The Absence of Islands in Akan: The Role of Resumption

Abstract

The precise nature of Ā-dependencies that terminate in a pronoun has been a long-standing subject of cross-linguistic research. Traditionally, it has been assumed that there are two derivational strategies to form resumptive Ā-dependencies: movement and base generation. Island configurations have played a crucial role in determining which derivational strategy is employed in a given language, as islands effects are expected to arise from dependencies created by movement but not by base generation. The body of cross-linguistic research on resumption has shown that the situation is more complicated once other diagnostics are taken into account, as languages can have mixed resumption profiles. In this paper, we discuss resumption in Ā-dependencies in Akan, a Kwa language spoken in Ghana, and illustrate that, despite their general insensitivity to islands, resumptive dependencies also show many classic hallmarks of movement. We situate these findings in the broader context of a general understanding of resumption cross-linguistically and discuss how the conflicting diagnostics might be reconciled with a movement-based analysis.

1. Introduction

Since their discovery by Ross (1967), syntactic islands have played a central role in research on long-distance dependencies in natural language. Whether or not an island boundary may intervene between the head and tail of a dependency is arguably the main diagnostic used to argue that a particular dependency involves movement. While this is most often applied to filler-gap dependencies, as research on long-distance dependencies has taken data from a wider range of languages into account, it became clear that many languages have productive strategies for forming long-distance dependencies that terminate in a pronoun rather than a gap (e.g., Borer 1984, Koopman and Sportiche 1986, McCloskey 1979). For this reason, diagnostics such as island-sensitivity are crucial in determining the nature of resumptive dependencies, as it is not a priori clear whether they involve movement or base generation. While it initially appeared that there may be a neat dichotomy between resumptive dependencies created by movement and those created by base generation and binding, subsequent research has shown that the cross-linguistic picture is, in fact, far more complicated and that the presence or absence of island effects with resumption does not always neatly align with other diagnostics for movement (see, e.g., Salzmann 2017b and Hewett 2023). Reconciling these conflicting diagnostics is the current major challenge facing analyses of resumptive Ā-dependencies.

In this paper, we will discuss how Akan, a Kwa language spoken in Ghana, fits into the empirical landscape of resumption. Building on previous work (Korsah and Murphy 2020), we argue that Akan has a rather unique profile among languages employing a resumptive strategy to form Ā-dependencies. Despite the general absence of island effects with resumption in Akan, we will show that there are several other diagnostics that have been traditionally assumed to point to a movement derivation of resumption, such as crossover effects, cyclicity effects, and reconstruction effects, in addition to some language-specific arguments. In what follows, we aim to provide a detailed picture of the empirical situation in Akan and discuss its broader theoretical consequences for the theory of resumption and island sensitivity. While several of the discussed diagnostics can also be accommodated under a non-movement analysis, we will show that alternative approaches such as base generation or mixed chains ultimately struggle to capture the full range of data, cycliclity effects in particular. We ultimately conclude that a movement-based derivation still offers the most promising solution to the puzzling resumption profile of Akan; however, the exact way in which resumption obviates island effects in the language is still an unresolved issue. We will discuss two potential explanations for this involving island repair by resumption (Korsah and Murphy 2020) and category-sensitive island sensitivity (Hein and Georgi 2021) and the outstanding challenges facing each approach.

2. Islands and Resumption in A-Bar Dependencies

Much of the early generative work on syntactic displacement focused on the properties of filler-gap dependencies and the restrictions imposed on them (e.g., Ross 1967). However, the idea that a long-distance dependency could also terminate in a pronoun was discussed in Ross’ seminal work, while gapped dependencies are subject to the island constraints identified by Ross, such as the adjunct island in (1a); Ross (1967, p. 433) observed that a resumptive pronoun at the tail of the dependency appears to avoid an island violation (1b).

(1)	a.	*King Kong is a movie which1 you’ll laugh yourself sick [CP if you see____1 ]
	b.	King Kong is a movie which1 you’ll laugh yourself sick [CP if you see it1]

A tempting conclusion to draw from (1) would be that island violations diagnose a movement derivation in the syntax. This would imply that the gapped dependency in (1a) involves movement, while the resumptive dependency in (1b) does not. Consequently, there would therefore be two different grammatical strategies available for forming an Ā-dependency such as the one found in relative clauses: genuine syntactic movement versus base generation and binding of a pronoun, as schematized in (2).

Subsequent work on the kind of resumption that Ross identified for English has concluded that resumption in English does not represent a genuine grammatical option in the language, but instead it is a case of ‘intrusive resumption’, e.g., an extra-grammatical repair strategy used to facilitate processing (Beltrama and Xiang 2016; Sells 1984; Shlonsky 1992; however see Ackerman et al. 2018; Kroch 1981). Cross-linguistic research, however, has uncovered that many languages do indeed have genuine grammatical strategies for forming resumptive dependencies. In fact, this option sometimes co-exists alongside the more familiar gap-formation strategy.

One such language is Irish. McCloskey (1979; 2002) shows that Ā-constructions in Irish, such as those found in relative clauses, may terminate either in a gap (3) or a resumptive (4). Importantly, only gaps are island-sensitive, therefore conforming to the aforementioned hypothesis that gapped dependencies involve movement, while resumptive dependencies involve base generation (also see Borer 1984 on Hebrew).

(3)	Gaps are island-sensitive in Irish (McCloskey 1979, p. 32)
	*bean1 nachN  bhfuil fhios agam [CP an  bpósfadh  duine  ar bith____1 ]
	woman neg.comp I know    int.prt would marry person any
	‘a woman who I do not know if anyone would marry’

(4)	Resumptives are not island-sensitive in Irish (McCloskey 1979, p. 33)
	bean1  nachN  bhfuil fhios agam [CP an   bpósfadh   duine  ar bith í____1 ]
	woman neg.comp I know    int.prt would marry person any her
	‘a woman who I do not know if anyone would marry her’

This dichotomy is further supported by the observation that other diagnostics for movement show the same distinction. So-called ‘weak crossover’ (WCO) effects, in which a pronoun bound by the moved phrase is ‘crossed’ by a dependency (i.e., is not c-commanded by the tail of the dependency), are often taken to be a hallmark of (Ā-)movement (see Section 3.4 for further discussion). In Irish, only gapped Ā-constructions show WCO effects (5).

(5)	No WCO with resumptives in Irish (McCloskey 2011, p. 110)
	a. *[DP fear1 [CP a d’fhág [a1 bhean]____1]]
	man aL left his wife
	‘the man who their wife left’
	b. [DP fear1 [CP a d’fhág [a1 bhean] é1  ]]
	man  aN left his wife them
	‘the man who their wife left him’

The consistent patterning together of (in)sensitivity to islands and other movement diagnostics has been observed in other languages, too. In Vata, a Kru language spoken in Côte d’Ivoire, Koopman and Sportiche (1982, 1986) illustrate that resumptive dependencies show both sensitivity to islands (6) and weak crossover effects (7).

(6)	Resumptives are island-sensitive in Vata (Koopman and Sportiche 1986, p. 370)
	*Álɔ̍1 ǹ   nyla̍ nynı̍ [CP nā   ɔ̍1  dı̍  mέ ] la̍?
	who you wonder  comp 3sg cut it    wh
	‘Who do you wonder whether he cut it?’

(7)	WCO with resumptives in Vata (Koopman and Sportiche 1982, p. 143)
	*Álɔ1 [ɔ̍1 nɔ̍]   gùgù [CP nā   ɔ̍1   mlì ] la̍?
	who his mother think  comp 3sg left   wh
	‘Who did their mother think left?’

Given this state of affairs, one might also expect to find other putative movement diagnostics patterning together with island sensitivity. Various kinds of reconstruction effects may be taken as evidence for a movement derivation since, taken at face value, a movement-generated resumptive implies that the resumed phrase is in some sense syntactically represented at the tail of the dependency, while a simple base generation and binding strategy may not.1

With this in mind, we find some supporting evidence of this correlation from deto-relative clauses in Bulgarian. As Krapova (2010) shows, movement gaps are not possible inside complex NP islands, while overt resumptive pronouns are (8).

(8)	Resumptives are not island-sensitive in Bulgariandeto-relatives (Krapova 2010, p. 1250)
	Tova e [DP edin film1 [CP deto [DP vsicki [CP koito sa go1   / *____1 gledali ]]
	this is  a  film   that  all     who.3pl are it.cl.acc    seen mnogo go   xaresvat]]
	a.lot   it.cl.acc like.3pl
	‘This is a film that all those who have seen it like it a lot.’

In addition to not exhibiting WCO effects (Krapova 2010, p. 1250, fn.16), resumptive dependencies in Bulgarian deto-relatives fail to exhibit another apparent effect of movement, namely allowing for reconstruction to the position of the resumptive. Since the head of the relative clause contains a pronoun bound by a quantifier inside the relative clause, it would have to reconstruct to its base position for purposes of interpretation. As (9) shows, this is possible with a gap but not with a resumptive pronoun.

(9)	No reconstruction to position of resumptive (Krapova 2010, pp. 147–1248)
	[DP [snimkata  na deteto  si1]2  [CP deto vsjaka majka1 2 / *ja2
	picture.the of child.the her.refl    that every mother    her.cl.acc
	nosi   vs. portmoneto si    ]]
	carry.3sg in  purse.the  her.refl
	‘the picture of her1 child that every mother1 carries (it) in her purse’

Given that reconstruction is often taken to be a defining property of movement dependencies, we would expect to find reconstruction effects with island-sensitive resumption. There are actually surprisingly few clear-cut cases of this, however. Perhaps the only potential example we are aware of comes from Welsh where both resumptive and gapped dependencies into relative clauses are ungrammatical, indicating some sensitivity to strong islands.2

(10)	Resumption in Welsh is island-sensitive (Borsley et al. 2007, p. 148)
	*Dyma ’r  ffenest1 darais   i [DP ’r bachgen [CP dorrodd   hi1 ddoe   ]]
	that.is the window hit.pst.1sg I    the boy    break.pst.3sg it yesterday
	‘That is the window that I hit the boy who broke (it) yesterday.’

According to Rouveret (2008), Welsh allows for reconstruction for variable binding to the position of the resumptive, as in (11). Taken at face value, this would seem to imply that a different kind of derivation is involved than with resumption in Bulgarian deto-relatives, for example.3

(11)	Reconstruction to position of resumptive in Welsh (Rouveret 2008, p. 182)
	Mae gan Siôn [farn   ar ei1   lyfr]2 y mae pob awdur1 yn   ei2 pharchu
	is   with Siôn opinion about their book C is   each author prog it respect
	‘Siôn has an opinion about their book that each author respects.’

