Languages vary in whether evidentialsâthat is, linguistic markers of information source (Aikhenvald 2004)âcan appear in clausal complements. Embedding of evidentials is possible in some languages, such as Turkish (Turkic: Turkey; Ćener 2011) (evidence for treating such clauses as instances of genuine embedding is provided in section 3).1In some other languages, such as Cuzco Quechua (Quechuan: Peru; Lefebvre and Muysken 1987), evidentials are banned from clausal complements.The semantic literature often views the variation in embeddability of evidentials as evidence for the semantic heterogeneity of evidentiality as a category (Faller 2002, 2007, Garrett 2001, Matthewson, Davis, and Rullmann 2007, McCready and Ogata 2007, Peterson 2010). Specifically, embeddability, and the contrast between (1) and (2), has been regarded as a diagnostic of the semantic status of respective markers. Embeddable evidentials, such as Turkish miĆ (Ćener 2011), have been treated as epistemic modals. Nonembeddable evidentials, such as Cuzco Quechua mi (Faller 2002), have been analyzed as dealing with speech acts. However, the assumption that nonembeddability of evidentials indicates their speech act status can be challenged on theoretical grounds alone. A growing body of syntactic (Aelbrecht, Haegeman, and Nye 2012, Haegeman 2012, 2014) and semantic (Krifka 2014) literature argues that speech act material is in fact not limited to root clauses, so the behavior of evidentials needs further scrutiny (see Korotkova 2017, Thomas 2014).In this squib, I engage with the very idea of the semantic heterogeneity of evidentials, and the larger question of semantic variation (Bochnak 2013, von Fintel and Matthewson 2008, Matthewson 2001, 2013). I show that even though (non)embeddability is a matter of crosslinguistic variation, it is not a case of genuine semantic variation in evidentiality. This result provides further empirical support for the view that evidentials have a unified semantics across languages (Korotkova 2016, 2019, Matthewson 2011, 2012, Murray 2010, 2017). According to this view, the existing variation in evidentiality is not semantic. Drawing on data from Turkish, where evidentials can appear in tensed but not in nominalized complements, I propose that restrictions on the embedding of evidentials are due to the syntax of clausal complementation. It it has been suggested before that (non)embeddability is due to morphosyntax (Matthewson 2012, Murray 2010, 2016, 2017), but this claim has not been argued for in detail. To this end, I offer the first detailed investigation of syntactic constraints on evidentials in clausal complements and put forth the following generalization: evidentials are embeddable only in those languages that have complements with enough structural space to host them. In Korotkova 2016, I discuss other arguments against the semantic heterogeneity view, such as the fact that the variation is not unidimensional.The next two sections present core data from Turkish. Section 2 provides a brief description of complementation strategies, and section 3 focuses on the distribution of evidentials across those strategies. Section 4 examines a broader crosslinguistic picture.Like any Turkic language, Turkish has a variety of complementation strategies (George and Kornfilt 1981, Kornfilt 1997). The scope of this squib is restricted to tensed and nominalized complements, described in detail below (tensed and nominalized are descriptive labels; see section 3.2 on tense in nominalizations).2 They differ both in morphological makeup and in clausal architecture, and there is no one-to-one mapping between those strategies and the embedding predicateâs semantic type.Tensed complements with nominative subjects, exemplified in (3), are the closest counterpart of that-clauses in English. This complementation strategy is available to speech predicates, such as demek âsayâ and söylemek âtellâ, and to many predicates of mental attitude, such as dĂŒĆĂŒnmek âthinkâ and unmak âhopeâ. In terms of external syntax, the complement, being an object, typically precedes the verb. The internal syntax is the same as in root clauses: (a) verb-final word order, (b) nominative-accusative case alignment, (c) full verbal morphology, including tense and agreement.Although the complement in (3) resembles a root clause, it instantiates genuine subordination. This is evidenced by the interpretation of the in-situ wh-phrase in (4) as a matrix question, which would be ruled out if the bracketed clause were quoted or a parenthetical.Other tests for the subordinate status of such clauses include long-distance licensing of polarity items and the