There are varying views about the directionality of agreement, with some researchers arguing for strictly downward probing (Chomsky 2000; 2001; Preminger 2013; Rudnev 2021), cyclic downward probing with reprojection of the probe (Béjar & Rezac 2009; Clem 2021), or upward probing (Zeijlstra 2012; Wurmbrand & Haddad 2014; Bjorkman & Zeijlstra 2019). This paper adopts a bidirectional agreement model proposed by Arregi & Hanink (2021).13 In a bidirectional agreement model, the probe first searches down in its c-command domain. If it fails to find an appropriate goal, it probes upwards upon additional merge of other elements until it finds its goal.

Arregi and Hanink motivate bidirectional agreement from the complementizer agreement facts of Washo (Hokan/isolate). Complementizer agreement is especially interesting in the context of agreement directionality due to its crosslinguistic variation. Many Germanic languages show the embedded complementizer agreeing with the subordinate subject (van Koppen 2005), while in Lubukusu it agrees with the superordinate subject (Diercks 2013). In Washo, the embedded complementizer shows switch-reference marking: it is realized as the suffix -š when the superordinate and subordinate subjects denote different entities (different subject or DS marking), while it is phonetically null when the two subjects are co-referent (same subject or SS marking). (19a) and (19b) demonstrate the DS and SS pattern, respectively.14

Given that the complementizer is sensitive both to the identity of a structurally higher subject and a structurally lower one, Arregi and Hanink adopt a bidirectional model of agreement conditioned as in (20).

Upon merge, the unvalued F′ searches downward into its c-command domain. If it fails to find an appropriate goal, or if it is specified for multiple agreement as is the case of the Washo complementizer, the probe searches upwards until it finds one.

I adopt (20) in analyzing Korean honorification. In Section 4.2 I demonstrate that (u)si is realized in a position that c-commands the subject, while (su)pni is realized in a position that is c-commanded by the addressee representation. This means that in Korean, subject honorification involves downward probing, while addressee honorification involves upward probing. I provide a principled account for both types of honorification under the bidirectional agreement model.

This paper joins a recent line of research that analyzes addressee honorification as a type of allocutive agreement (Oyharçabal 1993; Miyagawa 2012; 2017; 2022; Zu 2017; Alok & Baker 2018; Alok 2020; 2021; Ceong & Saxon 2020; McFadden 2020; McFadden & Sundaresan 2021). The term allocutive agreement refers to the verb or auxiliary agreeing with the addressee, as shown in the Basque example in (21).

Addressee honorification can be seen as a type of allocutive agreement. In (21), the Basque auxiliary reflects the gender of the addressee and the formality of the utterance situation. Addressee honorification reflects the relation of the speaker and the addressee as perceived by the speaker.

A peculiarity of allocutive agreement is the absence of an overt goal. Linguists have had the intuition that allocutive agreement is agreement “with the addressee”, but the addressee is not overtly pronounced (arguably barring vocative expressions). In providing a syntactic explanation, researchers have postulated functional projections that represent discourse participants. Speas & Tenny (2003) put forth a pioneering approach by drawing a parallel between the argument structure of the event denoted by vP and the argument structure of the speech event. They implement this intuition by postulating the saP (speech act phrase) which includes representations of the speaker and addressee.

Subsequent studies have recognized the advantage of postulating a speaker/addressee representation in the syntax. The position of the representation varies among researchers. Haegeman (2004); Haegeman & Hill (2013); Miyagawa (2012; 2017; 2022); Zu (2017); Portner et al. (2019a; b); Yamada (2019); Kaur (2020); McFadden (2020); McFadden & Sundaresan (2021) all posit some representation of context or speaker/addressee at the root of the clause. Haddican (2018) and Sigurðsson (2004) assume a representation lower down at the IP/FinP area in explaining Basque cliticization and Icelandic facts, respectively. Alok (2020; 2021) argues for both the FinP and the root to bear representations of discourse participants, and adopts such a structure in deriving the crosslinguistic variation in allocutive agreement with a focus on Magahi.

Korean shows addressee honorification only in the matrix clause. This makes the language amenable to positing a speaker/addressee representation at the root. For concreteness, I postulate a Speaker Phrase (SpkP) and Addressee Phrase (AddrP), borrowing from Miyagawa (2012; 2017; 2022) and Zu (2017) – although I view many of the other proposed structures compatible with theirs. Adopting the feature system put forth by Harley & Ritter (2002), the identities of the speaker and the addressee are represented using the part(icipant) and auth(or) features.15

A special mention is due of Portner et al. (2019a), who also focus on analyzing Korean addressee honorification. The authors postulate two functional projections at the clause periphery: SentMoodP and cP (the c stands for context). SentMoodP is similar to ForceP: it encodes information about the clause type. It bears the s-mood feature that can be specified as decl (declarative), interr (interrogative), etc. cP bears the feature status, which encodes information about the speaker-addressee relation, and the feature formal, which encodes the formality of the utterance. There are five possible values of the status feature (S ≥ A, S > A, S = A, S < A, and S ≤ A, where S stands for Speaker and A for Addressee), and two possible values of the formal feature (+ or –).