So far, a reasonably neat picture emerges. On the one hand, we have languages that have what we might call a ‘Type I’ resumption profile, where island insensitivity correlates with a lack of other positive diagnostics for movement such as crossover effects or reconstruction (e.g., Irish and Bulgarian). In addition to these, we have a ‘Type II’ profile, where island sensitivity aligns with other potential movement diagnostics. Vata, and perhaps Welsh, is an example of this kind of language. Both of these can be seen in the table below.

(12) Types of resumption profiles (preliminary)
	Is resumption island-sensitive?	Does resumption have properties of movement?
Type I (e.g., Irish, Bulgarian)	✗	✗
Type II (e.g., Vata, Welsh?)	✓	✓
Type III?	✓	✗
Type IV?	✗	✓

At this point, one might wonder if there are languages with ‘mismatched’ resumption profiles, e.g., a ‘Type III’ language in which we find island-sensitivity but otherwise a lack of evidence for movement. Or perhaps the reverse case in which there is evidence for movement in spite of a lack of island effects with resumption. Determining whether such profiles truly exist is a difficult and involved matter. For one thing, we will see that languages can show conflicting results for the various diagnostics for movement (potentially casting some doubt on their validity as such). Nevertheless, a detailed comparison of the varying properties of resumptive dependencies does reveal some interesting mixed cases (also see Schurr et al. 2024 on conflicting evidence from islands and other movement diagnostics for non-resumptive Ā-dependencies in Shupamem).

A potential Type III language could be Scottish Gaelic. Following the description in Adger and Ramchand (2005), Ā-dependencies in the language are island-sensitive.

(13)	Ā-dependencies in Scottish Gaelic are island-sensitive (Adger and Ramchand 2005, p. 178)
	*am fear1   a phòg mi [DP a’bhean   [CP a   phòs   pro1 ]]
	the man c.rel kissed I    the.woman    c.rel married
	‘the man who I kissed the woman who married’

The tail of the dependency in (13) is not occupied by an overt resumptive pronoun, from which one might want to conclude that there is movement rather than resumption here. However, Adger and Ramchand (2005, pp. 167–171) provide a number of empirical challenges for a trace-based account, e.g., from non-identity effects, where the gap position fails to exhibit the kind of connectivity effects we would expect from movement. Examples of this include selectional mismatches between the ‘extractee’ and its base position, the lack of definiteness agreement on adpositions adjacent to Ā-gaps and the absence of case connectivity effects with participles. In addition to this, Adger and Ramchand (2005) point out that we find a lack of reconstruction effects, for example, for Principle C. The example in (14) is grammatical, which would not be expected if there was reconstruction of the wh-phrase.

(14)	No reconstruction for Principle C in Scottish Gaelic (Adger and Ramchand 2005, p. 171)
	[Dè   an  dealbh de dh’Iain1]2 a   cheannaich e1 pro2 an de
	what the picture of Iain   c.rel bought   he    yesterday
	‘Which picture of Iain did he buy yesterday?’

While Adger and Ramchand (2005) reject a movement-based account in favour of Agree, it still shows some of the hallmarks of a Type III profile. Ultimately, though, this is perhaps not the clearest example of a potential Type III resumption profile, as it still requires the assumption of a null resumptive pronoun, which can, in practice, be difficult to distinguish from a genuine gap4. Furthermore, Adger and Ramchand (2005) report that we find the kind of morphological cyclicity effects that are normally assumed to be typical of movement. We return to this issue in Section 4.

Finally, we may ask whether there is a different kind of mismatched profile, i.e., a Type IV language. This would be a language with island-insensitive resumption that nevertheless shows evidence of movement. As we will see, despite the fact that the empirical picture is generally somewhat murky and incomplete, we believe that the resumption profile of Akan currently makes the strongest case for a Type IV profile, and while resumption with nominal extractees is typically impervious to islands in Akan, it nevertheless shows other effects of what are typically considered to be hallmarks of movement. This breaks the neat dichotomy that was originally hinted at by languages such as Irish and Vata.

In the remainder of this paper, we will lay out the empirical landscape of resumption in Akan and, building on Korsah and Murphy (2020), attempt to establish the full resumption profile of Akan. We will argue that Ā-dependencies in Akan are island-insensitive and yet still derived by movement. In rejecting alternative approaches involving base generation and mixed chains, we will discuss two potential ways in which one can reconcile a movement derivation with island insensitivity.

3. Resumption and Ā-dependencies in Akan

Akan has two main kinds of Ā-constructions that will be relevant for the following discussion. The first is the so-called na-focus construction which is used for ex situ wh-questions, such as (15), in addition to focalization of some argument or adjunct in a declarative clause.5 The structure we assume for the na-focus construction is given below.

The second major construction of interest to us is the relative clause. As can be seen in (16), relative clauses in Akan contain the invariant relative marker áa and also contain the so-called ‘clausal determiner’ (CD) nó, which has the same form as regular determiners (on clausal determiners, see Korsah 2017 and Kandybowicz and Torrence 2019). As the structure below illustrates, we assume that the CD is adjoined to the relative CP which is in turn extraposed within the DP.6

Going back to the earliest work on wh-movement and focus in Akan (e.g., Boadi 1974; Saah 1988, 1994; Schachter 1973), it was noticed that Ā-dependencies must terminate in a resumptive pronoun if the displaced phrase is animate. Taking the ditransitive sentence in (17a) as our baseline, we observe that wh-movement of the animate indirect object leads to an obligatory resumptive pronoun nó in the base position (17b).7 If we instead move the inanimate direct object, then no pronoun may appear in this position (17c).

(17)	a.	Yaw ma-a Saka siká.
		Yaw give-pst Saka money
		‘Yaw gave Saka money.’
	b.	Hwáń1 na   Yaw má-a   {*_____1 / no1 } sika?
		Who foc Yaw give-pst    3sg money
		’Who did Yaw give money to?’
	c.	Déέn1 na Yaw má-a   Saka { _____1 / *no1 }?
		what foc Yaw give-pst Saka     3sg
		‘What did Yaw give to Saka?’

At first glance, this may suggest that Akan has two distinct ways of forming Ā-dependencies, a gap strategy and a resumptive strategy. This initial impression is misleading, however. As shown in Korsah (2017), the gaps we find with the extraction of inanimates are best viewed as the result of a more general rule of ‘pro drop’ that typically blocks the overt realization of inanimate object pronouns (both anaphoric and resumptive). The reason for this thinking comes from three contexts in which inanimate pronouns are obligatorily pronounced. These are (i) preceding clause-final adverbs (Saah 1994), (ii) the object of change-of-state verbs (Boadi 1971; Osam 1996), and (iii) the subject of secondary predicates (Korsah 2017). As each of the examples in (18) through (20) show, gaps and resumptives alternate even in positions associated with an inanimate extractee.

(18)	Inanimate resumption with clause-final adverb
	a.  Aduane nó1 na Kofí pέ (*no1)
	food   def foc Kofi like    3sg
	‘It’s the food that Kofi likes.’
	b. Aduane nó1 na Kofí pέ *(no1) anɔpá
	food def foc Kofi like 3sg morning
	‘It’s the food that Kofi likes in the morning.’

(19)	Inanimate resumption with change-of-state verbs
	a.  Akonwa nó1 na  Kofi kŕá-a   (*no1)
	chair   def foc Kofi import-pst 3sg
	‘It’s the chair that Kofi imported.’
	b.  Akonwa nó1 na Kofi bú-u *(no1).
	chair def foc Kofi break-pst 3sg
	‘It’s the chair that Kofi broke.’

(20)	Inanimate resumption with secondary predication
	a.  Aduane nó1 na   Kofí pέ  [SC *(no1) hyehyééhyé ]
	food  def foc Kofi like    3sg very.hot
	‘It’s the food that Kofi likes very hot.’
	b.  [DP Aduane nó1 [CP áa Kofí pέ *(no1) hyehyééhyé nó ]] nie
	food  def   rel Kofi like  3sg very.hot  cd   this
	‘This is the food that Kofi likes very hot.’

Given these observations, Korsah (2017) and Korsah and Murphy (2020) conclude that both na-focus constructions and relative clauses always trigger a resumptive pronoun, which subsequently undergoes pro drop if the extractee is inanimate. The three contexts above are exceptions to the inanimate pro drop rule. What exactly singles out these contexts as exempt from the pro drop rule is an interesting question that need not immediately concern us here (see Korsah 2017 for a possible analysis involving object shift).

The final thing to mention about resumption in Akan pertains to the nature of subject vs. object resumption. As we have seen above, object extraction leads to a resumptive pronoun nó in the base position of the object. Subject pronouns, both resumptive and anaphoric, appear as a bound morpheme on the verb. This is shown in (21) for both anaphoric pronouns (21a) and resumptive pronouns (21b).

Syntactically, we assume that Ā-dependencies with subjects involve a null resumptive in subject position. We treat the overt morpheme on the verb as an agreement marker cross-referencing the features of the null resumptive pronoun (22). Some supporting evidence for this comes the observation in Korsah (2017) that subject resumptives, unlike object resumptives, undergo an optional alternation with the inanimate pronoun E-. This can be understood in terms of (optional) anti-agreement, which is already a well-known reflex of subject extraction (see e.g., Baier 2018; Henderson 2013; Ouhalla 1993; Schneider-Zioga 2007).

(23)	Anti-agreement with local subject extraction (Korsah 2017, p. 118, 121)
	a. Kofi1 na  ɛ/ɔ1-káń-n   kŕataá nó
	Kofi  foc 3-/3sg.sbj-read-pst book  def
	‘It is Kofi who read the book.’
	b. [DP Abɔfrá nó1 [CP aá ɛ/ɔ1-káń-n    kŕataá nó ]] nie
	child  def  rel 3-/3sg.sbj-read-pst book  def   this
	‘This is the child who read the book.’

The idea that this is anti-agreement is supported by the fact that the unagreeing E- marker is not possible under long-distance extraction (Korsah 2017, p. 120). This is entirely parallel to what we find with anti-agreement in Berber, for example (Ouhalla 1993).

3.1. Island Effects

With these preliminary observations about Ā-constructions and resumptive pronouns in place, we can now turn to island effects in Akan. As has been established in much previous research, island effects are absent with extraction of both animate and inanimate noun phrases (Goodluck et al. 1995; Saah 1994; Saah and Goodluck 1995). This is irrespective of whether the dependency terminates in a resumptive or a gap. This is illustrated below for four kinds of syntactic islands, namely complex NP islands, wh-islands, adjunct islands and sentential subject islands, for both resumptive dependencies (24) and gapped dependencies (25), respectively.

