When hypotaxis looks like parataxis: embedding and complementizer agreement in Teiwa

Teiwa, an Alor-Pantar language of the Trans-New Guinea family, has been characterized as expressing speech reports not with complementation, but with combinations of two clauses juxtaposed under a single intonation contour with no morphological indication for integration (Klamer 2010: A Grammar of Teiwa, Mouton de Gruyter). We argue, contra Klamer, that speech and attitude reports in Teiwa should be analyzed as embedding (or hypotaxis). We present evidence from intonation, syntax and semantics that speech reports are expressed by a single, monosentential structure in Teiwa with embedding of the speech report. Our results also show that purely morphological diagnostics can be unreliable for distinguishing between a monosentential or bisentential structure of speech reports. We describe several formal experiments from our fieldwork that provide more reliable tests. Our result has implications for both the ongoing theoretical discussions of clausal complementation, complementizer agreement, grammaticalization of complementizers and the historical evolution of complementation.


Introduction
Grammatical analysis of embedded clauses has been a central theme of modern theoretical linguistics since Chomsky (1957) used data from complex clauses to gain deeper insight into the formal properties of natural language. Recently, researchers have gained new insight into complementation through close examination of syntactic and semantic properties across a wide range of languages (Aboh 2005;Arsenijević 2009;Bogal-Albritten 2016;Elliott 2017;Jayaseelan 2014;Kastner 2015;Kratzer 2006;Moulton 2015;and others). Our primary empirical interest in this paper is embedding in the Teiwa language (Klamer 2010a). Understanding embedding in Teiwa is of general interest because Teiwa was claimed to lack complementation entirely and to instead use juxtaposition under a single intonational contour (Klamer 2010a). (1) is a Teiwa example of this type of juxtaposition from our own fieldwork: (1) Natan a na-walas a wa bas a Qalambas ma gi Natan 3sg 1sg-tell 3sg say tomorrow 3sg Kalabahi come go 'Natan told me that he will go to Kalabahi tomorrow. ' [LT.CD1.008] 1 Example (1) contains both the verb walas 'tell' and following it, the verb wa 'say'. However, in the English translation we provide, the speech report expressed by (1) requires only one verb of speech, told. In other words, the Teiwa (1) seems to correspond to the English sentence Natan told me that he said that … with double embedding, but this interpretation is actually very marginal in Teiwa. Therefore an analysis of (1) as embedding seems initially implausible. Since at the same time, several substrings of (1) can form independent sentences as shown in (2), it seems plausible that (1) consists of more than one sentence. An analysis as a sequence of the two independent sentences (2a) and (2b) is the alternative proposed by Klamer to which we refer in this paper as the bisentential analysis.
(2) a. Natan a na-walas Natan 3sg 1sg-tell 'Natan was talking to me. ' [sentence 1] b. a wa bas a Qalambas ma gi 3sg say tomorrow 3sg Kalabahi come go 'He said he is going to Kalabahi tomorrow. ' [sentence 2] We argue here that (1) must be analyzed as a single sentence where the sequence a wa acts as a complementizer agreeing with the matrix clause subject. We refer to our proposal as the monosentential analysis. As far as we know, the way speech and attitude reports are expressed in Teiwa is unique among living languages. The complementizers are related to the verb say in many languages, e.g., Malayalam (Jayaseelan 2014), Matses (Munro et al. 2012), or Usan and other Papuan languages (Reesink 1993), but only a handful of languages link speech complements with the fully inflected verb say. One such language is Nanti, where the complementizer (glossed as quot) agrees with the matrix subject (see (12) below). Agreeing complementizers are reported for a number of Bantu languages (Kawasha 2007;Diercks 2013) several Mande languages of West Africa (Idiatov 2010;van Koppen 2017), and in Nilotic (Diercks & Rao 2019). Idiatov (2010) argues that such agreeing complementizers, at least in Mande, appear to be always related to quotative predicators and ultimately verbs, sometimes still present in the same form. As an example, we will discuss Lubukusu (Section 3 below). In Teiwa, the verb say linking speech and attitude reports requires a subject pronoun. 2 The language most similar to Teiwa that we are aware of is the extinct Akkadian (Deutscher 2000), a parallel to which we return in Section 3. The main goal of our paper is to propose a monosentential analysis of Teiwa examples like (1). Under our analysis a wa essentially fulfils the role of a complementizer. We first present background on the relevant aspects of Teiwa in Section 2. In Section 3, we discuss a number of languages with complementizer agreement. In Section 4.1, we address the status of the utterance initial a wa, as in (2b), where a wa does not follow another propositional attitude verb. We argue that in such cases, wa is an independent full verb taking a clausal complement following it. In Section 4.2, we then consider examples like (1), in which a wa is non-initial and prosodically integrated with the preceding propositional 2 One reviewer raised the question of whether forms such as a in (1) should be described as pronoun or clitic.
For our main argument this question is irrelevant and it is only important that the same form is used in (1) and in an independent sentence like (2b). We use the term pronoun for two reasons: to follow Klamer (2010a) and because the impersonal pronoun hala can occur in this position (see (30b) below). Hala is bisyllabic, can bear stress, and in object position is followed by agreement, and therefore not a clitic. Note furthermore that other pronouns such as a do not undergo reduction in rapid speech, even when the second syllable of the sequence a wa does for some speakers (further details in Section 4.2.5). attitude verb; we argue that a wa in such occurrences is a complementizer. We present three arguments in favor of the monosentential analysis and argue for an agreement relation between the speech verb (agreement controller) and the complementizer wa (agreement target). Our conclusion is that, despite initial appearances to the contrary, the monosentential analysis is correct based on evidence we present from prosody, syntax, and semantics. We return to the parallel between languages discussed in Section 3 and Teiwa in Section 5 where we state the implications of the Teiwa embedding on our understanding of the evolution of complementizers and complementation.

Background and the puzzle
In 2.1 we provide information about the Teiwa language and its speakers. The treatment of Teiwa complex sentences in Klamer (2010a) is presented in 2.2.