availability of de re construals (Ăzyıldız 2012, Ćener and Ćener 2011). Optional with bilmek âbelieveâ and absent with demek âsayâ, the complementizer diye, derived from demek (Ăzyıldız, Major, and Maier to appear), is obligatory with some predicates, such as dĂŒĆĂŒnmek âthinkâ and duymak âhearâ. Its presence does not affect the syntax of the embedded clause.Tensed complements with accusative subjects, exemplified in (5), are reminiscent of English exceptional case-marking constructions. This complementation strategy is available to a subset of mental attitude predicates that take complements discussed in section 2.1, such as unmak âhopeâ and sanmak âbelieveâ. Speech predicates are excluded. Such complements, ungrammatical at the root level (6), precede the verb and have root temporal morphology, but, in contrast to the construction in (3), have accusative subjects and lack agreement.The type of syntactic domain in such clauses, as well as the structural source of the accusative case, is a matter of debate (Aygen 2002, Kornfilt 1984, 1996, 2007, Zidani-EroÄlu 1997). For my purposes, it is sufficient to say that those are clausal complements with a special syntactic status and, unlike their English counterparts, regular verbal tense.Turkish makes productive use of nominalizations (7), which differ substantially from root clauses in the following respects: they (a) are case-marked by the matrix verb, (b) have genitive, instead of nominative, subjects, and (c) have possessive agreement morphology.Nominalizations are used with a wide range of predicates. For some, such as bulmak âdiscoverâ, nominalization is the only available complementation strategy. There are also many predicates that take both nominalized and tensed complements: for example, unmak âhopeâ, söylemek âtellâ, beklemek âexpectâ. Some of those predicates (e.g., bilmek âthink/knowâ and hatırlamak ârememberâ) trigger what Ăzyıldız (2017) calls the factivity alternation: the interpretation is nonfactive with tensed complements (8a) and factive with nominalized ones (8b), as evidenced by the fact that nominalizations cannot follow up a contradicting claim.For an analysis of the pattern, see Ăzyıldız 2017. Of importance here is the fact that nominalizations are inherently epistemically neutral. There are many predicates (e.g., söylemek âtellâ and dĂŒĆĂŒnmek âthinkâ) that do not participate in the alternation and produce no semantic difference between the tensed and the nominalized complement. For example, the factivity inference is absent in the noncontradictory example (9).As is common in the Anatolia-Balkans-Caucasus region, evidentiality in Turkish is part of the tense system and is expressed by perfect morphology (Izvorski 1997). I concentrate on the behavior of miĆ, an indirect evidential denoting hearsay (10a) or inference (10b).When used as a perfect, but not as an evidential, miĆ is incompatible with temporal frame adverbials such as âyesterdayâ or âat five oâclockâ, a typical property of perfects across languages (Klein 1992, Pancheva and von Stechow 2004). In what follows, I will use such adverbials to disambiguate between the aspectual and the evidential interpretation.PaceJohanson 2000, miĆ can be embedded. In (1) and (12), the adverbial, which modifies the lowerâbut not the matrixâclause due to its tense, ensures the evidential interpretation.The presence of miĆ does not affect the syntactic status of the clause ((13), (14)).3 For example, wh-in-situ can be interpreted as a matrix question in nominative-subject complements with miĆ (13) and without it (4). Likewise, accusative-subject complements are transparent for anaphor binding (Kornfilt 2007), and miĆ does not affect it (14).However, in contrast with tensed clauses, miĆ is banned from nominalizations.If the ungrammaticality of (15) were due to a morphological quirk, one would expect no embedding differences between the evidential and the aspectual guise of miĆ. This expectation is not borne out: miĆ can appear inside a nominalization as a perfect (16).4To recapitulate, the evidential miĆ can occur in tensed complements but not in nominalizations. Schenner (2010), the first to notice the ban on miĆ in nominalizations, advocates a semantic story in the spirit of Hooper and Thompson 1973. Evidentiality is argued to be semantically compatible only with nonfactive attitude predicates, which, according to Schenner, are the ones not taking nominalizations. This view assumes a strict mapping between the verb, complement type, and factivity. As demonstrated in section 2.3, there is no such mapping. The factivity inference with nominalized complements is present with some, but not