(24)	Island insensitivity with resumptives (Saah 1994, p. 172; Korsah 2017, p. 117)
	a. Hwáń1 na  wo-hú-u  [DP onípá ko [CP áa ɔ-bɔ̍-ɔ   no1 no ]] ?
	who  foc 2sg-see-pst    person def    rel 3sg-hit-pst 3sg cd
	‘Who did you see the person who hit him?’    (CNP island)
	b.  Ám!má1 na Kofí bísá-a  [CP sέ hwán na ɛ-dɔ̍   nó1 nó ]
	Amma foc Kofi ask-pst    that who foc 3sg-love 3sg cd
	‘It is Ama who Kofi asked who loves her.’   (wh-island)
	c.  Ám!má1 na Yaw ré-sú   [CP ésánesέ Kofi dɔ̍   nó1 nó ]
	Ama  foc Yaw prog-cry    because Kofi love 3sg cd
	‘It is Ama that Yaw is crying because Kofi loves her.’.’  (adjunct island)
	d. Hwáń1 na  [CP sέ  Kofi dɔ̍  nó1 nó ] á-ma   abusuá mmienú nó
	Who  foc   that K  love 3sg cd   perf-give family two   def   á-bóḿ.   perf-reconcile
	‘Who is such that [that Kofi loves her] has made the two families reconcile.’
	(sentential subject)

(25)	Island insensitivity with gaps (Saah 1994, p. 172, 197)
	a. Déέn na wo-níḿ  [DP onipa  ko  [CP áa ɔ-tɔ̍-ɔ-έ   ____ 1 nó ]] ?
	who foc 2sg-know    person def    rel 3sg-buy-pst-ye    cd
	‘What do you know the person that bought?’   (CNP island)
	‘What do you know the person that bought?’   (CNP island)
	b.  Deɛn1 na Mary bisa-a [CP sέ hwán na ɔ-yɛ-e   ____1 nó ] ?
	what foc Mary ask-pst    that who foc 3sg-make-pst    cd
	‘What did Mary ask who made?’    (Wh-island)
	c.   Siká1  na  Yaw ré-sú  [CP ésánesέ Kofi dɔ̍ ____ 1 nó ]
	money foc Yaw prog-cry    because Kofi love    cd
	‘It is money that Yaw is crying because Kofi loves.’     (Adjunct island)
	d. Déέn1 na [CP sέ Kofi dɔ̍ ____1 nó ] á-ma   abusuá mmienú nó
	what foc that Kofi love    cd    perf-give family two def
	á-bóḿ
	perf-reconcile
	‘What is such that [that Kofi loves (it)] has made the two families reconcile.’
	(sentential subject)

Given the fact that there appears to be no obvious difference in acceptability between displacement associated with a gap or a resumptive with regard to islands, Saah (1994) proposed that both grammatical strategies are instances of base generation, with binding of a null pro in the case of a gap (cf. the Scottish Gaelic example in (13)).

The situation regarding island effects is more complicated, however. Korsah and Murphy (2020) point out that the extraction of certain kinds of categories does appear to show island sensitivity. First, consider the fact that PPs can undergo long extraction out of an embedded CP:

(26)	Long-distance PP extraction is possible
	[PP Akonwá nó mú ] na Ama níḿ   [CP sɛ   Kofí dá____PP ]
	chair def in foc Ama know that Kofi lie
	‘Ama knows that Kofi lies in the chair.’

When this CP forms part of an island, e.g., a complex NP island (27b) or a wh-island (27d), then extraction is impossible.

(27)	PP extraction is island-sensitive
	a. Amma nim  [DP neá ńtí   [CP áa  Kofi dá [PP akonwá nó   mú ] ]]   Ama know    thing because    rel Kofi lie    chair def in
	‘Ama knows the reason why Kofi lies in the chair.’
	b. *[PP Akonwá nó mú ] na Ama níḿ [DP neá ńtí [CP áa Kofi dá   chair   def in foc Ama know   thing because    rel Kofi lie   ____PP ]]  ‘Ama knows the reason why Kofi lies in the chair.’    (Complex NP island)
	c. Amma bisả-a  [CP sέ   bré   bέn na  Kofi dả [PP akonwá nó mú ]]
	Ama  ask-pst  that time q  foc  Kofi lie    chair def in
	‘Ama asked when Kofi lies in the chair.’
	d. *[PP Akonwá nó mú] na   A!ma bỉsả-a [CP sέ  bré  bέn na  Kofi dả____PP ]
	chair def in foc Ama ask-pst   that time q foc Kofi lie
	‘Ama asked when Kofi lies in the chair.’       (wh-island)

The same pattern is also found for VP movement, as Hein (2017) has shown. Hein points out that VP extraction may proceed as long-distance extraction:

(28)	Long-distance VP extraction is possible (Hein 2017, p. 9)
	[VP dán  sí-é    ] na Ám!má ká-a   [CP sέ Kofí á-yɔ̍  ____VP ]
	house build-nmlz    foc Ama   say-pst    that Kofi perf-do
	‘Ama said that Kofi built a house (not bought a car)’

Nevertheless, it shows sensitivity to the same kinds of islands as PP movement does (29).

(29)	VP fronting is island-sensitive (Hein 2017, p. 10)
	a. ?*[VP dán sí-é    ] na   mé-ń-té-e    [DP atétésέm bíárá [CP sέ   house build-nmlz foc 1sg-neg-hear-pst   rumour.pl any    that  Kofí á-yÓ  ____VP ]]  Kofi prf-do  ‘I did not hear any rumours that Kofi has built a house.’   (Complex NP island)
	b. ?*[VP dán  sí-é   ] na Ám!má bísá-a  [CP sɛ   dabέn na  Kofí yɔ̍-ɔ-έ   house build-nmlz   foc Ama ask-pst    that when foc Kofi do-pst-ye ____VP ]
	‘Ama asked when Kofi built a house.’     (Wh-island)

The conclusion that Korsah and Murphy (2020) draw from this is that sensitivity to islands is linked to the possibility of (covert) resumption. Neither PPs nor VPs allow for resumptive pronouns.8 This can be seen for PP extraction in (30b). Even though the tail of the Ā-dependency is in a context that should block pro drop (a clause-final adverb), no resumptive is possible here. The impossibility of a resumptive (even one that has undergone pro drop) is directly linked to the re-emergence of island effects according to Korsah and Murphy (2020).

(30)	Extracted PPs lack resumptives
	a. Kofí da [PP akonwá nó mú ]
	Kofi lie    chair   def in
	‘Kofi is lying in the chair.’
	b. [PP Akonwá nó mú ] na Kofí dá { ____PP / *hɔ   } anɔpá.
	chair def in    foc Kofi lie  there    morning
	‘Kofi lies in the chair in the morning.’

An important complication to this otherwise straightforward picture was raised by Hein and Georgi (2021). They note that non-referential NPs of various kinds do not allow for overt resumptives, similar to PPs and VPs. They provide data showing that idiom chunks, predicate nominals and non-specific indefinites do not leave resumptives when extracted, even in those contexts where the pro drop is otherwise blocked.

The example that Hein and Georgi (2021) provide for a predicate nominal is given in (31). Despite the presence of a clause-final adverb in B’s answer, no overt resumptive is possible.

(31)	Context: Kofi is about to graduate this year. Kwame claims that
	A: Kofi bɛ-yɛ  dɔkota afe  yi
	Kofi fut-be doctor year this
	‘Kofi will become a doctor this year.’
	However, Ama knows that this is not correct and says
	B: Tíkyani na   Kofi bɛ-yɛ  {____1 / *no1 } afe yi
	teacher foc Kofi fut-be    *3sg.obj year this
	‘It’s a teacher that Kofi will become this year.’    (Hein and Georgi 2021)

The example in (32) shows the same effect for extraction of a non-specific indefinite.

(32)	Context: You are a new student at a school and tell a classmate that you are planning to rent a school uniform instead of buying one. However, you do not know if that is possible. Your classmate asks
	A: Wo-be-bisa headmaster no?
	2sg-fut-ask headmaster def
	‘Will you ask the headmaster?’
	However, you did not want to bother the headmaster with this so you say
	B: Daabi, ɔkyerɛkyerɛni1 na me-be-bisa   {____1 / ??no1   } kane
	no, teacher   foc 1sg-fut-ask   3sg.obj first
	‘No, I will ask a (random) teacher first.’ (one of many teachers around)		(Hein and Georgi 2021)

Due to the impossibility of an overt resumptive here, one would expect displacement of these nominals to show sensitivity to islands like PP or VP extraction does. As Hein and Georgi (2021) show, however, these extractions are insensitive to islands (33). Only a gap in the extracted position is licit, any kind of resumptive is still impossible.

(33)	a. Tíkya1 na m-á-té   [DP atésέm nó [CP sέ Kofi bέ-yέ {____1 /
	teacher foc 1sg-perf-hear    rumour def   that Kofi fut-be
	*nó1    } afe yí   ]]
	*3sg.obj   year this
	‘It is a teacher that I heard the rumour that Kofi will become this year.’
	b. Npípa na wo-té-e   [DP atésέm nó [CP sέ Kofi súró {____1 /
	person foc 2sg.sbj-hear-pst    rumour def    that Kofi fear
	*nó1    / *wón1 } páa ]]
	3sg.obj    3sg.pl really
	‘It’s people that I have heard the rumour that Kofi really fears.’
	(Hein and Georgi 2021)

This seems to go against the empirical generalization in Korsah and Murphy (2020) that whether a particular resumptive dependency is insensitive to islands in Akan correlates directly with the possibility of having an overt resumptive pronoun in that dependency. We address this issue in some detail in Section 4.2.2.

With the discussion of island sensitivity in Akan in place, we now turn to other potential diagnostics for movement. In the following sections, we will discuss three in particular: cyclicity effects (Section 3.2), reconstruction effects (Section 3.3) and crossover effects (Section 3.4). In addition, Section 3.5 will present a language-internal argument against treating all kinds of Ā-movement in Akan as the base generation.

3.2. Cyclicity Effects

The first diagnostic test to discuss comes from so-called ‘cyclicity effects’. As we will show, Akan shows a robust cyclicity effect of tonal overwriting that also persists into islands. This will lead to two possible conclusions: (i) tonal overwriting does not diagnose movement and (ii) movement out of islands is possible. Ultimately, we will argue in favour of the latter conclusion in Section 4.2.1, given the difficulties associated with maintaining a base generation analysis.

Since Chomsky (1973), it has been assumed that movement-based Ā-dependencies are subject to locality constraints. This means that long-distance dependencies must be broken up into a series of smaller movement chains. In their most recent instantiation, these locality constraints take the form of phases. If we follow Chomsky (2000, 2001), both vP and CP constitute phases and therefore require a long-distance dependency to make an intermediate ‘stop-over’ at each of these projections, as in (34).