Teiwa language
Teiwa is spoken on the island of Pantar, located in the Nusa Tenggara Timur province of Eastern Indonesia, shown in Figure 1. Teiwa has about 4,000 speakers, who reside in two villages, Madar and Lebang. Most of the Teiwa speakers in Madar are male because the society is predominantly patrilocal (Klamer 2010a: 17). Virtually all Teiwa speakers are at least bilingual, speaking Teiwa and Malay. They are predominantly Christian, received at least a few years of elementary education in Indonesian and virtually everyone is fluent in Alor Malay. Christian worship has been conducted in Indonesian since the 1970s (Klamer 2010a: 3). Teiwa is an endangered language whose speakers are shifting towards Alor Malay. In the coastal village of Madar, children do not acquire Teiwa as their first language while in Lebang, located in the mountains, children still spoke the language and acquired it in most domains at the time of our fieldwork. Given the strong influence of Malay and Indonesian, it is possible that the language may have recently undergone changes in the areas of interest due to the influence of Malay/Indonesian. For this reason, wherever possible, we have made an effort to include speakers of various ages in the experimental studies. Despite our effort we did not detect a significant effect of age and conclude that the phenomena we report on exhibit some stability across native Teiwa speakers of different ages. Teiwa has a nominative-accusative alignment and limited morphology. Nouns agree with possessors (alienable vs. inalienable); verbs agree with human objects and can combine with realis suffixes. Teiwa is overwhelmingly head-final, the most common constituent order being OV, with negation and conjunctions following the predicate. Verb serialization is common (Klamer 2010a: 30-31).

Clause combinations in Teiwa
This section summarizes the analysis of Teiwa clause combinations, following (Klamer 2010a: 361-395), which provides the following description: Teiwa discourse is characterized by strings of clauses that can be connected by a conjunction or discourse linker, or separated by an intonational break, while there are also clauses that are merely juxtaposed to each other under a single intonation contour (Klamer 2010a: 363).
Specifically, Klamer reports that relative and complement clauses are not attested in Teiwa. Klamer further addresses subordinate adverbial clauses, as well as adnominal purpose clauses, which we will not discuss here. At the same time, Klamer clearly intends her discussion of complement clauses to be a preliminary one, stating that further detailed investigations of Teiwa clause combinations therefore remain necessary (Klamer 2010a: 363).
Serial verb constructions are distinguished from clause combinations by the empty subject position of one of the serialized verbs, as shown in (3). (3) Ba a ta tup-an a pin Ø gi seq 3sg top get.up-real 3sg [hold Ø go] serial.verb 'So he got up and left.' (Klamer 2010a: 308) As for adverbial clauses, Klamer shows that two temporal connectives exist in Teiwa: si 'sim' for temporal simultaneity (4a) and ba 'seq' for temporal sequentiality (4b). Ba is also used to describe causal relations.
(4) a. Qui yaa ma xa'a si i wan tad. caterpillar descend come this sim 3pl surround 'Caterpillars came down surrounding them/swarming around them.' (Klamer 2010a: 383) b. A uy i wan hamar ma walas ba a bis. 3sg person prox prayer come tell seq 3sg healthy 'He prayed for this person [so] he would get better.' (Klamer 2010a: 384) Si 'sim' and ba 'seq' must occur between two clauses or (for ba 'seq') two phrases. However, Klamer reports that there is no evidence of an asymmetry between two clauses connected by si 'sim' or ba 'seq': Both conjunctions function to express a temporal connection between two equivalent clauses (p. 386). So, syntactically, they behave like the coordinating conjunctions ata 'and' and le 'or' when they are used to connect clauses. The only evidence Klamer presents for clausal subordination in Teiwa is from an adnominal purpose clause construction illustrated in (5): qar ol-an in (5) must clearly be a subordinate clause because of its clause medial position. 3 (5) Ni'in sen qar ol-an ga-x wan maan 1pl.excl Location [money [rice buy-real] VP 3sg-possession] Theme be neg 'We do not have money to buy rice.' (Klamer 2010a: 377) (Lit. '[Possession of money to buy rice] does not exist (at) us').
With speech reports, Klamer finds no evidence of subordination in Teiwa (Klamer 2010a: 364-369) but does make a distinction between speech reports with the verb wa 'say' and those with other speech verbs. Consider first the verb wa. Klamer points out that the verb wa can be used to express not only the concepts 'say' and 'think', but also 'go' as a deictic motion verb. However, only the wa 'go' combines with the realis suffix. 4 Furthermore, when used to report speech or thought, wa can be followed by either direct or indirect speech. One clear example of indirect speech is (6).
(6) iman a wa iman paat, pi-soxai le, pi-pantum le they 3sg say they not.know 1pi-dance or 1pi-poem(ind) or 'They say they do not know our dances or our songs.' (Klamer 2010a: 275) Klamer does not commit herself to a specific syntactic structure for the analysis of speech reports with wa, but refers to them as 'quotative constructions'. One potential reason for a special status of wa may be word order: While complements in Teiwa generally precede the verb, in all of Klamer's examples with wa, the speech report follows wa, as shown in (6). However, the order of the verb paat and its object in (6) illustrates, the word-order of object and verb exhibits actually some flexibility, as Klamer also reports. We return to the analysis of wa below.
Of the other utterance verbs, Klamer's discussion of regan 'ask', bangan 'request/ask for something', and walas 'tell' is most relevant for our current purpose, though Klamer also discusses taxani 'inquire', soi 'order', and liin 'invite'. Klamer proposes that all of these verbs occur in juxtaposition with the reported utterance itself (p. 364) and describes the structures in which they occur as multi-clause combinations (p. 364). With regan 'ask', she provides the two examples in (7) to show that the content of the question reported can either follow regan 'ask' and its nominal arguments, as in (7a), or precede them, as in (7b).
(8) Iman mis-an wan om mai om ga-regan taxaran si pi ta in 3pl sit-real be inside store inside 3sg-ask how sim 1pl.incl top thing i er? forthc make 'They sat planning thinking: "How shall we do it?"' (Klamer 2010a: 373) In the case of bangan 'request/ask for something', Klamer points out that this verb has unusual word order properties: Object nominals can occur with bangan either in the usual preverbal position, shown in (9a), or postverbally, as in (9b), which she reports as otherwise impossible in Teiwa. Further, Klamer also suggests that the clause following bangan in (9c) serves as an object.
(9) a. Na sen bangan. 1sg money ask.for I ask for money. (Klamer 2010a: 366) b. Na bangan na-bangan qau. 1sg ask.for 1sg-life good 'I ask for a good life.' (Klamer 2010a: 366) c. Na bangan Kri Bas n-un-mulax geneg. 1sg ask.for Mr Bas 1s-appl-help a.bit 'I ask Mr Bas to help me a bit.' (Klamer 2010a: 367) (Lit. 'I ask for Mr Bas [he] helps me a bit.') Finally, the most interesting class of utterance verbs for our purpose consists of those followed by wa. Klamer uses walas to illustrate this class, as we do in (1) above, and reports that in most cases walas is followed by the quotative construction consisting of a wa 's/he says' and the speech report (direct or indirect speech), as illustrated in (10a) (indirect speech) and (10b) (direct speech).
(10) a. a tup-an a emaq u ga-walas, a wa a mau tewar 3sg get.up-real 3sg wife dist 3sg-tell 3sg say 3sg want(ind) walk por awan ta gi. island far top go 'he got up [and] said to his wife, he said he wanted to go to an island far away' (Klamer 2010a: 367) b. … a-yivar ga-walas a wa: "Ha siga' ga'an" … 3sg-dog 3sg-tell 3sg say 2sg be.quiet 3sg '… [he] told his dog saying: "You be quiet"' (Klamer 2010a: 367-68) Following on Klamer's investigation of complex clauses in Teiwa, our own inquiry is focused on the speech reports introduced by wa and on the status of examples like (1) and (10) where wa follows another speech verb. Before examining relevant cases in Teiwa in detail, we take a brief look at the typological and historical findings from other languages.