all, predicates. In particular, dĂŒĆĂŒnmek âthinkâ does not give rise to it (9). The evidential miĆ is licensed inside tensed (12) but not nominalized (15) complements of dĂŒĆĂŒnmek. The analysis in Schenner 2010 does not predict the observed contrast.I propose a syntactic explanation of the pattern: nominalizations lack structural space to host evidentiality. To this end, I adopt Borsley and Kornfiltâs (2000) view of nominalizations in Turkish and remain agnostic about the exact location of the evidential miĆ in the clausal spine. It is sufficient to say that evidentiality must be introduced high.Nominalizations across languages are traditionally described as verbal nouns. They are often analyzed as mixed categories headed by a nominal projection whose place in the structure determines the degree of âverbinessâ of the entire complex (Kornfilt and Whitman 2011b and references therein). Turkish indicative nominalizations have been argued to be verbal all the way to TP, at which level a D head is introduced (Borsley and Kornfilt 2000, Kornfilt and Whitman 2011a). Arguments for this view are presented below.First of all, the verbal spine cannot be truncated lower than vP, as evidenced by object case and modifiers. Turkish nominalizations, just like regular verbs, assign the accusative case to objects. They also take verbal modifiers, including temporal adverbials, such as dĂŒn âyesterdayâ, and adverbs, such as dikkatlice âcarefullyâ, which cannot modify nouns (cf. dikkatli(*ce) insan âcareful(*ly) personâ). On the other hand, there is evidence from agreement and subject case that the nominalizing does not happen higher. Nominalizing morphology, dik (nonfuture) and açak (future), occupies the tense slot. Morphology that is linearly to the right of dik and açak is nominal. Importantly, just like regular nouns, nominalizations use possessive, and not verbal, agreement morphology. Turkish is a language where finiteness, understood as the ability to assign the nominative case, has been linked to agreement and/or mood (Aygen 2002, Kornfilt 2007). The lack of verbal agreement is thus responsible for the genitive, not nominative, subject case in nominalizations. Finally, the overall complex has the distribution of a noun. It can serve as an argument to postpositions, a property that regular tensed clauses do not have, and is case-marked. (17) illustrates these defining characteristics of Turkish nominalizations.The nominalization in (17), formed by the marker dik, exhibits the following verbal properties: (a) object case (aÄaça (ACC), not aÄaçın (GEN)) and (b) modifiers (dikkatlice (adverb), not dikkatli (adjective)). Its nominal properties are (a) agreement (possessive in, not verbal ) and (b) subject case (Bestenin (GEN), not Beste (NOM)). The nominalization also bears the accusative case assigned by the matrix verb.This discussion of Turkish nominalizations makes the following point: nominalizations have some, but crucially not all, structural layers of root clauses. I argue that this is critical to understanding the distribution of the evidential miĆ in clausal complements. I propose that the behavior of miĆ can be explained if its syntactic position is taken into account and that nominalizations lack precisely those layers where miĆ is introduced.The argument proceeds as follows. First, following Speas (2004, 2010), I assume that evidentiality, and point of view in general (Speas and Tenny 2003, Sundaresan 2018, Zu 2018), is part of the left periphery.5 Second, the analysis of Turkish nominalizations presented above entails that they lack the left periphery because the cutoff of the verbal spine happens lower. I propose that the evidential miĆ is banned from nominalizations because they lack structural space to host it. The contrast between Turkish nominative-subject complements and nominalizations falls out naturally under this proposal. Only the former have the internal structure for evidentiality. It is expected that miĆ is licensed in nominalizations in its aspectual guise (16): being below TP, aspect is preserved in nominalizations.The contrast between nominative-subject tensed complements and nominalizations may suggest that the evidential miĆ is only licensed in finite clauses. Such a claim would be inaccurate. Nonfinite accusative-subject complements also license miĆ (12b). The correct empirical generalization is that miĆ is only licensed in tensed complements, both with nominative and with accusative subjects. Leaving an articulated syntactic analysis of evidentiality in Turkish for the future, I draw the following theoretical conclusion: miĆ is introduced high in the structure, and high layers are