It has been shown that there are various different kinds of evidence supporting this derivation (see, e.g., Georgi 2014, van Urk 2020), with perhaps the most compelling evidence coming from languages which show dedicated morphological reflexes of movement taking place within that clause (see e.g., Chung 1982; Georgi 2014, 2017; Lahne 2008; Schneider-Zioga 1995). These reflexes can be on C (see e.g., Georgi 2017; McCloskey 2002) or they can be on the verb, i.e., in v (e.g., van Urk and Richards 2015). Presumably, while movement dependencies are forced to comprise a series of smaller connected steps, morphological cyclicity effects could be taken to diagnose movement over resumption, since variable binding is typically not subject to locality constraints.9

A potentially revealing example of a morphological cyclicity effect comes from Defaka. As Bennett et al. (2012) argue, the suffix -ke in Defaka surfaces on verbs in clauses in which there has been overt Ā-movement (35b). Importantly, when something is extracted from an embedded clause, as in (35c), the -ke morpheme is found on both the embedded and the matrix verb. It therefore seems to mark the path of movement, assuming that movement applies cyclically with an intermediate stopover in the embedded clause.

(35) Morphological reflex of movement in Defaka (Bennett et al. 2012, pp. 296–297)

Korsah and Murphy (2020) argue that the Asante Twi dialect of Akan shows a similar cyclicity effect in the realm of tone. Asante Twi has a process of high-tone overwriting on verbs in certain Ā-constructions. This process was originally noticed by Schachter and Fromkin (1968) and has subsequently been corroborated in further research (e.g., Fiedler and Schwarz 2005; Genzel 2013; Marfo 2005). As the following example shows, the underlyingly low verb wO (‘be’) in (36a) receives a high tone in the corresponding focus construction in (36b).

(36)	a. Kofí wɔ έnyirési
	Kofi be England
	‘Kofi is in England.’
	b. Kofí1 na ɔ1-wɔ̍    έnyirési
	Kofi foc 3sg.sbj-be England
	‘It is Kofi who is in England.’    (Schachter and Fromkin 1968, p. 209)

Korsah and Murphy (2020) point out that tonal overwriting has the signature of a reflex of successive cyclicity. In particular, it applies to all verbs that are along the path of movement. Importantly, this effect is also found with an overt resumptive pronoun. This is the case for both focus constructions (37) and relative clauses (38).

(37) Overwriting affects all verbs in a long-distance dependency (Korsah and Murphy 2020)

(38) Long distance relativization shows movement reflex

a. Me-nim  [CP sέ  óbíárá  á-te   [CP sέ   Kofi á-ka   [CP sέ    1sg-know   that everybody perf-hear    that Kofi perf-say    that     ɔ-dɔ   ɔbáá nó   ]]]    3sg.sbj-love woman def    ‘I know that everybody has heard that Kofi has said that he loves the woman.’

b. Me-hu-u [DP Obáá nó [CP Op1 áa óbíárá á-té [CP sɛ Kofi    1sg-see-pst woman def rel everybody hear-pst that Kofi    ả-ká [CP sɛ ɔ-dɔ̍ nó1 nó ]]]]    perf-say that 3sg.sbj-fut-love 3sg.obj cd    I saw the woman whom everybody has heard that Kofi has said that he loves her.’

While the unbounded nature of the process may seem to provide a convincing argument for movement, it is important to establish that movement reflexes are also found inside islands. This is because it is possible that there could be a movement dependency up to the edge of the island boundary with a non-movement dependency reaching into the island itself (see Section 4.2.1). We still find movement reflexes inside islands, as shown in (39). This supports the idea that whatever the source of this cyclicity effect is, it does not distinguish between dependencies into islands and non-islands. If it is the result of successive-cyclic movement via vP, as Korsah and Murphy (2020) argue, then we must conclude that such movement is possible out of an island (see Section 4.2.2 for further discussion). The other possible alternative, which we will ultimately dismiss, is that tonal overwriting does not diagnose movement after all. We will argue that there are several difficulties associated with maintaining this assumption, however (see Section 4.2.1).

3.3. Reconstruction Effects

Now, let us turn to reconstruction effects. Reconstruction effects have traditionally been assumed to imply movement, as a representation of the moved item is assumed to be present at the tail of a movement dependency (especially given the Copy Theory of Movement; Chomsky 1995), while this is not necessarily the case under base generation. As we will see in Section 4, however, whether reconstruction effects are themselves indicative of movement is still controversial and should therefore be treated with caution.

Recall from the examples discussed in Section 2 that there are languages with island-insensitive resumption like Bulgarian which allow the c-command requirement for variable binding to be satisfied by a moved element containing a bound variable only if the dependency terminates in a gap rather than a pronoun, as in (40).

(40)	No reconstruction to position of resumptive (Krapova 2010, pp. 1247–1248)
	[DP [snimkata na deteto si2]1  [CP deto vsjaka majka2____1 / *ja1 picture.the of child.the her.refl    that every mother    her.cl.acc nosi   vs. portmoneto si  ]]
	carry.3sg in purse.the   her.refl
	‘the picture of her1 child that every mother1 carries (it) in her purse’

On the other hand, there are languages such as Lebanese Arabic, as discussed by Aoun and Benmamoun (1998), which allow reconstruction for variable binding at the position of an overt resumptive pronoun (also, see Cinque 1977 for evidence for reconstruction in French left dislocation structures). In (41a), the extracted phrase containing a bound pronoun may be interpreted by reconstructing it to its base position, i.e., interpreting the extracted phrase as if it occupied that position. As (41b) indicates, the surface position of the resumptive pronoun must be c-commanded by the binder in order for the sentence to be grammatical, further supporting the idea that there is indeed reconstruction to this position.

(41)	Reconstruction to resumptive in Lebanese Arabic (Aoun and Benmamoun 1998, p. 581)
	a. [Mʕalləmt-o1]2 fakkarto  Ɂənno kəll  walad1 ʕaṭee  -ha2 hdiyye
	teacher.f-his thought.2sg that every boy gave.3s -her gift
	b. *[Mʕalləmt-o1]2 fakkarto   Ɂənno ʕaṭee -ha2 kəll walad1 hdiyye
	teacher.f-his thought.2sg that   gave.3s -her every boy   gift
	‘His teacher, you thought that every boy gave her a gift.’

In Akan, we see a similar effect. A moved phrase containing a bound variable reconstructs to the position of the pronoun in order to license variable binding, as can be seen in (42b).

(42)	Reconstruction for variable binding in Akan
	a. Abán   bíárá1 dwéne [DP ne1-máńfóɔ̍ yíe-yɔ̍ hó ] dáá
	government every think   poss-people well-be self   every day
	‘Every1 government thinks about the well-being of its1 people every day.’
	b. [DP ne1-máńfóÓ yíe-yÓ hó ]2 na abán   bíárá1 dwéné no2
	poss-people well-be self    foc government every think 3sg.obj
	dáá
	every day
	It’s the well-being of its1 people that every1 government thinks about every day.’

When testing for reconstruction, it is important to control for various other factors. For example, Schneider-Zioga (2009) notes that reconstruction for variable binding in Kinande is possible for local extraction but not for long-distance extraction. This leads her to propose that long-distance extraction in Kinande is in fact not formed by movement (see Section 4.2.1). In Akan, however, we also find reconstruction effects with long-distance movement (43), meaning that we do not have to be concerned about this potential complication.

(43)	Reconstruction across clause boundary
	[DP Ne1-máńfóɔ̍ yíe-yɔ̍ hó ]2 na Kofí níḿ [CP sɛ abán   bíárá1 dwéné
	poss-people well-be self    foc Kofi know    that government every think
	no   2 dáá    ]
	3sg.obj every day
	‘It’s the well-being of its1 people that Kofi knows that every1 government thinks about
	every day.’

There is yet another potential confounding factor, however. Aoun et al. (2001) demonstrate that, in Lebanese Arabic, there are actually two available strategies for resumption: base generation and movement (or ‘true’ vs. ‘apparent’ resumption in their words). They argue that reconstruction effects are only found with resumptives generated by a movement derivation. This option is blocked, however, when the resumptive is located inside a strong island. In this context, the base generation strategy must instead be used, leading to the lack of reconstruction into islands. Akan, however, does not share this property. As (44) illustrates, reconstruction for variable binding persists into islands.

(44)	Reconstruction for variable binding into an island
	[DP Ne1-máńfóɔ̍   yíe-yɔ̍ hó ]2 na m-á-té    [DP atésέḿ bí      [CP sɛ
	3sg.poss-people well-be self    foc1sg-perf-hear    rumour indef   that
	abán   bíárá1 dwéné no2   dáá  ]]
	government every think   3sg.obj every day
	‘It’s the well-being of its people that I have heard a rumour that every government thinks about everyday.’
	thinks about everyday.’

The preceding examples also show reconstruction to the position of a resumptive pronoun for the purposes of variable binding. We also find reconstruction effects for other kinds of binding. For example, the anaphor hó is subject to Principle A (Saah 1989). As (45a) shows, it must be bound by the closest c-commanding antecedent within its local clause. Under long-distance movement, the anaphor may only refer to the embedded antecedent, thereby indicating reconstruction (45b).10

(45)	Reconstruction for Principle A with resumption
	a. Kofí1 dwene [CP sέ Ám!máj bɛ-pírá [DP ne hój/*i ]]
	Kofi think    that Ama   fut-hurt   3sg.obj refl
	‘Kofii think that Amaj will hurt herselfj/*i’
	b. [DP Ne  hój/*i ]1 na Kofíi dwéné [CP sɛ Ám!máj bɛ-pírá no1   Okyena  ]
	possrefl  foc Kofi think    that Ama fut-hurt 3sg.obj tomorrow
	‘It is herselfj that Kofi thinks that Amaj will hurt tomorrow.’

Similarly, Akan shows Principle C effects in non-movement constructions, as (46a) shows. When a phrase containing an R-expression is extracted, we still observe a Principle C effect with respect to the position of the resumptive pronoun (46b).

(46)	Reconstruction for Principle C with resumption
	a. *ɔ1-p [DP Kofí1 m̉fỏnỉrỉ yỉ]
	3sg-like Kofi picture this
	‘He1 likes the picture of Kofi1’
	b. *[DP Kofí1 m̉fỏnỉrỉ yỉ ]2 na Ám!má níḿ [CP sɛ ɔ1-pέ no2   paa   ]
	Kofi picture this foc Ama  think   that 3sg-like 3sg really
	‘It’s this picture of Kofi1 that Ama thinks he1 really likes’

As the two examples in (47) illustrate, reconstruction for Principle C is also found with resumption into islands.11

(47)	a. Principle C reconstruction into complex NP island
	[DP Ám!mái adúané nó ]1 na m-á-té    [DP atetésέm bí   [CP sέ
	Amma food def foc 1sg-perf-hear    rumour indef    that
	ɔ̍*i/j-dí   no1 dáá     ]]
	3sg.nom-eat 3sg every day
	‘It’s Ama’si food that I have heard a rumour that s/he*i/j eats every day.
	b. Principle C reconstruction into adjunct island
	[DP Ám!mái adúané nó ]1 na Yaw ré-sú   [CP ésánesέ ɔ̍*i/j-dí   no1
	Amma food def foc Yaw prog-eat because 3sg.nom-eat 3sg
	dáá]
	every day
	‘It’s Ama’si food that Yaw is crying because s/he*i/j eats every day.’