Say as a complementizer
Research on several other languages has described the use of forms related to the verb say to express propositional attitudes. There is ample evidence that complementizers commonly originate in a general verb of speech (Heine & Kuteva 2002;Güldemann 2008).
More specifically this has been demonstrated for some Austronesian languages (Klamer 2000), Creoles like Bislama (Crowley 2004: 182), Chadic (Frajzyngier 1996), Dravidian (Jayaseelan 2014), Ewe (Lord 1993), and some Papuan languages (Reesink 1993). In difference to Klamer's analysis, such uses have often been described as Complementizer forms of the verb say. In this section, we show that Teiwa represents a uniquely interesting case to understand when a verbal form should be analyzed as a complementizer.
For the following, we introduce a descriptive terminology: We write that the verb say occurs (or occurred) as Independent Occurrence expressing the verb of speech if no other propositional attitude verb occurs nearby that say might be construed with. In the Teiwa data we have seen up to now, (2b) and (6) contain such independent occurrences of wa. Furthermore, a form related to the same verb say can also occur as Integrated Occurence where it is construed with another propositional attitude verb. Of the Teiwa examples up to now, (1), (10a), and (10b) contain integrated occurrences of wa. The form of wa used in independent and integrated occurrences is actually identical in Teiwa, 5 but more frequently it seems that integrated forms of the verb say are reduced.
We focus on languages where the integrated uses of say exhibit agreement with the subject of the other attitude verb they are construed with. In European languages, complementizers either do not agree at all (as in English) or agree with the subject of the embedded clause, as in some dialects of German and Dutch (Zwart 1997: 137-141). However, the Germanic complementizer agreement is never controlled by the subject of the matrix clause. Complementizer agreement with matrix subjects has been observed only in Nanti (Michael 2008) and in a number of Bantu languages (Kawasha 2007;Diercks 2013), several Mande languages of West Africa (Idiatov 2010;van Koppen 2017), and in Nilotic (Diercks & Rao 2019). One such language is Lubukusu (Bantu, Kenya). Diercks (2013) introduces the notion complementizer agreement using Lubukusu examples such as (11). 6 The subject of the speech verb babolela 'say' in (11) is babandu 'people', a class 2 noun. It is the controller of the subject agreement prefix ba-attached to both the verb and the complementizer li. In (11b), the agreement controller is a class 1 noun Alfredi.

(11)
Lubukusu (Bantu, Kenya) a. Ba-ba-ndu ba-bol-el-a Alfredi ba-li a-kha-khil-e. 2-2-people 2s-said-ap-fv 1Alfred 2-that 1s-fut-conquer 'The people told Alfred that he will win.' b. Alfredi ka-bol-el-a ba-ba-ndu a-li ba-kha-khil-e. 1Alfred 1s-said-ap-fv 2-2-people 1-that 2s-fut-conquer 'Alfred told the people that they will win.' (Diercks 2013: 358) Diercks (2013) shows that the complementizer agreement in Lubukusu is productive: various noun classes control agreement on the complementizer (p. 364-365) and the agreement is not affected by negation (p. 366). Lubukusu has a number of complementizers, among which agr-li appears to be restricted to verbs of speech/manner of speech, knowledge/belief, and desire; it is incompatible with verbs that presuppose the truth of their complement (Diercks 2013: 397-398). We do not know whether the complementizer li of Lubukusu is related to a verb of saying. Complementizer agreement is also reported for Nanti (Arawakan, Peru). In Nanti, we find both independent occurrences of kaNt 'say' and integrated occurrences thereof as ka, but Michael's analysis is different from Klamer's analysis of Teiwa. Michael (2008: 111) analyzes the Nanti ka in (12) as a marker of quotative embedded clauses (a complementizer in our terms) that agrees with the matrix subject. Specifically, ka occurs with first person agreement in (12a) and with third person masculine agreement in (12b). Michael (2008) identifies the verb kaNt 'say', shown in (12b), as the diachronic source of the marker ka and reports that similar morphemes are also attested in other Campa languages (Michael 2008: 111-112).
(12) Nanti (Arawakan, Peru) a. no=kenkitsa-ak-i no-ka no=gonke-haa-a 1sg=tell.story-perf-realis.i 1sg-quot 1sg=arrive-cl:water-real.a Shanpinkihari place name 'I told a story of my arriving in Shanpinkihari.' b. i=n-kant-e i-ka-ha tomi no=ha-i 3mS=irr-say-irr.i 3ms-quot-neg.irr son 1sg-go-real.i no-n-kamoso-e 1sg=real-visit-real.i 'He would say: I will not go and visit my son.' The closest parallel to Teiwa is presented by Akkadian (Semitic, Ancient Mesopotamia), whose long written record captured the development of a general complementizer (Deutscher 2000). Akkadian is the earliest known Semitic language, which was used in writing from around 2500 BC until the beginning of the current epoch in a great variety of genres. Akkadian letters are available from the earliest period and reflect well the development of the colloquial language because they were dictated by illiterate clients to scribes to be read out to the recipient, a property which makes the letters well suited for the investigation of grammaticalization (Deutscher 2000: 23).
As shown in Table 1, Old Akkadian is characterized by (Deutscher 2000) as a language without embedding where enma S(-ma) and later umma S(-ma) encode an independent clause paratactically joined with a speech report attributed to the speker (marked as S here). From Early Old Babylonian umma S-ma is attested with other speech verbs and the S marking disappears in the Middle Babylonian period. During the Neo-Babylonian period, umma appears as a complementizer with verbs such as 'fear' or 'hear', as shown in Table 1. In Early Old Babylonian umma may form an independent clause as in (13). 7 The subject pronoun anaku 'I' combines with the particle ma and has to follow umma. Deutscher (2000: 70) describes ma as a particle used for emphasis and coordination.