Crucially, they view the clause-final morphology of Korean as a single morpheme that realizes the combination of the status, formal, and s-mood features. In other words, the SentM and c heads are realized as one vocabulary item. (23) shows some examples of how the feature combinations are realized in their system, and (24) shows the structure of a sentence containing supnita. (Examples are cited from Portner et al. 2019a: 10–11.)

Portner, Pak, and Zanuttini’s analysis effectively captures the complex interaction of addressee honorification and clause type. Their analysis addresses the polite/intimate register eyo and e, which I set aside in this paper (see Section 2.2 for a brief discussion of eyo). They also straightforwardly explain why addressee honorification is unembeddable in Korean, by limiting the distribution of c to the matrix clause. My analysis presented in Section 4.2 shares some ideas with theirs, including a discourse-related representation limited to the root. However, I present two reasons below on why I diverge from their proposed structure.

First, treating supnita as one morpheme undermines its morphological transparency. As shown in Table 1 and (25), ta and kka are independent clause-type morphemes that can stand alone without (su)pni. Treating supnita and supnikka as single morphemes on a par with ta/kka does not accurately reflect the independence of ta and kka nor the fact that supni is simply attached to these particles when the addressee is of high status.

Second, it is expected under (23) that supnita and eyo cannot co-occur in a sentence because they realize the same set of features at the same syntactic positions (SentM and c). They only differ in the value of the feature formal. However, it is certainly possible to see both supni and eyo in a sentence.

According to An (2022: fn 19), sentences like (26) reflect an older stage of Korean where a speaker would use both supni and eyo to express extreme deference towards an addressee.17 While the supni-ta-eyo construction sounds rather archaic to the contemporary speaker, it is still productively used in an often comical or sarcastic tone. (26) provides evidence that supnita and eyo are not of equal morphosyntactic status, contra what is suggested in (23).

These two issues are straightforwardly addressed in my proposed structure in (22). There, (su)pni is realized at Agr_Addr and the clause-type morphemes ta/kka at Force. Since they are realizations of two separate heads, and the presence of Force is not dependent on the presence of (su)pni, the syntactic and morphological independence of the two morphemes is captured in (22). As for (26), a good candidate for the position of eyo is the Addr head itself. This position mirrors the linear order of the elements supni-ta-eyo. It is also in line with Ceong and Saxon’s (2020) proposal that eyo be linked to a higher structural position than (su)pni. In Section 4.2.2, I demonstrate an analysis of Korean addressee-honorific (su)pni as allocutive agreement based on the structure in (22).

Following Chung (2009), Choi & Harley (2019), Yamada (2019), Ikawa & Yamada (2021), and Ikawa (2022) among others, I adopt the Distributed Morphology framework (Halle & Marantz 1993; 1994) to explain the effect of honorification on the morphology of case markers and the verb. In other words, I assume that the syntax operates on formal features instead of lexical items. After spell-out to the interfaces, vocabulary items are inserted to the terminal nodes that bear the features. Insertion of vocabulary items is regulated by Vocabulary Insertion rules (VI rules), which map a set of features to a vocabulary item. In accounting for Korean honorification within the Distributed Morphology framework, the concept of contextual allomorphy is useful (Bobaljik 2000; Marantz 2013). When the realization of a feature set is subject to contextual allomorphy, the feature set can be mapped to different vocabulary items depending on the environment surrounding the node bearing the feature set. For example, the Korean nominative case marker shows honorification-dependent allomorphy: it is realized as kkeyse next to high-status NPs, and realized as i or ka otherwise (the choice between i and ka is phonological). This alternation can be straightforwardly captured by a contextual allomorphy rule such as (27) which maps the nominative feature [case: nom] to kkeyse only in the context of a [hon: +] in the node to its immediate left. (Following Bittner & Hale (1996a; b) among others, I represent case-marked nominals as KPs where the functional head K selects for an NP or DP complement.)