In addition to this, we find evidence for reconstruction with idiom chunks, too. Akan has a class of idiomatic VP constructions that are known as inherent complement verbs (ICVs) in the literature (Essegbey 1999; Korsah 2016). For example, the VP to ndwom (‘throw song’) has the non-compositional interpretation ‘to sing’ (48a) (Kandybowicz 2015, p. 266). This meaning is preserved when the complement of the ICV is focused and a resumptive pronoun fills the object position (48b).

(48)	Reconstruction for idiomatic interpretation
	a. Kofí to-o   ndwóm έnóra
	Kofi throw-pst song yesterday
	‘Kofi sang yesterday’
	b. Ndwóm1 na Kofí tó-o    no1  έnóra
	song   foc Kofi throw-pst 3sg.obj yesterday
	‘It was singing that Kofi did yesterday.’

Taken together then, it is clear that Akan shows a range of evidence of reconstruction in resumptive dependencies. As mentioned in the introduction, reconstruction effects have traditionally been taken as evidence for a movement derivation. In our contemporary understanding of resumption, however, this conclusion is arguably less straightforward, as we will discuss further in Section 4. As with other diagnostics, the presence of reconstruction alone is not a conclusive argument for movement. It must be considered alongside other potential diagnostics, too.

3.4. Crossover Effects

The final potential diagnostic for movement we will discuss is crossover effects. Crossover effects involve a configuration in which a displaced element is also co-indexed with a bound variable pronoun. A distinction is typically made between two types of crossover situations: Strong Crossover (SCO) (49) and Weak Crossover (WCO) (50).

(49)	Strong Crossover
	a. Ẉho1 does he1 love ____1 ?
	b. Who1____ 1 loves him*(self)1?

(50)	Weak Crossover
	a. *Who1 does [his1 mother] love ____ 1 ?
	b. Who1____ 1 loves [his1 mother] ?

In each of the ungrammatical cases, (49a) and (50a), the movement gap does not c-command the bound pronoun. The main difference is that in WCO configurations like (50a), the pronoun does not c-command the movement gap, unlike in SCO configurations (49a). This is relevant because it is possible that SCO effects could be subsumed under reconstruction for Principle C (see, e.g., Chomsky 1981). As we saw in the previous section, Akan exhibits reconstruction effects with resumption, so SCO would potentially not tell us anything new.

For this reason, we will focus on WCO effects. What exactly WCO effects show is still a matter of some contention. Traditionally, WCO has been used to motivate the idea that the tail of a movement dependency, whether realized as a pronoun or a gap, has a special status as a ‘syntactic variable’ (Chomsky 1981; Reinhart 1983; Safir 1984; also see Koopman and Sportiche 1982). These variables are then subject to additional constraints, including Weak Crossover. Importantly, this differs from simple bound pronouns without movement, which do not count as syntactic variables. As the following grammatical example shows, a similar binding configuration to WCO does not lead to degradation when movement is not plausibly involved:

(51)	Every boy1 told [his1 mother] that the teacher praised him1

As we have seen, this is relevant in the context of resumptive Ā-dependencies. Recall that Irish, a language with island-insensitive resumption, does not exhibit WCO effects in the relevant cases (52a), whereas island-sensitive resumption in Vata does (52b).

(52)	a. [DP fear [CP Op1 a  d’fhág [a1 bhean] é1 ]]
	man  aN left    his wife   them
	‘the man who their wife left him’                (Irish)
	b. *Àɔ̍1 [ɔ̍1 nɔ̍]   gùgù [CP nā ɔ̍1 mlì ] la̍?
	who his mother think   comp 3sg left   wh
	‘Who did their mother think left?’              (Vata)

If the resumptive pronoun in (52a) is the result of base generation and binding, similar to (51), while the resumptive in (52b) counts as a syntactic variable, then the presence/absence of WCO effects in each case is predicted.

Let us now turn to Akan. Since nominal resumption in Akan is insensitive to islands, like in Irish, we would expect not to find WCO effects, all else being equal. Korsah and Murphy (2020) provided the following examples that they claim illustrate a WCO effect in Akan. They claimed that (53a) is grammatical, while (53b) is ungrammatical (at least for some speakers). The presence of a WCO effect in (53b) was disputed by Titov (2019) and, indeed, it seems that this conclusion was too hasty. At present, we are no longer confident in our previous claim that (53b) reliably exhibits a WCO effect. The conclusion at this point would seem to be quite straightforward then—Akan lacks WCO effects, and therefore, resumption in Akan fails this potential diagnostic for movement, patterning with base-generated resumption languages like Irish.

There is a general complication with using WCO effects to test whether resumption is movement-derived, however. An example such as (53b) is structurally ambiguous in that it is not immediately clear which of the two pronouns is the resumptive pronoun (see Hewett 2023, pp. 332–334 for discussion). In principle, a WCO configuration with a resumptive pronoun can be viewed as a true crossover configuration in which the lower pronoun is the resumptive of XP, while the possessor in YP is a variable bound by XP (54). Alternatively, the possessor position could instead be analyzed as the resumptive pronoun (55), assuming that the lower pronoun can be bound from (within) a c-commanding A-position (Büring 2005, Reinhart 1983).12 In this case, there is no crossover configuration to begin with. This means that the absence of WCO effects could simply be due to the fact that there is an alternative derivation which avoids the crossover configuration entirely, providing an alternative way of satisfying the binding requirements of the non-resumptive pronoun. On this view, the absence of WCO effects would not necessarily provide an argument in favour of base generation over movement.

Indeed, we suspect that this is the situation in Akan. Previous discussions of WCO in Akan have failed to take into account the fact that Akan allows extraction from the possessor position of a noun phrase (56), in violation of the Left-Branch Condition (Ross 1967). This then opens up the derivation possibility in (55) for apparent WCO configurations such as (53b). These can instead be analyzed as involving resumption in the possessor position, which we treat analogous to subject resumption as involving null resumptive triggering agreement (57). This corresponds to the analysis in (55) in which there is no crossover at all. For this reason, Hewett (2023) argues that this kind of primary crossover configuration is inconclusive for determining whether movement or base generation is involved.

In response to this, different options for ruling out the derivational possibility in (55) have been explored. One option, for example, would be to replace the higher pronoun with an element that can function as a bound variable but not as a resumptive element. This is the strategy employed by McCloskey (1990, 2011) in testing for SCO in Irish. The problem here is that although epithets are not possible resumptives in Akan, they also may not function as bound variables either, as the ungrammaticality of (58) under a bound interpretation indicates.

(58)	*[ ɔdaduani biara ]1 bɔ  mpae s    dabidabi aban   bɛ-hu [ gyimifoO
	prisoner every pray prayer comp someday government fut-see fool
	no ]1 mɔbɔ
	def mercy
	‘Every prisoner1 prays that the government will have mercy on the fool1 someday.’

For this reason, the fact that the example in (59) is ungrammatical is still relatively uninformative about whether there are WCO effects in Akan, as the epithet can function neither as a resumptive element nor a bound variable.

There is, however, an alternative strategy we can use to try to test for WCO in Akan which equally rules out the possibility that the resumptive pronoun is in the possessor position. This comes from what Postal (1993) called ‘secondary crossover effects’, originally discussed by Higginbotham (1980) (also see Safir 1984). The examples in (60) exhibit the secondary WCO paradigm.

(60)	Secondary Weak Crossover
	a. [Whose1 mother]2____2 hates him1 ?
	b. *[Whose1 mother]2 does [his1 sister] hate ____2 ?

What distinguishes secondary WCO from the regular WCO configuration is that the quantifier that binds the pronoun is contained inside the moving phrase rather than being the moved phrase itself (compare *Who₁ does his₁ sister hate?). Nevertheless, we find unacceptability just in case the bound pronoun is ‘crossed’ by movement (60b), i.e., if the pronoun is not c-commanded by the base-position of the moved phrase containing the binder.

The analytical challenge posed by such data is that appealing to binding of the pronoun by a trace in a c-commanding A-position is not straightforwardly possible as the trace now bears the ‘wrong’ index, i.e., two. Nevertheless, (60a) is acceptable. What we can assume is at play here is that a pronoun may be bound ‘indirectly’ from the possessor position of a phrase in an argument position (see Büring 2004 for an explicit proposal for how this could work). If we assume that movement reconstructs, then possessor binding from an argument position will be possible in (60a) but not in (60b), where the base position is lower than the phrase containing the bound pronoun.

If we recreate the secondary WCO paradigm in Akan, we find the same pattern as we do for English. We have an apparent secondary WCO effect in (61b) where the resumptive pronoun is an object, while binding from the possessor is possible if the position of the subject resumptive pronoun c-commands the bound pronoun (61a).

(61)	Secondary WCO in Akan
	a. Me-bisá-a ɛsέ   [ hwan1 ba   ]2 na ɔ2-hú-uɛɛ no1
	1sg-ask-pstcomp who child foc 3sg.sbj-see-pst 3sg.obj
	‘I asked whose1 child saw him1.’
	b. *Me-bisá-a ɛɛsέ   [ hwán1 ba ɛɛ]2 na n1-adanfo sómá-a no2
	1sg-ask-pstcomp    who child    foc poss-friend send-pst 3sg.obj
	‘I asked whose1 child his1 friend sent.’

This example also allows us to avoid the confounding possibility of extraction from the possessor, but in a slightly different way. We can rule out the possibility of the possessor position in (61b) being the resumptive, as it bears a different index to the moved phrase. For this reason, we can be certain that we have a crossover configuration in (61b) and that the reason for the ungrammaticality must be that there is no c-commanding argument position from which possessor binding can take place. Given these findings, it seems that one might be able to uphold the claim that Akan exhibits sensitivity to WCO after all.