(13)
Early Old Babylonian umma anaku-ma alp-um sa taqabbi'-u [u-la i] z-za-az umma 1-part ox-nom rel you.say-sub is not available 'I said: "the ox you are talking about is not available".' (Deutscher 2000: 72) Several examples of integrated uses of Early Old Babylonian umma with speech verbs are given in (14). Deutscher (2000: 72) states that the speech verb and umma are in paratactic relation. Note that the subject pronoun following umma 'say' is first person singular in (13a) and (13b) but third person plural in (13c). The subject of umma is the same as the subject of the speech verb, resembling the Lubukusu and Nanti agreement pattern.
(14) Early Old Babylonian a. ašpurak-kum umma anāku-ma litma'ā … I.wrote-to.you umma 1-part let.them.swear … 'I wrote to you, (and) this is what I said: "…"' (Deutscher 2000: 72) b. a-ap-pá-al-su [um-m]a anāku-ma … [5 lines] I.answered-him umma 1-part 'I answered, (and) this is what I said "…"' (Deutscher 2000: 72) c. iq-bi-ú ni-im-ma umma sunu-ma … they.said-to me-part umma they-part … 'They talked to me, (and) this is what they said "…"' (our translation) (Deutscher 2000: 73) As the translations given in (14) indicate, Deutscher (2000) does not consider the Early Old Babylonian data to be instances of embedding, but of two independent paratactically joined sentences (see also Deutscher 2005), i.e. a similar analysis to the one Klamer (2010a) adopts for Teiwa examples like (6). In sum, the previous literature suggests that structures with integrated occurrences of say (or an agreeing complementizer) receive different analysis across languages: a monosentential, complementation structure when the integrated form of say is morphologically distinct from the independent forms as in Nanti and all languages without agreement on integrated say, but a bisentential, paratactic structure when the integrated and independent forms of say are identical as in Early Old Babylonian and Teiwa. Teiwa is the only living language of the latter type as far as we know, and therefore it is important to apply further diagnostics for complementation to it including prosodic and semantic tests. We present our results in the following section.

The analysis of Teiwa wa ('say')
This section presents new evidence concerning both independent and integrated occurrences of Teiwa wa. In Section 4.1, we argue that independent occurrences of wa are full lexical verbs taking a speech report as the object, and in Section 4.2 we show that integrated occurrences of wa are complementizers that follow another speech verb and agree with its subject in person and number. Both uses are distinguished by a number of features summarized in Table 2 and will be discussed in the remainder of this section.
In our discussion we frequently refer to a number of experiments conducted with the assistance of several young Teiwa men, to whom we refer with their first names as Ben, Bimbo, Lorens, Natan and Orias. All of them are friends and live in the village of Madar.
In several experiments they acted out scenarios which included reporting each other's speech, thoughts, and intentions.

Independent occurrences of wa 'say'
In this section, we discuss uses of wa 'say' in contexts lacking another speech verb, such as (15). We show that independent wa is a full lexical verb with the interpretation 'say, think'. We imply that wa is not an evidential marker, to which speech verbs such as 'say' commonly grammaticalize (Heine & Kuteva 2002: 265 The independent wa is a transitive verb taking a speech report as its object argument and the speaker of the speech report as the subject. In other words, the analysis of wa is exactly that of say in English.

Encoding of the subject
As mentioned in Section 2.1, Teiwa does not exhibit much verbal morphology. The pronoun a immediately preceding wa in (15) is phonologically a clitic. Its occurrence is obligatory, and the form varies with the person and number of the subject, as (16) illustrates.
(16) iman i wa, saxa' ga-tof la un tei g-om me' they 3pl say chicken 3sg-egg foc cont.dst box 3sg-inside be.in 'They think that there is a chicken egg in the box.' [AM.FB2.003] The pronoun doubling in (16) encodes focus on the third person plural subject. Further details about this form can be found in (Klamer 2010a: 84-85). In Teiwa, subjects are expressed by pronouns, and prefixation is reserved for objects (Klamer 2010a: 31). Verbs that mark their subject with a prefix are rare in Teiwa: Klamer (2010a: 98) reports that there are only three verbs which follow such pattern: -o'on 'hide' (17), -ewar 'return', and -ufan 'forget'. All three roots are vowel-initial and encode events with an affected subject.
Notably, the S prefix of the vowel-intitial verbs has a reduced form consisting of a single consonant, unlike the subject agreement found with the complementizer wa (see Section 4.2.5). Furthermore, the independent wa is compatible with negation, as seen in (18), a property expected from a lexical verb but not from an evidential particle or a complementizer. The speech report verb wa can be a prosodic head, forming a prosodic word with the subject proclitic a and the subject noun Natan. The pitch contour of (19) is shown in Figure 2. [BT.EFB3.008] The intonation contour indicates that the speaker's pitch rises at the end of (20), as can be seen in Figure 3. This pattern is common for declarative questions (Van Heuven & Van Zanten 2005: p. 88) and is cross-linguistically widespread (Ohala 1984: p. 5  The question may be fragmented because in the following turn, shown in (21), the other participant first supplies a noun phrase, then after 10 seconds reiterates his answer as a full statement.
(21) yirkua, … (10s) … a wa xa'a yirkua la gom mia' turtle 3sg say this turtle top 3s-inside fill 'A turtle… She thinks there is a turtle inside.' [BT.EFB3.009] In sum, the syntactic and prosodic facts univocally demonstrate that wa is a full-fledged lexical verb, marking subject, allowing object ellipsis and compatible with negation. Next, semantic evidence is presented showing that the speech verb wa and its object form a single clause. The speech report is embedded and syntactically subordinate to the verb wa.