Under the Subset Principle, the VI rule with a more specific contextual condition applies when the contextual condition is met – this is (27a). Otherwise, the more general rule (27b) applies. (See Gribanova & Harizanov 2016 and Bonet & Harbour 2012 on the Subset Principle pertaining to contextual allomorphy.) Hence the feature [case: nom] is realized as kkeyse only when the NP bears [hon: +]; otherwise it is realized as i/ka. Contextual allomorphy occurs not only with the nominative case marker but also with the dative case marker as well as verbal morphology. In sections 4.2.1 and 4.2.2, I make use of similar VI rules in capturing the morphology of Korean honorification in a straightforward way.

I propose that all NPs of Korean enter the derivation with a valued honorific feature. I emphasize here again that whether NPs bear [hon: –] or [hon: +] is not tied to the lexical entry of the NP but determined by the speaker. Based on the value of [hon] on the NP, its case marking is realized according to the VI rules in (29). The case features [case: nom] and [case: dat] are subject to contextual allomorphy when the adjacent NP bears [hon: +].

In addition to allomorphy in case marking, a high-status subject also triggers the presence of (u)si. I propose that Agr_Subj bears [hon:___] which searches in its c-command domain for a valued hon feature. Since the external argument is at Spec, VoiceP, it is within the c-command domain of the probe. Hence the probe successfully finds the subject in its initial search and agrees with it. If the probe is valued as [hon: +], it is realized as (u)si.

The VI rule (31) states that [hon: +] is realized as (u)si in the context of a tense feature in the node to its immediate right. The context guarantees that (u)si is only realized for subject honorification; the same [hon: +] involved in addressee honorification would not trigger insertion of (u)si, since it is not in the context of [tense]. Notice that no VI rule targets [hon: –] in Korean since non-high-level honorification does not trigger overt verbal morphology.

The proposed analysis explains why the presence/absence of (u)si tracks the status of the structurally highest argument. The probe is simply searching downwards and agreeing with the first argument that it finds. Hence the high-status subject apeci-kkeyse triggers realization of (u)si on Agr_Subj in (32a). In contrast, the high-status indirect object apeci-kkey in (32b) does not affect the valuation of the probe, which instead agrees with the non-high-status subject Mina-ka.

The structures of (32a) and (32b) are visualized in (33) and (34) respectively. (I follow Pylkkänen (2008) in positing an Applicative Phrase (ApplP) in ditransitive constructions.)

Furthermore, high-status passive and unaccusative subjects are expected to trigger realization of (u)si despite being internal arguments since they are the only available goal for the probe on Agr_Subj. This prediction is borne out: passive/unaccusative subjects trigger regular subject honorification as shown in (35–36).

Lastly, when a ditransitive construction is passivized, the probe on Agr_Subj agrees with whichever internal argument is promoted to subject position. (37) shows two ways of passivizing ditransitive (32b). When the high-status indirect object apeci-kkeyse is promoted to subject as in (37a), (u)si is present on the verb. In contrast, when a non-high status direct object like senmwul-i is promoted as in (37b), (u)si cannot appear. Again, the presence or absence of (u)si is determined by the status of the structurally highest argument.

In this section I propose an analysis of Korean addressee honorification as allocutive agreement. In doing so I join Ceong & Saxon (2020), who also call for a similar approach. This paper provides a more fully worked out syntactic agreement mechanism in line with their earlier proposal. The analysis builds on that of McFadden & Sundaresan (2021), who made the connection between allocutive agreement, flexible direction of agreement, and proposal of the IntP phase by Carstens (2016) – modulo two points of divergence. The first is the position of the probe. McFadden and Sundaresan locate the allocutive probe in Force, while I locate it in Agr_Addr. The second is the model of agreement assumed. McFadden and Sundaresan cite Béjar & Rezac (2009) and Clem (2021) in adopting the Cyclic Agree model, but the agreement mechanism they describe is more flexible than what is possible under Cyclic Agree. When a probe fails to find a goal in its c-command domain in McFadden and Sundaresan’s system, the probe needs to search upwards across multiple projections. However, in the Cyclic Agree model, re-projection of a probe on a head X is limited to projections within XP. As a result, there is a limit to how “high up” a probe can search. The agreement phenomena described in this paper and in McFadden and Sundaresan’s work is better captured by a more flexible model of agreement such as bidirectional agreement (Arregi & Hanink 2021), introduced in Section 4.1.1. In this model, the probe actually searches upwards to agree with a goal c-commanding it from a separate projection. This eliminates the need for re-projection into multiple c-commanding projections, which would not be possible in a Cyclic Agree model without additional assumptions.