With that said, there is an alternative view that would say that secondary crossover effects are not reliable diagnostics for movement either. Hewett (2023) has recently demonstrated that secondary crossover is found with island-internal resumption and can be derived under base generation given the theory of binding developed in Büring (2004). The crucial assumption of this theory that Hewett (2023) highlights is that the operator responsible for indirect binding out of an XP ($\Sigma$) is excluded from Ā-positions (Hewett 2023, p. 399). This means that there is no way to bind a resumptive from inside of a base-generated or a moved XP in an Ā-position, thereby giving rise to secondary crossover effects in non-movement derivations. On this view, secondary crossover effects do not diagnose movement, as they are also expected to arise in a base-generated Ā-dependency in which there is no intervening A-position from which binding could take place. Primary crossover effects are also analytically ambiguous given the aforementioned complication regarding identifying which pronoun is the resumptive one. This places increased importance on the particular analysis of crossover effects that one adopts, as Hewett’s argument against secondary crossover as a movement diagnostic very much rests on Büring’s (2004) theory of crossover. Alternative analyses may not necessarily have the property of deriving secondary crossover with base-generation ‘for free’, in which case this diagnostic may not necessarily be inconclusive after all.13

3.5. Evidence for Base-Generated Ā-constructions in Akan

In addition to the arguments we have seen above, which may be interpreted as supporting the claim that resumptive dependencies in Akan involve movement, there are also language-internal reasons to doubt that base generation is involved in all Ā-constructions. The reason for this is that there is evidence that Akan does in principle have a strategy for forming resumptive Ā-dependencies with base generation and these dependencies crucially lack the kind of cyclicity effects that we have identified previously. This divergence therefore provides an argument against treating all Ā-dependencies as involving base generation.

The first relevant construction is the déέ-construction. Korsah and Murphy (2020) note that, in contrast to the na-focus construction and relativization, the déέ-construction does not show any reflexes of movement—neither tonal overwriting nor anti-agreement. This difference can be seen in (62).

(62)	No tonal overwriting or anti-agreement in thedéέ-construction
	a. Kofi1 na ɛ/ɔ1-káń-n    kŕataá nó
	Kofi foc 3-/3sg.sbj-read-pst book def
	‘It is Kofi who read the book.’
	b. Kofi1 déέ, *ɛ/ɔ1-kan-n   kŕataá nó
	Kofi top *3-/3sg.sbj-read-pst book   def
	‘As for Kofi, he read the book.’

The conclusion that Korsah and Murphy (2020) draw is that the pivot of the déέ-construction is base-generated, whereas the na-focus construction involves movement. If this is the case, then we have good reason to believe that Akan does in fact have an independent base generation strategy for Ā-dependencies that has a distinct empirical footprint to the other Ā-constructions we have discussed (i.e., na-focus constructions and relativization). Importantly, however, neither type of construction shows sensitivity to islands with nominal extraction.

Further evidence for base generation comes from the construction in (63b) (also discussed in Korsah and Murphy 2020). A possible interpretation of (63b) is that it involves raising of the embedded subject to the matrix clause with concomitant subject resumption.

(63)	a. Kofí nim  [CP sɛ Ám!má pɛ Yaw ]
	Kofi know   that Ama   love Yaw
	‘Kofi knows that Ama loves Yaw.’
	b. Kofí nim   Ám!má1 [CP sɛ O1-pɛ Yaw ]
	Kofi know Ama   that 3sg.sbj-love Yaw
	‘Kofi knows Ama to be someone who loves Yaw.’

Two things are unexpected about (63b) if movement is at play, however. We do not find tonal overwriting on the embedded verb pɛ, and the unagreeing subject pronoun E- is not possible, as (64) shows.

(64)	Kofí nim Ám!má1 [CP sɛ {ɔ1-/*ɛ1-}pɛ   Yaw ]
	Kofi know Ama    that {3sg.sbj-/*3-}love Yaw
	‘Kofi knows of Ama that she loves Yaw.’

Since both of tonal overwriting and anti-agreement can be taken to be diagnostics of movement, Korsah and Murphy (2020) propose that this construction involves a kind of prolepsis, i.e., base generation and binding (Salzmann 2017a), rather than movement. This is further supported by the fact that this construction shows idiosyncratic predicate restrictions, and while it is possible with ‘know’, it does not work with ‘think’ and ‘say’ (65b). This lexical restriction makes sense if the higher object is actually an object of the higher verb, but remains rather puzzling if movement was involved (note that all of these are bridge verbs).

(65)	a. *Kofi dwene Ám!má1 [CP sɛ ɔ1-pɛ   Yaw ]
	Kofi think  Ama    that 3sg.sbj-like Yaw
	Int: ‘Kofi thinks of Ama that she loves Yaw.’
	b. *Kofi ka-a Ám!má1 [CP sɛ ɔ1-pɛ   Yaw ]
	Kofi say-pst Ama   that 3sg.sbj-like Yaw
	Int: ‘Kofi said of Ama that she loves Yaw.’

What both of these constructions show is that Akan does in fact have a strategy for forming Ā-dependencies using base generation and binding. Importantly, however, the canonical properties of Ā-constructions we have identified elsewhere are absent in these constructions. This provides an argument against treating na-focus constructions and relativization as involving base generation, as one would then have to find an alternative explanation for their divergent behaviour with regard to tonal overwriting and anti-agreement, which does not rely on movement (see Section 4.2.1 for further discussion).

4. Discussion

4.1. The Resumption Profile of Akan

As we have seen, the picture that emerges from a close look at resumption in Akan is a mixed one, while nominal resumption does not show sensitivity to islands, it does show a number of other effects which one might want to classify as evidence for movement, e.g., cyclicity effects, reconstruction effects and crossover effects. As we will discuss in this section, conflicting diagnostics of movement within resumptive languages is not unusual, but we believe that Akan stands out in that it seems to pass almost every putative movement diagnostic except sensitivity to islands.

How does this fit into the cross-linguistic landscape of resumption? Some of the data discussed in Section 2 and that have been mentioned in the course of the preceding discussion are summarized in the table in 14, which we have organized according to the four different types of resumption profile we outlined at the outset (however, see Salzmann 2017b and Hewett 2023 for more comprehensive overviews).

(66) Some resumption profiles cross-linguistically 14
	Island-sensitive?	Evidence for Ā-movement?
		Crossover effects	Cyclicity effects	Reconstruction effects
Type Ia (Irish)	✗	✗	(✗)	—
Type Ib (Bulgarian)	✗	✗	—	✗
Type IIa (Vata)	✓	✓	—	—
Type IIb (Welsh)	(✓)	—	✓	✓
Type III (Scottish Gaelic?)	✓	—	✓	✗
Type IVa (Akan)	✗	(✓)	✓	✓
Type IVb (Jordanian Arabic)	✗	—	—	✓
Type IVc (Swedish)	✗	%	—	(✓)
Type IVd (Hebrew)	✗	(✓)	—	(✓)

As (66) makes clear, alongside what we have called Type I and II languages, in which island-sensitivity mostly patterns together with other movement diagnostics, there are some mixed profiles. As discussed in Section 2, it is unclear whether there is even a true Type III language, i.e., a language with island-sensitive resumption that fails other diagnostics for movement. Scottish Gaelic, as analyzed by Adger and Ramchand (2005), is perhaps the best candidate for such a language, but the evidence for resumption is somewhat more indirect as the resumptives in question are more often than not null in their crucial examples, and while Scottish Gaelic lacks reconstruction effects, it does show cyclicity effects on complementizers similar to what is found in Irish, giving rise to a mixed profile.

Aside from Akan, there are a few other languages that come into question for a Type IV profile, however, i.e., a language without island effects but that shows a range evidence of movement. These are listed in 14. As mentioned above, Jordanian Arabic has been argued to show reconstruction effects in the absence of islands (Guilliot and Malkawi 2011), but other diagnostics remain unclear. Another language with a potential Type IV profile could be Swedish. However, the situation surrounding resumption is complicated; while it has been claimed that resumptives show both WCO amnesty (Engdahl 1985, p. 9) and reconstruction for Principle A to the site of a resumptive pronoun (Zaenen et al. 1981), Asudeh (2012) presents several important qualifications about these data that may serve to weaken the case. Crossover is subject to speaker variation and the purported example of reconstruction by Zaenen et al. (1981) is subject to a serious confound, as the resumptive involved is arguably not a genuine resumptive (see Asudeh 2012, pp. 35–36). Finally, as noted by a reviewer, Hebrew could also be argued to have a Type IV profile. However, as with the other potential Type IV cases we mentioned, there are numerous complications here, too. Even though Hebrew resumption is generally agreed to be island-sensitive (Borer 1984), it has been argued that we find WCO effects in restrictive relative clauses (e.g., Demirdache 1991; Shlonsky 1992). The relevant examples use epithets to avoid the confound noted by McCloskey (1990); however, Sichel (2014, 667) argues that when factors such as register and information structure are controlled for, the relevant examples do not show WCO effects after all (though, see Hewett 2023, 426, fn. 78 for a qualification of this). In addition, resumption in Hebrew relative clauses has been argued to show reconstruction effects (Sichel 2014). Importantly, however, Sichel (2014) shows that this only holds for obligatory resumption contexts. Resumptive pronouns that alternate with gaps do not exhibit reconstruction, which Sichel attributes to the fact that both movement and base-generation derivations are available with resumption, with a preference for gaps where possible. This makes it significantly harder to diagnose when exactly movement is involved in resumption.

For this reason, Akan appears to be the clearest Type IV profile that has been identified thus far.15 According to the results of our investigation, Akan passes all three relevant movement diagnostics (albeit only with secondary crossover effects detectable), yet still lacks island sensitivity. There appears to be a genuine mismatch between island effects and other movement properties. The question that now follows from these findings is how one should then analyze resumption in Akan. This will be addressed in the following section.

4.2. Analytical Consequences

If Akan constitutes a genuine example of a mismatched Type IV resumption profile, then how can we go about reconciling the tension between the lack of island sensitivity with the other apparent evidence for movement? Given the current state of our theoretical understanding of resumption, it seems that there are essentially two options: either Akan resumption is derived by movement and some other process is responsible for the lack of island effects, or resumption in Akan must not involve movement, thereby giving us the lack of island violations, and all apparent evidence for movement must then receive an alternative explanation. We will argue that the latter approach, utilizing pure base generation or mixed chains, is not a sufficient explanation of the Akan pattern. Instead, we believe that a movement-based account is currently the best analysis available, while it does admittedly leave open the issue of how to derive island-insensitivity, for which we will consider two possible solutions.

4.2.1. Akan Resumption as Base Generation?

The first approach we will discuss is to take island-sensitivity at face value, i.e., as indicative of a base generation derivation, and try to find other explanations for the putative evidence for movement that appears to be in conflict with this conclusion. It is certainly true that the full cross-linguistic picture involving phenomena like reconstruction and crossover effects is murky, as indicated by the table in 14. Both island-sensitive and island-insensitive languages have been argued to show reconstruction and crossover effects, so it is rather unclear why this variation should exist if they are straightforward diagnostics for movement vs. base generation (see Salzmann 2017b, pp. 196–206 for discussion). It seems that, taken on their own, neither of these effects can be said to conclusively indicate that movement or base generation is involved.