Semantic evidence: true reports of a false statement
We have designed an experiment to probe the truth values of the reported speech in both the non-embedded analysis proposed in (Klamer 2010a) and the embedding analysis proposed here. The non-embedded analysis predicts that the truth of the reported speech is entailed, but the embedding analysis does not make this prediction. The judgements of our primary consultants indicated that the prediction of the embedding analysis is correct. The materials consisted of short dialogues which were recorded by two of our consultants. Each test item consisted of a pair of sentences spoken by Natan and Orias. The first sentence, spoken by Natan, was either clearly correct or incorrect. Examples of both are  Participants in the experiment were presented with 30 such pairs (the first six were practice items) and instructed to report whether what Orias said was accurate. The participants responded either atab ('yes, true, correct') or man ('no').
The prediction of the non-embedding and embedding analyses for the four types of items are presented in tabular form in Table 3: The non-embedding analysis predicts that (23) should be understood as two sentences. The first, Natan a wa. ('Natan talked.'), should be judged true in all four scenarios, and therefore the truth of the second, a Bandar/Madar me 'He lives in Bandar/Madar' should determine the acceptability of the whole utterance (23). On the other hand, the embedding analysis predicts that (23) should be judged in the same way as the English translation 'Natan said that he lives in Bandar/Madar', i.e., only according to whether it is a correct report of the prior utterance.
The judgments of the speakers who assisted with the construction of the material corresponded to the embedding analysis and are confirmed in the experiment we conducted (see Supplement 1, Experiment 3). The experiment included 21 participants aged 25 to 81 years (mean age: 46.3). We did not find an effect of age. This indicates that the Teiwa grammar of independent wa has not undergone a recent change.
In sum, we argued in this section that the independent wa should be analyzed along the lines of the English verb say, a verb that takes a (usually clausal) complement. We presented experimental evidence corroborating this conclusion.

Integrated occurrences of wa
In this section, we consider cases where the sequence of a subject pronoun and wa follow another speech verb, such as -walas 'tell' in (1), repeated here as (24).
(24) Natan a na-walas a wa bas a Qalambas ma gi Natan 3sg 1sg-tell 3sg say tomorrow 3sg Kalabahi come go 'Natan told me that he will go to Kalabahi tomorrow.' [LT.CD2.008] Initially a bisentential analysis of such sequences seems plausible because the integrated wa could receive the same interpretation as the independent wa. Such an analysis would correspond to the English paraphrase Natan talked to me. He said that he'll go to Kalabahi tomorrow. But we demonstrate that the bisentential analysis cannot be correct, and that instead the monosentential analysis is correct. This entails that integrated wa must receive a different interpretation than the independent wa discussed in 4.1. We argue that wa in uses such as (24) is a complementizer.
In the following examples, integrated wa mostly occur with the speech verb walas 'tell'. The speech complement of walas 'tell' does not have to be always expressed with the complement clause linked with wa. Another possibility is to increase the valency of walas 'tell' with the oblique marker ma 'come' adding a nominal phrase describing the information, as shown in (25). 9 The basic valency of walas includes the subject (agent = teller) and the object (recipient = tellee). When used with the verbs listed above (27) integrated wa is optional (i.e., (28) is also grammatical). But these sentences are perceived to be less clear without wa, and therefore, the use of wa is preferred. In addition to the five verbs above, we found a single occurrence with ma bi'in 'praise' and multiple uses with om 'think', as in (29). 11 9 For details about the oblique marker ma, see (Klamer 2010a;b). 10 Klamer (2010a: 369, 386) shows, furthermore, that liin 'invite', soi 'order', and taxani 'inquire' combine with speech report clauses but does not state whether those can be linked with wa. 11 Klamer (2010a: 94) discusses om only as a noun within what she describes as phrasal verbs. Other experiencer verbs can be formed with om followed by an adjective or verb (Klamer 2010a: 94-95). But, the adjective in (ia) exhibits no agreement, and the verb in (ib) exhibits third person agreement while (29) shows that the subjects of wa and om agree.

Agreement of wa with the subject of the main verb
While Klamer (2010a) only reports integrated uses of wa with the pronoun a, we found that integrated wa agrees in person and number with the subject of the main speech verb. The agreement is productive and regular, as shown in (30). Integrated wa agrees with the second person plural subject of walas 'tell' in (30). In (30b), integrated agrees with impersonal hala 'others, unknown people' (for more details on hala see Klamer (2010a: 30)).
(30) a. yi'in la na-walas maan yi wa insi yi-yaf waad wan kariyan 2pl foc 1sg-tell neg 2pl say fut 2pl-house big be work 'You did not tell me that you will perhaps build a big house (for yourselves

Compatibility of wa with aspectual markers
Integrated wa cannot be preceded by an aspectual modifier such as the distal continuative un. As shown in (32b), where wa is preceded by the aspectual marker un, its interpretation is that of a speech verb and not a complementizer.
(32) a. Ben un a yivar a wa a un tuax hufa' Ben cont.dst 3sg dream 3sg say 3sg cont.dst palm.wine drink 'Ben is currently dreaming that he is currently drinking tuax (palm wine).' [LT. CD2.151] b. Ben un a yivar Ben a un a wa a tuax hufa' Ben cont.dst 3sg dream Ben 3sg cont.dst 3sg say 3sg palm.wine drink 'Ben is currently dreaming that he is talking and that he drinks tuax (palm wine).' [LT.CD2.152]

Syntactic evidence for a monosentential analysis
In English, and even more so in some other European languages, the presence of complementizers such as that provide unequivocal morphosyntactic evidence for embedding. In Teiwa, the supporting evidence for embedding comes from prohibitives and long distance extraction.

Prohibitives (negative imperatives)
Klamer (2010a: 301) describes two types of prohibitives in Teiwa. The first type is marked by the clause final gaxai 'do not'. The second type, shown in (33), contains gaxai and a positive imperative marker: ha-dan 'your part, your obligation'. The sentence's final gaxai 'do not', however, marks a prohibitive force and (33) is therefore interpreted as a prohibitive overall.
(35) Ha-dan Ben ga-walas gaxai ha wa na tabako ma iga 2sg-part Ben 3sg-tell do.not 2sg say 1sg tobacco come hide 'Do not tell Ben that I hid the tobacco.' [MM.E789.0122] The prohibitive interpretation of (35) is difficult to reconcile with the bisentential analysis, which would give rise to the reading: 'Do not talk to Ben. Say I hid the tobacco.' The monosentential analysis, on the other hand, offers a rather straightforward explanation: the clause na tabako ma iga is embedded and therefore under the scope of the prohibitive marker.

Long-distance movement
Syntactic movement from one clause to another is one of the strongest arguments for a complex monosentential structure, long-distance movement of question words being a classic case. However, it has also been established that not all languages allow long-distance movement from complement clauses -Northern German dialects, in particular, tend to not allow long-distance movement of question words from finite complement clauses (Fanselow & Weskott 2010 and others). Teiwa, like English, fronts question words in non-polar (or wh-) questions (Klamer 2010a: 280-293). Example (36) shows that question words can move from the complement of independent wa.