only present in tensed complements.Evidence that both types of tensed complement share at least some high projections, absent from nominalizations, comes from epistemic modality. Epistemic but not root (e.g., deontic and ability) modals are often argued to be structurally high (Cinque 1999, Hacquard 2010; though see Rullmann and Matthewson 2018). The Turkish possibility modal ebil can have both a root and an epistemic interpretation. In embedded clauses, the latter is available in tensed complements with nominative (18a) and accusative subjects (18b), but not in nominalizations (19). Both (18a) and (18b) can have an ability interpretation (âEstragon thought that Godot was able to comeâ) and an epistemic one (âEstragon thought that Godot would possibly comeâ). In nominalizations, only the ability interpretation is available.As with miĆ (16), this is not a morphological quirk. Ebil can show up as a root but not as an epistemic modal in the same morphosyntactic environment. These facts illustrate that nominalizations lack certain layers that are present in both types of tensed complement. A proper understanding of what those layers are requires more research on the fine-grained structure of the left periphery in Turkish. Below I outline several analytical options.One possibility is that evidentiality in Turkish is introduced at the same layer as speech act information, which would be a syntactic counterpart of Fallerâs (2002) analysis of evidentials as speech act modifiers. Another possibility is that evidentiality has a designated projection lower than the speech act layer, as Cinque (1999) proposes (see Simeonova and Zareikar 2015 on an implementation for Azeri, Bulgarian, and Persian). Recent research also argues for syntactic encoding of perspective above TP but below speech acts; see, for example, Zu 2018 on egophoric agreement in Newari (Tibeto-Burman: Nepal) and Sundaresan 2018 on perspectival anaphora in Tamil (Dravidian: India). To this end, it may be the case that the syntax of evidentiality is similar to that of other perspectival expressions, as Speas and Tenny (2003) argue (see Irimia 2018 for an implementation of this view for Romanian). Finally, evidentiality and epistemic modality may be structurally bundled together, which is compatible with Ćenerâs (2011) account of Turkish and which would be a syntactic correlate of Izvorskiâs (1997) analysis of evidentials as epistemic modals.Given that miĆ can occur in accusative-subject complements, it is unlikely to be introduced at the topmost layer, as such complements do not have the full verbal spine. At the same time, under any of the above options it would not be surprising that miĆ only appears in complements of speech and doxastic predicates. Such predicates are known to be more permissive in general: many phenomena are more likely to appear in their complements than in the complements of factives, for example; see the implicational hierarchy in Cristofaro 2003, Culy 1994, and Noonan 1985. In the cartographic tradition, this translates into the presence of higher projections in the complements of speech predicates and doxastics but not in the complements of other clause-taking predicates (cf. Deal 2017, Sundaresan 2018, Zu 2018). To establish the structural position of miĆ, one needs to look at other phenomena with limited distribution in embedded environments. In English, such phenomena include high adverbs (frankly, sincerely) and the so-called root transformations, such as negative inversion and argument fronting (Haegeman 2012). In Turkish, one case of interest would be constraints on indexical shift (Ăzyıldız, Major, and Maier to appear; see also Major 2019 on the genetically related Uyghur). Crucially, if one were to argue that miĆ occupies the same projection as speech act modifiers, epistemics, or other perspectival expressions, independent evidence on the syntactic distribution of those high elements would be needed.The data from Turkish offer a new perspective on the crosslinguistic variation in embeddability of evidentials. I argue above that the distribution of the evidential miĆ in embedded environments is due to the syntax of clausal complementation. In this section, I generalize this proposal to other languages, unrelated genetically and geographically.I propose that embeddability of evidentials in a given language depends on the availability of complements that can host them. For example, it is possible to recast in syntactic terms the nonembeddability of Cuzco Quechua evidentials, long considered to be rooted in their semantics (Faller 2002, 2011). Languages of the Quechuan family only have nominalized complements (Cole