock and tease apart the components of the verbal spine. The mysterious appearance and disappearance of the nasal with the different tense affixes which has troubled many grammarians can now be clearly accounted for and easily understood in our account.

Conversely, the agentive exponent -kk is not visible in the past tense forms since it creates consonant clusters that are not permitted by Malayalam phonotactics when it co-occurs with the consonant-initial past tense affix. The agentive affix is, however, visible in all the other tense forms because resyllabification separates the clustered consonants using the available onset position of an adjacent vocalic or vowel-initial verbal suffix as shown in (25).

The complementary behaviour of these affixes in the various inflectional contexts not only supports our analysis, but demonstrates how phonotactics and morphophonology often obscure morphology, causing some underlying morphosyntactic properties to become invisible on the surface.

We summarise the preceding discussion briefly. Unlike what has been averred in earlier analyses, -kk and -N are agentive and intransitive affixes, respectively. The v in Malayalam is covert and Voice can be unspecified or bear the features [±transitive]. While Voice_[+tr] has no phonological exponent, it always induces phonological changes on the verb stem and the feature is made visible by such changes. Voice_[–tr] is instantiated by the exponent -/N/. √agent which can adjoin at either v or Voice, at both or neither is realised by the exponent -kk. The presence or absence of overt exponents in v and Voice is directly correlated with the selection of the past tense exponent; when the categorised root is linearly adjacent to T at Spell-Out, the affix -i is selected. When any overt exponents at v and Voice intervene, the affix closest to T determines the past tense allomorph at T, which is uniformly -Tu.

We must also point out that the past tense affix -Tu has several surface phonetic variants (which contribute to the surface complexity of the past tense morphology) in two series: (a) those without a nasal [-t̪t̪u, -ccu, -ʈʈu, -ʈu, -d̪u, and -ttu], and (b) those with a nasal [-n̪n̪u, -ɲɲu, -ɳɖu, -ɳu, -n̪u, and -n̪d̪u]. The variants are the result of a series of phonological accommodations that are driven by Malayalam phonotactics and that apply at morpheme boundaries after the insertion of the affixal material -kk -N, and -Tu. We do not discuss these variants of -Tu further in this paper, but simply note that the phonological accommodations follow straightforwardly given the underlying morphosyntax, which is our core concern here.

3.1 Form 1

Form 1 may be considered the default in Malayalam and includes unaccusatives, unergatives, and transitives. The semantics of the root is fixed when it merges with v, and varying argument structures (or valencies) are possible as seen in Section 2.2. The roots that have inherent agentive semantics can generate Form 1 unergatives and transitives. The sample sentences below range over unaccusatives (28) and (29), unergatives (30) and (31), and transitives (32) and (33), and the past tense marker that is selected is uniformly -i.

The functional heads v and Voice are both null in Form 1 and, at Spell-Out, the categorised roots come to be linearly adjacent to T after pruning, and can directly condition the allomorphy at T. The tense feature [+pst] is realised by the vocabulary item -i. Typically, the root is spelled-out with no change at all at TP, except for the intervocalic voicing of obstruents post-affixation as shown in the structure in (34).⁶ Examples of Form 1 verbs are given in (35).

Form 1 roots are, typically, phonologically heavy and, therefore, the selection of a vocalic affix is unsurprising. Open syllables are preferred in Malayalam and there is a strong tendency to break up heavy syllables by parsing a coda to a following onset as in (36).

There is, however, a small set of about ten monosyllabic verbs with the same default morphosyntactic structure as in (34), but with the past tense allomorph -Tu instead of the predicted -i as shown in (37). These (c)vc roots and a few (c)vi roots, all monosyllabic, show this exceptional pattern of past tense selection as in (38).

While the default roots that select -i may be either monosyllabic (e.g. pook- ‘go’) or bisyllabic (e.g. t̪u.ɭumb- ‘brim), none of them are of the syllable types (c)vc or (c)vi. With no difference in terms of the values on the functional heads, the selection of the past tense allomorph is the only difference between the typical Form 1 verbs and these exceptions which also range over different verb types (as is typical for this Form) as shown in (39) to (41).

The exceptional verbs demonstrate how phonology may interrupt morphology, and we offer here a brief discussion since it will help round out the analysis and provide some understanding of this exceptional behaviour. First, while vowel length is contrastive in Malayalam, the language has very few diphthongised, native (c)vi roots. Second, the coda position is typically restricted to the sonorant consonants [m, n, ɳ, l, ɭ, r]. When other consonants occupy the coda slots, they are usually parsed into an onset through [ɨ]-epenthesis, as in the examples in (42).