(36)
Yilag la Ben a wa ha g-ua? who foc Ben 3sg say 2sg 3sg-hit 'Who did Ben say you hit?' [LT.CD2.216] Movement from the complement clause of other speech verbs is difficult. In our attempts to elicit long extraction, speakers often produced examples like (37). Note that one of the arguments of ultaq 'tell' is marked as oblique with ma. This indicates that amidan 'what' in (37) is an argument of ultaq 'tell', and therefore (37) is not an example of long extraction.
(37) Amidan la Ben ma ha-ultag-an a wa bif ga'an sii? What foc Ben come 2sg-tell-real 3sg say child that bite 'About what did Ben tell you that it bit that child? ' [LT.CD2.225] Example (38), however, is also acceptable. In (38), neither of the arguments of walas 'tell' is marked with oblique case. The absence of oblique marking in (38) suggests that yilag 'who' is not an argument of walas 'tell', but is instead an argument of the verb ua 'hit'. Example (38) therefore requires the monosentential analysis of integrated wa.
(38) Yilag la Ben pi-walas a wa insi a n-ua who foc Ben 1pl.incl-tell 3sg say maybe 3sg 1sg-hit? 'Who did Ben tell us will hit me?' [LT.CD2.219] In sum, the syntactic evidence is in favor of the clausal embedding analysis for the integrated wa.

Prosodic evidence for a monosentential analysis
Klamer (2010a: 361) mentions that intonation connects juxtaposed clauses. Prosodic properties, in particular pitch movement (intonation), timing, and intensity signal sentence boundaries (Bolinger 1984;. In particular, a special intonation contour marks the end of the sentence. The question is whether the absence of a sentence final contour can then be taken to indicate subordination. While Klamer herself does not assume that intonation directly indicates sentence boundaries, Mithun (2009) shows that intonation in Mohawk distinguishes between monosentential and bisentential sequences. 12 Further, Kastner et al. (2014) argue that prosody alone can mark embedding in the emerging Kafr Qasem Sign Language. Klamer (2010a: 304) notes that Teiwa clauses are separated from each other by a falling intonation on the final word (usually a verb). We confirmed this in our data. Example (39) consists of two sentences separated by a long pause and shows the sentence final drop. In addition to the sentence final low tone (L%), (39) also shows that Teiwa exhibits pitch declination as well. In this section, we show that Teiwa speech reports with integrated wa generally exhibit a single intonation contour, as expected by the monosentential analysis. Specifically, we rely on falling intonation as a signal of a sentence boundary. The falling intonation contour, associated with the final unit, is absent before integrated wa; in fact, we see the opposite contour of the pitch rising just before wa, with an optional pause between wa and the rest of the sentence. Now consider examples of speech reports with walas 'tell'. For one, the walas a wa sequence is generally followed by a short pause, but there is no pause after walas 'tell'. The placement of the pause is not expected in the bisentential analysis, but it provides evidence for the monosentential analysis. The second argument related to pitch is based on the absence of sentence-final drops (i.e., intonation units with final L%) in the walas a wa sequence. Instead, we generally find a rise in pitch on wa. This, again, would be unexpected in the bisentential analysis, but it is consistent with the monosentential one. 13 The example in (40) from speaker MW exemplifies both of the general patterns we found. MW produced a 0.6s pause that follows a wa. Figure 4 further shows, there is a clear rise targeting the initial pitch level on a wa and the higher pitch is subsequently maintained across the pause.
(40) Natan a na-walas a wa a pemantik pin a-yaf ma gi Natan 3sg 1sg-tell 3sg say 3sg lighter(ind) hold 3sg-house come go 'Natan told me that he took the lighter to his house.' [MW.E789.18] In rapid speech, such as (41), the pause disappears, but the pitch rise on wa remains, as can also be seen in Figure 5.
(41) Natan a na-walas a wa a bai gaal Natan 3sg 1sg-tell 3sg say 3sg pig shoot.with.arrow 'Natan told me that he shot a pig.' [MW.E789.37] The intonation pattern is the same in examples where walas 'tell' is negated, as in (42). Figure 6 shows a slight pitch rise at integrated wa 'say' that is maintained even in the presence of the negator maan and targets the initial pitch level of about 95Hz. 13 In their cross-linguistic work, Hale & Selkirk (1987) point out that languages vary with respect to the prosodic phrasing of function words such as complementizers, and this variation is not fully determined by syntactic structure (see also Truckenbrodt 1999). If the monosentential analysis is correct, phrasing in Teiwa would need to be analyzed in a similar way as Kwakiutl (cf. Boas 1911) and other languages where the functional heads are always phrased with a preceding head. (42) atab, a ga-walas maan a wa, a gi tabaku ga-uyan true 3sg 3sg-tell neg 3sg say 3sg go tobacco 3sg-search 'That's true, he did not say that he went to get tobacco.' [AM. E6.148] Examples such as (43) show that the intonational contour may be the only indication of embedding. As shown in Figure 7 JL's pitch at the beginning of the sentence is roughly 250 Hz. The contour contains only one final fall to roughly 160Hz. Only at the beginning of next sentence is the pitch reset to the original level. In this example, because JL is speaking slowly, the two joined clauses are separated by a pause of roughly 0.7s separating the speech verb from its complement, analogously to (40).
(43) na Ben ga-walas Natan ga-baq muling 1sg Ben 3sg-tell Natan 3sg-body weak 'I told Ben [that] Natan is tired. ' [JL.E789.56] Finally, speaker JL also produced example (44), which consists of three clauses integrated under a single intonational contour. Figure 8 shows the contour starting at roughly 250Hz and closing at about 160Hz. There is no drop at the end of the first and second clause;  instead the pitch rises back to about 250Hz each time, signaling that the sentence is not yet finished. Figure 8 also shows that there are almost no pauses between the three clauses. In our analysis, this sentence contains two complement clauses.
(44) Natan ma na-walas Ben gawalas a un tii'-in quri Natan come 1sg-tell [Ben 3sg-tell [3sg prog lie.down-real sleepy]] 'Natan told me to tell Ben that he is going to sleep.' [JL.E789.88] Furthermore, we found one speaker, JM, a 45-year-old Teiwa female, to phonologically reduce integrated wa in spontaneous speech in (45). Figure 9 shows that JM produced the na wa sequence as a single syllable [nɑʊ]. In (45b), JM further reduced the na wa into just [nǝ].
(45) a. na ga-walas [nɑʊ] ga-baq muling 1sg 3sg-tell 1sg say 3sg-body weak 'I tell him he is tired. ' [JM. E789.51] b. na ga-walas nǝ, Natan tuax maya' 1sg 3sg-tell 1sg.say Natan wine spill 'I tell him that Natan spilled the wine.' [JM.E789.45]  Note that even JM's reduced forms still indicate person-number agreement with the subject of the main verb walas 'tell'. We found no reduced forms in the speech of other Teiwa speakers and our consultants consistently insisted on the complete a wa sequence. Therefore, JM's speech seems to represent a minority dialect where grammaticalization may have progressed. It should be noted that phonological reduction of wa does not occur in JM's speech when wa is the main verb, as shown in Figure 2. We conclude, therefore, that while JM's speech further supports our analysis of Teiwa as allowing embedding, it leaves intact our conclusion that embedding exists independent of complementizers.