and Hermon 2011, Lefebvre and Muysken 1987). Their syntax is similar to the syntax of Turkish nominalizations: Quechua nominalizations do not have the full verbal spine (Cole and Hermon 2011). I conjecture that the ungrammaticality of (2) is explained not by semantics (pace Faller), but along the lines advocated for Turkish in section 3.2: Quechua nominalizations lack layers where evidentials are introduced. The contrast in embeddability of evidentials in Turkish and Cuzco Quechua is then due to the fact that Turkish, but not Cuzco Quechua, has a counterpart of that-clauses.6The following languages, like Turkish, allow evidentials in clausal complements (ALLOW-languages): Bulgarian (South Slavic: Bulgaria; Sauerland and Schenner 2007), Georgian (South Caucasian: Georgia; Boeder 2000, Korotkova 2012), Standard Tibetan (Tibeto-Burman: China and Nepal; Garrett 2001), StâĂĄtâimcets (Salish: British Columbia, Canada; Matthewson, Davis, and Rullmann 2007), Tagalog (Austronesian: Philippines; Kierstead 2015, Schwager 2010). All these languages have complements with the morphosyntactic makeup of root clauses. On the other hand, a preliminary investigation of BAN-languagesâones that have been reported to ban evidentials in clausal complementsâreveals that complements in such languages have a dedicated morphosyntax. For example, only nominalizations are available in Ecuadorian Siona (Western Tucanoan: Ecuador; Bruil 2014) and Tariana (North Arawak: Brazil; Aikhenvald 2003, 2006). Furthermore, in some BAN-languages complements feature verb forms with dependent mood marking or reduced categorial distinctionsâfor example, in Abkhaz (Northwest Caucasian: Abkhazia (de facto and partially recognized) / Georgia; Chirikba 2003), West Greenlandic (Eskimo-Aleut: Greenland, Denmark; Fortescue 2003), and Maricopa (Yuman: Western United States; Aikhenvald 2004, Cristofaro 2013). In the Algonquian language Cheyenne (Montana, United States), evidentials in fact occupy the same slot as dependent moods and are thus banned not just from complements, but from all subordinate clauses (Murray 2016).Summing up, the BAN-languages simply lack complements that would have high projections. The contrast between BAN-languages and ALLOW-languages is then best defined as a difference in the inventory of complementation strategies.7 Note that it is not a given that evidentials occupy the same structural position across languages (see Blain and DĂ©chaine 2006, 2007) or even within one language (Speas 2004). Their embeddability would still depend on the availability of suitable embedders. That is, if an evidential is introduced at XP, the language needs to have complements with XP to embed it. In this scheme, X may vary across languages. This is not to say that constraints on evidentials in complement clauses lack semantic underpinnings altogether. For example, evidentials may turn out to be sensitive to the semantics of the matrix predicate (see Anand and Hacquard 2013 on the distribution of epistemics in attitudes). However, for the type of data I discuss, we can get a lot of mileage out of syntactic facts alone.Much empirical research is needed on the crosslinguistic behavior of evidentials in embedded environments. In this squib, I show that the crosslinguistic variation in embeddability of evidentials may be reduced to the variation in the syntax of clausal complementation. By offering a syntactic solution to the puzzle that has often been viewed as semantic, I make a larger point: semantic theories will be ultimately simpler when the variation is attributed to external, independently motivated factors. As a parallel, consider modal stacking in Germanic. Combinations like must can, restricted in English (Collins and Singler 2015, Hasty 2012), are standard in German. The difference is parameterized in the syntax (Wurmbrand 2003), not the semantics, of modality. Showing that there is no need to postulate semantic heterogeneity to explain the embedding facts, this squib argues for a similar division of labor between syntax and semantics in the evidential domain.I would like to thank Natasha Abner, Pranav Anand, Beste Kamali, Hilda Koopman, Deniz Ăzyıldız, Roumi Pancheva, Yael Sharvit, Vesela Simeonova, Dominique Sportiche, Tim Stowell, Maziar Toosarvandani, Igor Yanovich, and audiences at Carnegie Mellon, ILLC, Konstanz, Stanford, TĂŒbingen, UCLA, and Utrecht for comments and discussion. I am also grateful to the Squibs and Discussion editors and two anonymous reviewers for their feedback. This research was funded in part by the Alexander von Humboldt Foundation and by the SFB 833, âConstruction of Meaningâ (TĂŒbingen). All errors are mine.