As Kenstowicz (1994: p. 46) says, “The term diphthong is used more broadly to denote any sequence of tautosyllabic vowels. They need not necessarily share any phonological features and the entire sequence may count as a single time unit (mora).” Malayalam uses these distinctions in calculating syllabic weight as reflected, for example, in poetic metre Rajaraja Varma (1917). The behaviour of monosyllabic cvc and cvi roots shows that consonants other than [m, n, ɳ, l, ɭ, r] and the second element in the diphthongs (as opposed to long vowels) do not count as separate morae, and such syllables must be deemed to be light. The syllabic ‘lightness’ of these predicates is exceptional compared to the other roots of this class which, uniformly, are phonologically heavy/superheavy. This may be one reason to choose -Tu instead of -i. For the already /i/-ending cvi roots, a homophonous tense allomorph creates a morpheme identification problem. Further, a root final obstruent in cvc roots will become voiced in this context, effacing more of the root’s already abbreviated form and making it homophonous with other words in the language. For example, the roots n̪aʈ- ‘plant’ and iʈ- form the past tense n̪aʈʈu and iʈʈu, respectively, with -Tu and not -i, possibly because n̪aɖi ‘actress’ and iɖi ‘thunder’ already exist in the language. This reasoning is also supported by the fact that disyllabic roots with the same final syllable shape choose the past tense marker -i as appropriate to their morphosyntax, as in ɲeɾaɖ-i ‘mashed’, kaɾud̪-i ‘considered’, and peɾug-i ‘multiplied’. These roots may have enough overall phonological material to support the choice of an -i while keeping the root’s identity.

Finally, the cvi verbs trigger the intervocalic voicing property that is visible with the rest of the class (and more generally), and the cvc verbs show obstruent place-assimilation. That the shape of the past tense affix is -Tu is, thus, readily recoverable from the phonological shape of the verb with tense. It appears, then, that the exceptional roots being phonologically atypical for the class and not being driven by a phonological impetus to reset the syllable structure may have ended up selecting the affix -Tu, which is readily available in the vocabulary as a past tense exponent. In essence, this small set of monosyllabic predicates of Form 1 are the only true exceptions to our past tense selection account. We now return to a discussion of the other forms and the verbal features.

3.2 Forms 2–4

The structures of the verb spines that derive Forms 2–4 have a phonologically null v and Voice as in Form 1. The difference from Form 1 arises from the inclusion of √agent at different loci. √agent is adjoined at v in Form 2, at Voice in Form 3, and at both v and Voice in Form 4, presenting a continuum of verb building operations. In the examples (43) to (45), a single root n̪anay- ‘to water’ enters into each of these three different morphosyntactic combinations, building the verbs with different semantics. While (43) is transitive, (44) and (45) are both low causatives which are syntactically transitives. These low causatives can have a single agent which can be interpreted as both the ‘agent’ and the ‘causer’ (44). Alternatively, the external argument of the verb may only be the ‘causer’ and the ‘causee’ can be encoded by an oblique argument, an instrumental postpositional phrase (45) (similar to the by-phrase in English passives). We must point out that the phonological shape of each output verb form is different and systematically signals the structural differences. We discuss each structure with further examples in the following subsections.

3.3 Forms 5 and 6

Unlike the verb forms that we have discussed so far with phonologically null v and Voice, Forms 5 and 6 are different in that Voice has the feature [–transitive] that is instantiated by the exponent -N. In addition to a Voice_[–tr], Form 6 also involves the adjunction of √agent at v.

3.4 Forms 7 and 8

In both Forms 7 and 8, the Voice head bears the feature [+transitive]. They differ in whether √agent is adjoined at Voice. The [+transitive] feature on Voice induces phonological changes on the verb stem. The adjunction of √agent provides the causative interpretation, as we saw in Forms 3 and 4. However, both these forms are syntactically transitive. While Form 7 always yields transitive verbs, Form 8 yields morphological causatives.