In Section 4.1.2 I introduced the SpkP and AddrP as representations of the speaker and addressee at the root of the clause. Putting together these projections and my proposal in Section 2.2 that (su)pni is located in the functional head Agr_Addr, I propose that Agr_Addr bears an [hon:___] that agrees with a valued honorific feature on Addr. This means that the probe on Agr_Addr agrees with a goal that c-commands it, as schematized in (38). (The part and auth features on the Addr and Spk heads are omitted in subsequent trees for ease of presentation, unless they become relevant to the discussion.)

There are two pieces that make this agreement pattern possible. The first is the bidirectional agreement model introduced in Section 4.1.1, under which upwards probing is possible. The second is the view of Carstens & Diercks (2013) and Carstens (2016) that there is a phase boundary between FinP and ForceP, rendering elements in FinP unavailable for agreement from a higher position. (See McFadden & Sundaresan (2021) for earlier discussion of this phase boundary in the context of allocutive agreement.) Their argument is motivated by the asymmetry between two kinds of embedded clauses in Bantu languages, where some show phasal properties while others don’t. Lubukusu is an especially instructive language since it has both types of embedded clauses: ones headed by the complementizer -li (shown in (39)), and ones headed by mbo (shown in (40)). The embedded subject can undergo hyperraising out of an mbo-clause as in (40b), but a li-clause prohibits such hyperraising as shown in (39b).18

Furthermore, -li and mbo display different properties with regards to adjunct clauses headed by sikila ‘because’, which are considered ForceP. Mbo can appear inside a sikila clause, which suggests that mbo appears in a position lower than the one occupied by sikila. On the other hand, -li cannot appear inside a sikila clause.

In explaining (41), Carstens (2016) locates -li and mbo in different positions within the articulated clause periphery system of Rizzi (1997; 1999): mbo at Fin, and -li at Force. The version of articulated periphery adopted by Carstens also includes the Int(errogative) head between Fin and Force, posited by Rizzi (1999). He motivates Int from the fact that the Italian embedded polar question complementizer se allows a high topic such as a Gianni in (42) to either follow itself (42a) or precede itself (42b). Crucially, the declarative complementizer che, which is located at Force, must precede the high topic, as the contrast of (42c-d) shows.

Int is a separate projection from Focus, too, since se can be followed by a separate focused phrase like questo in (43).

Carstens captures the asymmetry in (39–41) by claiming that Int is a phase head.

An embedded -li-clause projects up to ForceP, which includes IntP. Since Int is a phase head, its complement FinP is spelled out and the elements within FinP are unavailable for movement due to the Phase Impenetrability Condition (PIC; Chomsky 2000). Therefore, hyperraising out of a -li-clause is impossible as shown in (39b). On the other hand, an embedded mbo-clause projects only up to FinP. When a raising predicate directly selects an mbo-clause, it is in fact selecting a FinP which does not include a phase boundary in its periphery. Due to this truncated structure, hyperraising out of an mbo-clause is possible as shown in (40b). Schematized representations of the Lubukusu articulated clause periphery are provided in (45).

As McFadden & Sundaresan (2021) demonstrate, this phase boundary allows us to view allocutive agreement as upwards-probing agreement. The specific agreement mechanism is visualized with tree structures in (46–47). The tree in (46) shows a structure completed up to Agr_AddrP. Upon merge of Agr_Addr, which is right below Force and hence above IntP, the probe on Agr_Addr probes downwards but is unable to find a goal in its c-command domain due to the PIC. Therefore, it remains unvalued.

Having failed to find a goal from its downward probe, [hon:___] searches upwards until it finds either [hon: +] or [hon: –]. It finds its goal upon merge of Addr and copies the goal’s features, achieving valuation as shown in (47).19

In (31), I proposed a context-sensitive VI rule where [hon: +] is realized as (u)si in the context of [tense] in the adjacent node to its right. Here I present a similar account for the realization of (su)pni. Rule (48) states that (su)pni can be a realization of [hon: +] only in the context of a force feature in the adjacent node to its right. (49) is a schematic demonstration of (48) being applied in a structure uttered to a high-status addressee.

Japanese subject honorification involves a much wider range of constructions than Korean or Magahi (Harada 1976; Yamada 2019). Depending on the context and the specific utterance, it may involve verbal suppletion, the suffix rare, or the o-noun-ni nar- construction which consists of the prefix o/go-, the dative particle ni and the light verb nar ‘become’. I direct the reader to Harada (1976) and Yamada (2019) for a detailed introduction of the Japanese subject honorification system, and focus here on verbal suppletion. Yamada shows that verbal suppletion can co-occur with addressee honorification as in example (54), repeated below as (70). The verb root ossyar- is the suppletive form of the root iw- ‘say’.