As discussed in Section 3.4, the presence or absence of crossover effects may ultimately prove inconclusive. Following the conclusions in Hewett (2023), primary crossover configurations are uninformative as to which element is the resumptive (in the absence of epithets as bound variables), while secondary crossover effects are compatible with base generation on a theory that prohibits indirect binding from Ā-positions. Ultimately, the strength of secondary crossover in Akan as a diagnostic for movement will depend on the particular binding-theoretic assumptions one adopts. In Büring’s (2004) theory, secondary crossover does not uniquely diagnose a movement derivation for resumption.

As for reconstruction effects, it has been argued that these can be derived in a base-generated Ā-dependency by assuming that the resumptive pronouns is derived by a process of deletion of NP within the DP that leaves the D head stranded. Applied to Akan, where the resumptive is homophonous with the definite determiner nó, a resumptive dependency would look as follows (see Arkoh and Matthewson 2013 and Hein and Georgi 2021 for analyses along these lines):

(67)	Resumption as NP deletion
	[… [DP NP nó … [ … [DP NP nó ]]]

On this view, the elided NP could contain the relevant material responsible for reconstruction effects for Principle C or variable binding, for example. It is also clear, however, that not all resumption can be derived like this (see, e.g., Bulgarian deto-relatives), which leaves open the question of how to determine which kind exists in a given language. Furthermore, Section 4.2.2 will present a language-internal challenge for upholding the analysis in (67) for Akan.

If certain tests such as reconstruction and crossover do not unambiguously diagnose movement, then it seems that we are left with cyclicity effects such as the tonal overwriting pattern in the Asante Twi dialect as perhaps the most compelling potential evidence for a movement approach to resumption in Akan. It is then incumbent on a base generation analysis to provide an alternative account of this pattern that does not tie it to syntactic movement at all.

The first possibility would be to consider the domain of tonal overwriting as being defined over the linear string, such that overwriting applies to all verbs in the sentence that are linearly crossed by certain types of Ā-dependencies (e.g., those in relative clauses and focus constructions). Examples in Korsah and Murphy (2020) such as (68) illustrate that movement from the matrix clause does not trigger tonal overwriting in the lower clause. However, this would also be compatible with the aforementioned view that it is about the verbs that are crossed by the dependency in a purely linear sense. This alternative hypothesis can be conclusively dismissed, however, by considering cases in which a verb is linearly crossed while not being along the path of syntactic movement. If we take a sentential subject as in (69a) as a baseline, movement of a lower object will, under standard assumptions, not pass successive-cyclically through the complement clause to the noun ‘rumour’ inside the subject DP; although, it does cross it linearly. As (69b) shows, the verb inside the noun complement clause does not receive any high tones. For this reason, it becomes clear that we must reject a linear-based approach to tonal overwriting.16

(69) Tonal reflex only affects verbs along the movement path

The fact that it is the syntactic path to the resumptive is relevant could, however, still be captured by an alternative approach to long-distance dependencies that uses a chain of base-generated binding dependencies. This approach is sometimes referred to as iterative prolepsis and has been proposed for long-distance dependencies in a number of other languages (e.g., Adger and Ramchand 2005; Boeckx 2008; Davies 2003; Finer 1997; McCloskey 2002; Schneider-Zioga 2009). This is arguably the only analytical option available for tonal overwriting on the base generation view, as far as we can tell.

To see how this might work, consider the following example from Kinande in (70). As Schneider-Zioga (2009) discusses, long-distance dependencies trigger wh-agreement (in this case kyo) in each clause containing the dependency.

(70)	Long-distance dependency in Kinande (Schneider-Zioga 2009, p. 47)
	ekihi1 kyo   Kambale asi [CP nga  kyo   Yosefu akalengekanaya [CP
	what wh-agr Kambale know    comp wh-agr Yosefu thinks
	nga kyo   Mary’ akahuka ____ 1]]
	compwh-agr Mary cooks
	‘What does Kambale know that Yosefu thinks that Mary cooks?’

Furthermore, Schneider-Zioga (2009) shows that long-distance movement in Kinande lacks reconstruction effects, unlike local movement. This leads her to conclude that long-distance dependencies in the language are not formed by successive-cyclic movement, but rather by a series of successive binding dependencies (which may themselves involve clause-local movement of the operator/pronoun within the clause), as shown in (71). On this view, the presence of wh-agreement could be linked to the presence of a binder in the specifier of C, for example, much like McCloskey (2002) assumes for a^N-chains in Irish. The question now is whether we can apply the same analysis to Akan. We could in principle assume that tonal overwriting on a verb in a given clause is triggered by the presence of a base-generated pro-binder within that clause, for example. Here, the major challenge would be to differentiate between those base generation structures which trigger tonal overwriting and those which do not (in the déέ-construction, for example).

As noted by a reviewer, this general line of analysis could give us a way to deal with the challenge of movement reflexes inside islands without actually needing to posit movement out of the island. As schematized in (72), a potential analysis could be that the lower operator moves to the edge of the island and is then bound by a higher operator which also moves within the higher clause. If the tonal reflex on a verb in Asante Twi tracks Ā-movement from, say, Spec-vP to Spec-CP, then we would correctly predict that we find the reflex both inside and outside of the island without actually having to posit movement out of the island.

One of the major issues with such an approach is that it is not obvious why both base-generated operators would be compelled to move. As we have seen, Akan does in fact allow base generation in principle, as indicated by the absence of tonal overwriting in the déέ-construction, for example, in (62). For this reason, we might expect to find ‘mixed chains’ of the kind described for both Irish (McCloskey 2002) and Selayarese (Finer 1997) once direct base generation without movement is permitted in the grammar. In other languages that have both base generation and movement strategies to form an Ā-dependency, there is apparently no requirement for movement in the lower half of the dependency, for example. This can be seen by the absence of the familiar morphological reflexes of successive-cyclicity in the lower clause. If long-distance dependencies were formed as in (72), we might expect Akan to allow mixed chains in which the operator in the lower clause does not move (73). The consequence of this would be that we find tonal overwriting in the higher but not in the lower clause.17 This prediction is not borne out. Long-distance dependencies always exhibit tonal overwriting in all subordinate clauses along the path of movement.

Furthermore, if a long-distance Ā-dependency terminates within an island, tonal overwriting on all verbs inside the island is also obligatory. What this would then entail is that base generation must be ruled out in (73), even when the lower CP is an island. If long-distance Ā-dependencies were formed by a series of proleptic binding relations, then there would have to be some additional requirement that each of the binders moves within its local clause. This raises the challenge of adequately restricting the theory. Once a language permits base generation as a strategy, as Akan clearly does, then it is unclear how to uniformly enforce movement of these base-generated operators, as would have to be the case.

One possibility could be to assume that the relevant binders are generated in Spec-vP and must move to Spec-CP in order to remain accessible (however, note that this raises non-trivial questions about the locality of binding relations, as binding is typically not subject to such restrictions). Again, in order to account for the lack of a tonal reflex in déέ-constructions, we would have to stipulate that base generation directly in Spec-CP is only possible in matrix clauses. Once again, this does not solve the restrictiveness problem. We would then predict mixed chains of the kind that is attested in Irish, namely base generation in the highest clause and movement in the lower clause (74). In Akan, this would correspond to a long-distance dependency with tonal overwriting in the lower clause but not in the higher clause This is not a pattern that we find. Akan does not, to the best of our knowledge, exhibit mixed chains of any kind. Assuming that Ā-dependencies in Akan are composed of a series of propletic binding relations raises the problem of then adequately constraining this mechanism. Once we have base generation and movement of operators, we seem to incorrectly predict the kind of mixed chains that we find in Irish and other languages.

Furthermore, there is another challenge faced by the approach in (72). On this view, the absence of island effects would be due to the fact that there is never actually any movement out of the island rather than any kind of island repair mechanism. However, recall that this is only true for movement of nominals. Ā-displacement of VPs and PPs is subject to island constraints, as we saw in (27) and (29). Furthermore, long-distance extraction of these categories leads to tonal overwriting, see (26) and (28), showing that they require a movement derivation. The problem is that, in order to derive the island-sensitivity of PP and VP extraction, the island-avoiding derivation in (75a) must be unavailable. Instead, movement must be to proceed out of the island, as in (75b). This leads to a further issue in constraining the prolepsis theory of Ā-dependencies in Akan, while admittedly the island-insensitivity of nominal extraction is also challenging for the movement-based analysis that we will discuss in the following section; neither of the solutions that we will discuss there (island obviation by resumption or category-discriminating islands) extends straightforwardly to a theory involving a series of binding dependencies between null operators.

In sum, the base generation alternative struggles to capture the distribution of the tonal reflex of displacement in a restricted way. Since it is clear that Akan must allow base generation as an option in principle, a theory employing iterative prolepsis must posit ad hoc restrictions on where operators may be base-generated, while also enforcing a general requirement that they move. As we have shown, this appears to predict that we should find mixed chains of the kind we do in Irish and Selayarese, contrary to the fact. In addition, there is no further supporting evidence for the base generation approach beyond the lack of island effects (e.g., the absence of reconstruction effects as in Kinande).

4.2.2. Akan Resumption as Movement

Given the fact that base generation cannot provide a fully satisfactory account of the properties of resumption in Akan, cyclicity effects in particular, we will argue that the best approach to the Akan data currently available is one involving a movement derivation of resumption. On this view, the island insensitivity that we observe with nominal extraction must then be explained by some other assumption(s).

The first analytical possibility to account for the apparently contradictory movement diagnostics is to assume, as Korsah and Murphy (2020) do, that Akan resumption is uniformly derived by movement and that the availability of a resumptive pronoun is directly responsible for obviating island effects. In their view, islands are PF constraints that penalize gaps in certain structural configurations. Furthermore, they assume that such constraints can be obviated by a resumptive pronoun at the tail of the dependency (a kind of grammaticalized intrusive resumption). Korsah and Murphy (2020) propose a rule of Pronoun Conversion that turns the lowest DP in a movement chain into a pronoun, as schematized in (76). This is essentially the equivalent of Fiengo and May’s (1994) notion of ‘vehicle change’, but applied to PF.

(76)	Pronoun Conversion
	[CP DP1 [TP … [ … DP1 … ]] ⟹ [CP DP1 [TP … [ … pro1 … ]]]

While this seems to give a straightforward reconciliation of the evidence for movement with island insensitivity, it complicates our view of islands significantly. Korsah and Murphy must assume that islands are representational constraints, i.e., that they hold at PF. This view does in fact have some precedent, see, e.g., (Boeckx 2012; Griffiths and Lipták 2014; Hornstein et al. 2007; Lasnik 2001; Merchant 2001; Perlmutter 1972; Pesetsky 1998). For example, Pesetsky (1998, p. 365) argues for a similar representational view of island constraints, as in (77).