3.5 Discussion

In the preceding, we have shown in great detail how the features on v and Voice combine with √agent adjunction to generate each of the eight possible verb structures which build the various verb forms. We also demonstrated how the allomorphy for past tense marking at T can be directly linked to the properties along the verbal spine and, specifically, the exponents that may be expressed within it. We provided evidence for the argument that √agent may be adjoined at v or at Voice or at both loci with specific impact on the verb semantics. Specifically, we showed that the locus of adjunction of √agent determines whether the external argument is an agent (adjunction at v) or a causer (adjunction at Voice). We also showed that when the causer and causee are two distinct entities, the causee may be present as an oblique (typically instrumental) argument.⁸

A single root can participate in derivations with different verbal spines and generate different verbs with varying syntax and semantics, as can be seen in (84) to (86) and the various examples that we have already presented and discussed. The root ɟay- ‘win’ derives the Form 2 transitive by the addition of √agent at v, and the Form 4 causative by the subsequent addition of √agent at Voice as shown in (84).

In (85), the root muŋŋ- absent v and Voice features in a Form 1 structure can build an unaccusative verb, while a Voice_[+tr] of Form 7 will build a transitive verb from the same root. When a √agent is adjoined to Voice_[+tr], the Form 8 low causative is built.

The alternations in (86) show how the root can participate in a Form 5 structure with a Voice_[–tr] to make unaccusative verbs, a Form 2 transitive verb with the adjunction of √agent at v, and the Form 4 low causative by the addition of √agent at both v and Voice.

The [±transitive] features on Voice are made visible by the exponent -/N/ for the feature [–transitive] (87), and by the phonological changes induced on the verb by the [+transitive] feature (88).

From the examples above together with the data provided in the preceding subsections, we can establish that whenever there are exponents between the root and T, the past tense allomorph selected is uniformly -Tu. When there is no intervening affixal material between root and T, the past tense suffix selected is -i, which is also the default marker. The selectional difference of the past tense affix follows from this context driven allomorphy. The only true exceptions to these processes, as we discussed in Section 3.1 are the few light monosyllabic roots which select -Tu instead of the expected -i. Their exceptional behaviour, as we have already outlined in some detail, may be attributed to the atypical light syllable structure of these predicates, in contrast to the typically, phonologically heavy roots that participate in this structure. We have also seen that disyllabic verbs with the same final light syllable shape do not exhibit such exceptional behaviour and bear the expected vocalic affix -i.

We have seen that v is always covert in Malayalam and that Voice can be either unspecified or can bear the features [±transitive]. We have also shown that √agent can be adjoined to v and/or Voice and yield agent and/or cause semantics to the verb. The morphosyntactic difference is also underscored by the difference in the phonological shapes of Form 2 (√agent at v) and Form 3 (√agent at Voice) verbs. Though both verbal spines involve a single √agent which is instantiated on the surface as a single -kk exponent in both, Form 3 verbs (90) have an epenthetic -i, while Form 2 verbs (89) do not. This surface structure difference corresponds systematically to the difference in the loci of adjunction. The morphophonology, then, unambiguously represents the underlying morphosyntax. We also see similar effects in Form 4 with √agent adjoined at both loci (91).

We provide the vocabulary insertion rules for the derivation of the various verb alternants in (92).

From the features available, as given in Table 1, there is a potential combinatorial possibility of Voice_[–tr] with √agent at Voice. Such a combination does not and in fact cannot be realised because the two features are semantically incompatible. The adjunction of √agent at Voice necessarily adds cause to the verb meaning and a verb cannot be both causative and intransitive at the same time.⁹ Low-causatives are always transitive in Malayalam.

Table 1 also contains two blank cells where Voice_[+tr] is incompatible with √agent at v and, by extension, at both v and Voice. We may recall that √agent at only Voice is compatible with Voice_[+tr]. However, adjunction of √agent at v is mapped to an exponent -kk which adds the same semantics of agent to an already agentive Voice_[+tr]. The semantic outcomes are the same, and with both present, unrealisable. These are two complementary and distinct strategies through which Malayalam can instantiate the semantic changes, creating what we earlier called lexical vs morphological causatives. A Voice_[+tr] typically triggers sound changes on the v, while the √agent at v is realised by an exponent. The presence of both √agent at v and Voice_[+tr] sets up a conflict between the need for Voice_[+tr] to be adjacent to the root which it will be unable to accomplish with the presence of an exponent between v and Voice that will interrupt such adjacency. Valence change is instantiated through either the [+tr] feature on Voice (with phonological changes) or through the adjunction of √agent at v, but never both. Such an outcome is, therefore, impossible in Malayalam. Our analysis in fact rules out these semantically and syntactically incompatible combinations of [Voice_[–tr] + √agent] and [v- √agent + Voice_[+tr]] and accounts fully for the two kinds of valence alternations that are found in Malayalam.