(71) demonstrates another example of verb suppletion. The use of the suppletive form reflects the speaker’s respect towards the subject sensei.

Addressee honorification is expressed by the presence of the suffix mas attaching to verbs, or the copula des.

The “low” pronunciation site of mas has been a puzzle for Japanese linguists. Despite expressing information pertaining to the entire utterance, Japanese mas appears lower than Neg and T.

This leads Miyagawa (2012; 2017; 2022) to view it as originating from T but LF-raising to C. This movement is triggered by the need for mas to undergo agreement with the addressee at AddrP. On the other hand, Yamada (2019) argues for a node-sprouting analysis of mas at Neg: mas is inserted to the left of Neg when it is c-commanded by an honorific addressee represented at the periphery of the clause.

This syntactically low realization of mas may seem problematic for my analysis, given the co-occurrence of 2nd person subject and addressee honorification in (70). However, the issue can be resolved if we follow Chomsky (2000; 2001) and especially Takahashi (2010) and Sakamoto (2016) among others in assuming that vP/VoiceP is a phase in Japanese. Yamada (2019) classifies Japanese subject honorification into three tiers: the tier of argument structure, the tier of high-applicative expressions, and the tier of expressions of vP-periphery. Verb suppletion is classified as the tier of argument structure, the structurally lowest tier linked to the vP area. If the suppletive subject honorification strategy shown in (70) is low enough to be below the vP/VoiceP phase boundary, while mas is realized at Neg or T, we expect both to be able to appear in a sentence since they belong to different spell-out domains.

Tamil is another language with an honorification system. First, it shows distinct verb agreement morphology for singular 3rd person towards whom the speaker expresses a polite relationship (the “3rd person polite” form in Table 3). Second, the morpheme ŋgæ, which marks plurality in the 2nd and 3rd persons (feminine/masculine/polite), is also used to reflect properties of the addressee. Specifically, it appears when the addressee is multiple persons or when an addressee is high-status. Table 3 presents the Tamil subject agreement paradigm, expressed on the verb ooɖŭ ‘run’.26

McFadden (2020) views the addressee-related ŋgæ as the morphological realization of allocutive agreement. (74) shows examples of the allocutive use of ŋgæ. We know that it is not a plural subject agreement marker, since the subjects are singular in these sentences.

Tamil is an interesting language in the context of the 2nd person honorification interaction. As shown in (56) repeated below as (75), 2nd person subject agreement cannot co-occur with allocutive agreement.

In this respect Tamil behaves similarly to Magahi. However, Tamil allocutive agreement is not as readily embeddable as in Magahi. It can appear in certain embedded contexts like (76), but McFadden & Sundaresan (2021) point out that embedded allocutivity is available only in a limited set of environments, similar to what Miyagawa (2012; 2017; 2022) describes for Japanese.27

This nuanced nature of Tamil allocutive agreement leads to different views on the position of the allocutive probe in Tamil. Depending on where one locates the probe, a different explanation is warranted for the 2nd person co-occurrence restriction. I merely sketch the two possible explanations here and leave a full-fledged analysis of Tamil to future research. Alok (2020; 2021) and Alok & Haddican (2022) locate the allocutive probe at Fin, in a parallel to Magahi. This approach allows for extending Alok’s account of the Magahi 2nd person co-occurrence restriction to Tamil. However, since Tamil ŋgæ is not as freely embeddable as Magahi allocutive agreement, more needs to be said about its licensing condition in embedded clauses. One possibility is to have Tamil AddrP be licensed only under complementizers of the appropriate featural specifications (e.g., an [+Addr] feature), as Alok & Haddican (2022) suggest for Japanese.28

For McFadden & Sundaresan (2021), it is Force that bears the probe in Tamil. The probe fails to find a goal beyond the IntP phase boundary and instead agrees with AddrP above it, which is only present in matrix and root-like embedded environments. The limited embeddability is explained more easily, if we follow them in viewing Tamil allocutive agreement as a main clause phenomenon that involves the AddrP instead of FinP. But adopting their analysis warrants some additional assumptions or amendments in explaining the ungrammaticality of (75). A potential explanation is to follow McFadden (2020) and view the 2nd person co-occurrence restriction as an instance of a generalized Obligatory Contour Principle (OCP) constraint (Mohanan 1994). McFadden suggests that the restriction may not be a deep syntactic constraint against two structurally adjacent heads agreeing with the same features but a surface-level constraint against multiple string-adjacent realizations of a linguistic expression. This is an attractive alternative explanation for Tamil, since the 2nd person plural subject agreement already includes the string ŋgæ as shown in Table 3.