Semantic evidence for a monosentential analysis
The primary semantic diagnostic for subordination is scope. The reason scope is a good diagnostic is that the way sentence meanings are combined in discourse is generally taken to amount to coordination (Stalnaker 1978;Sauerland 2016). In the following, we investigate four types of scopal phenomena: (i) the scope of negation, (ii) the scope of constituent questions, (iii) the scope of prohibitives, and (iv) the scope of polar question markers. 14

Scope of Negation
The scope of negation, which is used to distinguish between bi-and monosentential sequences, was first employed by Noonan & Bavin (1981) to study Lango. We constructed the example in (46) to investigate the scope of negation with the verb walas 'tell'. According to our primary consultants, this placement allows an interpretation in which negation takes wide scope over both walas and the following clause with integrated wa. Such interpretation is predicted only by the monosentential analysis.
(46) Natan a na-walas maan a wa a xaf kariman ol Natan 3sg 1sg-tell neg 3sg say 3sg boat small buy bisentential: 'Natan did not talk to me. He said he is buying a small boat.' monosentential: 'Natan did not tell me that he is buying a small boat.' 14 Other scopal phenomena, such as the scope of universal quantifiers, provide less conclusive evidence because of the phenomenon of telescoping (Poesio & Zucchi 1992). But telescoping does not challenge the claim that intersentential composition is always dynamic coordination, if Poesio & Zucchi (1992) are correct in their argument that telescoping involves silent universal quantifiers. This judgment is difficult to discern because the pragmatics of negation imposes a requirement on sentences such as (46) and its English translation that the content of the complement clause be salient (Horn 1989, and others). Typically, this requirement leads consultants to suggest a scenario out of the blue in which (46) could be uttered when the complement is actually true. This would correspond to the English paraphrase 'He is actually buying a small boat, but he did not tell me about it.' In this case, the complement clause would be salient because it is actually true. However, we found that our consultants also accept (46) in another scenario where the complement is salient, namely in short dialogues like (47), which involves a correction. In (47), the statement that Natan is talking about buying a small boat is made salient by a question.
(47) Q: Did Natan tell you that he is buying a small boat? A: No, he told me that he is buying a boat, but he did not tell me that he is buying a small boat.
In Teiwa, one situation we created to detect whether the wide-scope negation interpretation is available was acted out by Natan, Bimbo, and Ben. At first, only Natan and Bimbo are present, and before leaving the room Natan tells Bimbo the following: In this scenario, the following exchange between two observers, LT and AM, is felicitous and was actually observed in our fieldwork: (50) a. LT: xoran si, atab Natan a Bimbo ga-walas maan a wa, a gi then sim true Natan 3sg Bimbo 3sg-tell neg 3sg say 3sg go tabako gu-'uyan le maan? cigarette 3sg-search or neg LT: 'So then, Natan did not say to Bimbo that he went to look for cigarettes. ' [AM. E6.147] b. AM: atab a ga-walas maan a wa, a gi tabako gu-'uyan true 3sg 3sg-tell neg 3sg say 3sg go cigarette 3sg-search AM: 'Yes, he did not tell him that went to look for cigarettes. ' [AM.E6.148] In (50), both LT's question and AM's answer contain walas 'tell' followed by negation. The truth of (50b) in this scenario argues against the bisentential analysis since it predicts that (50b) will be understood as a conjunction or juxtaposition of two sentences. Only the first of these sentences would be negated, and the meaning of this statement would be roughly Natan did not talk to Bimbo. The second would have the positive meaning Natan said he went to look for cigarettes, which is also predicted to be false in the scenario. Only the monosentential analysis predicts AM's utterance to be felicitous because negation can then take scope over both clauses in (50b). The same reasoning also applies to the question (50a): Based on the monosentential analysis, the question should be interpreted as Did Natan not talk to Bimbo?, and it would therefore be answered negatively. Furthermore, the putative second sentence that makes up (50b) would be the claim Natan said he went to look for cigarettes, which is false in the scenario. The felicity of the positive response atab in (50b) argues, therefore, that the question is not interpreted as the bisentential theory predicts, but receives a wide-scope negation interpretation that is only consistent with the monosentential analysis. We confirmed these intuitions with informants and in an experiment (reported in the appendix) with 10 additional Teiwa speakers using six different items.

Scope of Constituent Questions
A second diagnostic for a complex monosentential structure which is widely used in language acquisition (e.g., de Villiers 1995) involves long constituent questions. When the interrogative word remains in situ, the distinction between a long monosentential interpretation and a bisentential, short-question interpretation is only semantic. A relevant long question in Teiwa would be (51), where the interrogative ita'a 'where' occurs in a sentence with an integrated wa and the attitude verb om.

(51)
Bimbo a om si a wa bui un ita'a la mia'? Bimbo 3sg inside sim 3sg say betelnut cont.dst where foc put bisentential: 'Bimbo was thinking. Where did he say that the betel nut is?' monosentential: 'Where does Bimbo think the betel nut is?' [MOT. E6.080] While the distinction between the two readings in (51) is rather subtle, a clearer distinction arises in (52), which must be interpreted as double embedding. Such interpretation is forced by the second wa: it cannot be integrated because its subject is Natan, not Bimbo. Teiwa speakers prefer the structure in (51) which suggests that the complementizer wa is left out if it is followed by the speech verb wa.
(52) Bimbo a om si (a wa) Natan a wa bui un ita'a Bimbo 3sg inside sim (3sg say) Natan 3sg say betelnut cont.dst where la mia'? foc put bisentential: 'Bimbo was thinking. Where did Natan say that the betelnut is?' monosentential: 'Where does Bimbo think Natan says the betel nut is?' Different answers for question (52) are predicted in the scenario where Bimbo is misinformed about what Natan believes. For example, Bimbo believes that Natan thinks the betelnut is in the basket, but Natan actually thinks it's in the box. In this scenario, the bisentential analysis predicts the answer in the box, but the monosentential analysis predicts in the basket. According to our primary consultants, the prediction of the monosentential analysis is correct. We successfully confirmed this judgment in an experiment with 6 different items and 10 additional Teiwa speakers reported in the supplement as Experiment 6.