(77)	*$\alpha$ … [island … β … ], where $\beta$ is the trace of α and unpronounced.

On this view, Pronoun Conversion would create a PF representation that complies with the general formulation of island constraints in (77). A complication to this approach is the fact that islands in Akan are obviated by resumption even if the resumptive pronoun undergoes a pro drop. Recall that we saw that the apparent animacy distinction in resumption was shown to be blocked in certain obligatorily pronoun contexts, which we took to show that movement of DPs always results in a resumptive pronoun. In order to satisfy the constraint in (77), the representational island constraint must be checked before pro drop applies (an instance of counter-bleeding). This would necessitate that the pro drop be a relatively late process on the PF branch, arguably complicating the architecture of the post-syntax.

Furthermore, recall from Section 3.1 that Hein and Georgi (2021) raised an empirical problem for Korsah and Murphy’s account; while both PPs and VPs lack resumptives and also give rise to island effects, Hein and Georgi showed that certain types of nominal expressions, namely idiom chunks, predicate nominals and non-specific indefinites, also lack overt resumptives in obligatory pronoun contexts (much like PPs and VPs), yet are still insensitive to islands (unlike VP and PP extraction). This serves to weaken the link between overt resumption and island insensitivity that underlies the proposal of Pronoun Conversion. In response to this, there are two options: one can abandon this approach to island obviation in favour of an alternative approach (such as Hein and Georgi’s, which we will discuss in a moment) or one could seek an independent explanation for why the class of nominals discussed by Hein and Georgi appears incompatible with overt resumption.

A possible way of pursuing the latter approach would be to try to argue that there is a more general constraint that disallows overt pronouns that are co-indexed with nominals of the kind discussed by Hein and Georgi (2021). Resumptive pronouns would then also happen to fall under this more general constraint. To show this, we would want to find contexts in which an overt non-resumptive pronoun may not refer to a nominal of the relevant type, yet a silent pronoun may. In practice, this is difficult to test because the nominals in question are not easy to construe as referential. Nevertheless, we were able to construct the relevant test sentences. Consider the example in (78) where we are trying to refer back to a predicate nominal, a noun-type that Hein and Georgi show to reject a resumptive pronoun.

(78)	Na Kwadwo pɛ   sɛ   ɔ-yɛ   odusini1, nanso ɔ-a-n-yɛ    ?bi1 /
	pst Kwadwo want comp 3sg-fut-be herbalist but  3sg-perf-neg-be   indef
	*no1   / ____1
	3sg.obj
	‘Kwadwo wanted to become a herbalist, but he did not become (one)’

Here, the only possible pronoun forms are a silent (dropped) pronoun (indicated as a gap) or, somewhat marginally, an overt indefinite pronoun bi. The overt anaphoric pronoun no is ruled out. This could be taken as indication that pronouns referring to predicate nominals must obligatorily dropped.18

In a similar way, we can construct cases in which the anaphoric pronoun is contained inside an adjunct clause. Taking a different kind of nominal here, that is a non-specific indefinite, we observe the same effect. The anaphoric pronoun in the adjunct must be null when it refers to the indefinite ‘people’ (79a). The pattern is reversed if the extracted NP is specific, where an overt pronoun becomes obligatory (79b).19

(79)	a. Nípa na Kofi súró ____1 [CP ésánesέ ɔ-fέré   {____1 / *nó1 } ]
	person foc Kofi fear    because   3sg.sbj-be.shy.of     3sg.obj
	‘It’s people that Kofi really fears because he is shy of (them).’
	b. Nípa  yi1  na  Kofi súró nó1  [CP ésánesέ ɔ-fέré    { *____1 /
	person this foc Kofi fear 3sg.obj because 3sg.sbj-be.shy.of
	nó1 }    ]
	3sg.obj
	‘It’s this person that Kofi really fears because he is shy of him.’

A possible way of interpreting this is that there is a more general process (e.g., an obligatory pro-drop rule) that applies to pronouns that are co-indexed with the class of nominal expressions that Hein and Georgi identified. Whatever process applies to block overt forms of anaphoric pronouns would then arguably equally apply to resumptive pronouns, too. While these findings are still tentative, this would potentially offer a way to reconcile Hein and Georgi’s (2021) findings with the Pronoun Conversion proposal outlined above.

An alternative route is to pursue an entirely different account of island obviation, as Hein and Georgi (2021) do. Hein and Georgi seek to derive island insensitivity in Akan from the nature of island constraints themselves. In particular, they argue that island constraints may be category-discriminating, presumably on a language-specific basis. The idea that islands may be category-sensitive has some precedent in work by Cinque (1990) and Postal (1998) who discuss so-called ‘selective’ islands that are sensitive to the type of extractee involved. Contrasts of the kind in (80) have been discussed in the literature, which may seem to indicate that extraction of PPs from wh-islands differs from extraction of DPs (though, see Szabolcsi 2006 for discussion of further factors).

80	a. [̣PP About whom ] do you wonder [CP whether to worry ____ PP ] ?
	b. [DP Who ] do you wonder [CP whether to worry about ____DP ] ?

In a similar vein, Hein and Georgi suggest that islands in Akan do not apply to the extraction of phrases with a ‘nominal core’, i.e., which contain an NP. This is because they also argue that the class of nominal expressions that do not show resumptives and yet are still island-insensitive (predicate nominals, non-specific indefinites and idiom chunks) are NPs rather than DPs; while a fully-fledged theory of category-sensitive islands has, to the best of our knowledge, not yet been worked out, it is clear that the proposal by Hein and Georgi (2021) would require significant complications, as it is not just the category of the extracted element which matters but also its internal structure (i.e., whether it contains a ‘nominal core’).20

On Hein and Georgi’s (2021) analysis, there is no direct link between the availability of a resumptive pronoun and sensitivity to islands. Hein and Georgi (2021) instead propose a deletion-based account in which resumptives in Akan are derived from deletion in the lowest copy applies to NPs, VPs and PPs, but not DPs. This leads to the view of resumptive pronouns as a stranded D head that is found only with DP extractees (81). On this view, predicate nominals, idiom chunks and non-specific indefinites are incompatible with (resumptive) pronouns because the deletion rule applies to NPs (in addition to VPs and PPs), meaning that there is no stranded determiner that could be realized as a pronoun. Nevertheless, since category-sensitive island effects apply to phrases with a ‘nominal’ core, both DP and NP extraction is island-insensitive regardless of the availability of a resumptive.

It is worth mentioning that adopting this particular analysis of resumption for Akan has been argued against by Korsah and Murphy (2020) on the basis of haplology effects. As pointed out by Saah (1994), a sequence of homophonous determiner nó’s in Akan is not possible. As (82a) shows, if the first nó is a resumptive pronoun, then this may occur adjacent to a clausal determiner. If the first of the determiner sequence is also a genuine determiner, however, then this is ruled out, and one of them must be deleted.21

(82)	Determiner haplology effect (Saah 1994, pp. 153–154)
	a. [DP AbOfrá [CP Op1 áa Kofí hú-u   nó1   nó ]] á-ba
	child    rel Kofi saw-pst 3sg.obj cd    perf-come
	‘The child that Kofi saw has come.’
	b. [DP Onípá [CP Op1 áa ɔ1-tó-o     [DP ndwóḿ nó ] (*nó) ]] yɛ-ɛ
	person rel 3sg.sbj-throw-pst    song   def    cd   do-pst
	adé
	something
	‘The person who sang the song did well.’

Korsah and Murphy argue that there is no good reason why the haplology constraint should care if the NP complement of the determiner is deleted. The relevant rule should still apply in cases such as (82a), even if the resumptive is a stranded D head. On the Pronoun Conversion analysis, however, the resumptive is still a phrasal constituent and would therefore be immune from a rule that deletes the second in a sequence of adjacent homophonous D⁰ elements.

For this reason, it seems that both an explanation in terms of Pronoun Conversion and category-sensitive islands faces some complications, and while both can adequately reconcile a movement-based derivation of resumption with the lack of island violations with nominal resumption, it is not clear to us which one should be preferred at this point. Nevertheless, it seems that a movement-based derivation of islands coupled with either one of these assumptions can deliver a more satisfactory account of the Akan data than its base generation competitor.

5. Conclusions

In this paper, we have discussed the properties of resumption in Ā-dependencies in Akan. While resumptive dependencies in Akan are island-insensitive, they also show various other properties that could be taken to indicate a movement derivation. For example, they show reconstruction, crossover and cyclicity effects. This profile seems to occupy a relatively unusual position in the typology of resumption at present. Despite robust island insensitivity with nominal extraction, these dependencies fulfil other diagnostics that are often taken to be indicative of movement. Among the possible interpretations of this state of affairs, one could argue that these diagnostics do not in fact identify movement after all or that this forces us to the conclusion that a movement derivation can still be maintained even in the face of island insensitivity.

Here, we have argued for the latter conclusion. We have shown that by attempting to reconcile the other diagnostics, the tonal overwriting property of displacement constructions in particular, with either a pure base generation approach or mixed chains analysis that involves base generation and movement, is ultimately not a tenable solution. Such accounts open up a wider range of derivational possibilities and therefore over-predict the patterns of tonal overwriting that we expect to find. What is more, Akan does in fact seem to have a class of base-generated Ā-constructions which have different properties than those indicated above, which further undermines this line of analysis.

Instead, we believe that an approach involving a movement derivation offers the best way of understanding resumption in Akan. On this view, the various evidence for movement is accommodated straightforwardly. The central challenge is then accounting for the absence of uniform island effects in the language. We have discussed two potential solutions: (i) following Korsah and Murphy (2020), resumption is directly responsible for island obviation on the assumption that islands are representational PF constraints, or (ii) islands in Akan should be treated as category-discriminating and solely prohibit extraction of DPs, as argued by Hein and Georgi (2021). While both seem to be viable solutions, we are currently not in a position to adjudicate between these two approaches.

In sum, Akan provides an interesting contribution to the typology of resumptive Ā-dependencies. We have argued that, at least in descriptive terms, it may well provide the strongest case for what we called a ‘Type IV’ profile of resumption where island insensitivity conflicts with other putative movement tests. Reconciling the potential tensions between island (in)sensitivity and other evidence for movement is still an ongoing task that requires further careful investigation of a wider range of languages. At this point, however, it is clear that the cross-linguistic picture of resumption that emerges requires us to go beyond the traditional dichotomy between movement and base generation.