Scope of Prohibitives
Teiwa prohibitives were introduced in 4.2.4. We argue that prohibitive provides additional evidence for the monosentential analysis of integrated wa. Consider the example in (53). The monosentential analysis predicts (53) to express a prohibition to tell the content following wa as in the English translation provided for (53). But the bisentential analysis predicts that the prohibitive should only take scope over the first verb, and the predicted interpretation should correspond to the paraphrase Do not talk to Ben, but say that I hid his lighter. According to our consultants, only the wide-scope prohibitive interpretation predicted by the monosentential analysis is available. We confirmed this judgment with 10 Teiwa speakers and report it in the appendix as Experiments 7, 8, and 9.

Scope of Polar Questions
Polar questions provide the last scope diagnostic for distinguishing between the monosentential and bisentential analysis. Klamer (2010a: 277-278) reports that polar questions in Teiwa are not marked by word order or by a special particle. Alternative questions, such as (54a), contain the sentence final le maan 'or not' and expect a positive or negative answer. In the following, we use the term polar question to describe this structure and distinguish it from alternative questions formed with le 'or', such as (54b).
(54) a. A xoran si yi ga-sar le maan? 3sg thus sim 2pl 3sg-find or neg 'If so, did you find her or not?' (Klamer 2010a: p. 279) b. Ina tau le yed? eat prf or prsp 'Have you eaten already or not yet?' (Klamer 2010a: p. 279) Polar questions may contain the negator maan. Importantly, the regular negator maan will immediately follow the speech verb, as in (55). The bisentential and monosentential analyses of integrated wa make different predictions for polar questions like (55).
The predicted interpretations are shown in (56a) and (56b), respectively. The bisentential analysis predicts no relationship between the negator maan following the speech verb and the sentence final le maan, while the monosentential analysis allows for an interaction.
(55) Bimbo a Ben ga-walas maan a wa Natan ga-warax sin le maan? Bimbo 3sg Ben 3sg-tell neg 3sg say Natan 3sg-wait first or neg (lit.) 'Bimbo did not tell Ben -he said to wait for Natan -or not?' [DSK.E789.088] (56) a. bisentential prediction: 'Bimbo did not talk to Ben. Did he tell him to wait for Natan?' b. monosentential prediction: 'Did Bimbo not tell Ben to wait for Natan?' The two accounts predict different responses to (55). In this experiment, Natan had to leave and asked Bimbo to pass a message to Ben to wait for him. In one scenario, after Ben arrived, Bimbo passed him Natan's message truthfully. In the alternative scenario, Bimbo made up a different message. The experiment participant, who witnessed the exchange, was then quizzed about what happened with sentences such as (55).
The bisentential account predicts (55) to be odd since it consists of a false assertion followed by a question inconsistent with that assertion. The prediction of the monosentential analysis depends on one additional factor: the way negative polar questions are answered in Teiwa. Well-studied languages like English and Japanese differ with respect to this property (Kuno 1973): In response to the question Isn't John running?, the English positive response yes entails that John is running. But the Japanese positive response hai entails that he is not running, and the negative iie must be used if John is actually running.
Our evidence indicates that Teiwa behaves like Japanese: In response to (55) in the scenario introduced above, the answer maan 'no' is the appropriate one. Again, we collected two types of data to corroborate this empirical claim: we constructed sentences that present (55) in the scenario mentioned above and tested them experimentally with 10 subjects (age-range: 16-65 years; mean age: 37). The design and results are reported in the appendix as Experiment 6 and confirmed that, indeed, maan 'no' is the felicitous response in the critical scenario and it is predicted by the monosentential analysis.

Conclusions
We established a contrast between the speech verb wa and complementizer occurrences of wa following other speech verbs: the speech verb wa can be combined directly with a finite clause complement, as in (57a). Other speech verbs can also embed finite clauses, but Teiwa speakers prefer them to be linked with the agreeing complementizer wa, as in (57b). In both cases, wa is prosodically integrated with the rest of the sentence, but only the complementizer wa may undergo phonological reduction for some speakers. Unlike the speech verb wa, the complementizer is not compatible with aspect or negation and must agree with the subject of the main verb. Semantic integration of the complement clause was demonstrated using a number of tests, including (i) the scope of negation, (ii) the scope of constituent questions, (iii) the scope of prohibitives, and (iv) the scope of polar question markers. The traditional analysis of speech and attitude reports assumes that a speech verb V and a finite clause CP stand in verb-complement relation in the syntax and in predicate-argument relation in the semantics (Hintikka 1962, and others). More recent proposals suggest neither a syntactic verb-complement relation nor a semantic predicate-argument relation have been obtained between V and CP (Aboh 2005;Arsenijević 2009;Kastner 2015;Moulton 2015;Elliott 2017). Our conclusion is that complementation can exist independently of complementizer morphemes (see also Kastner et al. 2014). Finally, our paper has implications for the discussion of complementizers and complementation in historical linguistics. In Section 3 we compared Teiwa with languages such as Old Babylonian, Nanti, and Lukubusu, and argued that wa may be in an early stage of complementizer development, but complementation precedes this development in Teiwa.
The presence of embedding in the Teiwa cases opens a new perspective on the discussion of structures in Akkadian and also Pirahã (Everett 2005) where parataxis has been proposed. For Akkadian, the types of evidence we discussed in Section 4.2 are likely to never be available (clearly so for prosodic evidence). For Pirahã, Sauerland (2017;2018) shows using a similar experimental designs that Pirahã speakers treat speech reports as semantically integrated with the speech verb and subordinated. Sauerland (2017) shows that Pirahã clause linker sai shows pitch differentiation of its conditional and nominalizing usage. Pirahã like Mohawk (Mithun 2009) represents a type of sentence integration detectable only at the semantic and prosodic level, which has to precede all subsequent stages of grammaticalization. 15 Semantic (sem) and prosodic (pros) integration is the necessary condition that allows for other grammatical signals of integration to develop. During the grammaticalization polysemy and marking of subjects (i.e. subject agreement, agr) gradually disappears. Table 4 schematically represents the attested stages of this process.