The grammar of self-talk. What different modes of talking reveal about language

2.1 The phenomenon and its significance

The significance of self-talk is perhaps best appreciated by considering the classic question as to whether human language is (primarily) a tool for thought or for communication (see Chomsky 2017 for the former view and Carruthers 2002; Jackendoff 2002; Levinson 2019 for the latter). Self-talk has the potential to be revealing in the context of this debate as – at least superficially – it appears to lie squarely between these two functions. On the one hand, self-talk has in common with thought that it is private (as opposed to the social dimension intrinsic in communication). On the other hand, self-talk has in common with communication that it is overt, or at least it can be (see below), whereas thought cannot (without turning into self-talk).^[4] This is summarized in Table 1.

Note for completeness that self-talk may remain covert. For example, Latinjak et al. (2023]: 356) conceptualize self-talk as “verbalizations addressed to the self, overtly or covertly” (see also Hardy 2006; Theodorakis et al. 2000). The terminology surrounding the phenomenon of self-talk is messy. The term self-talk is sometimes reserved for overtly talking to oneself (alongside the terms egocentric speech and private speech) whereas other terms are used for covert self-talk, such as inner voice, inner speech, verbal thoughts, covert speech, silent speech, verbal thinking, verbal mediation, inner monologue, inner dialogue, inner voice, articulatory imagery, voice imagery, speech imagery, and auditory verbal imagery (see Alderson-Day and Fernyhough 2015; Nedergaard and Lupyan 2024 for an overview). When self-talk remains covert, it is impossible to distinguish from thought by the criteria listed in Table 1. According to Alderson-Day and Fernyhough (2015]: 931), inner speech (a covert form of self-talk) “can be defined as the subjective experience of language in the absence of overt and audible articulation.”^[5]

The terminological conventions, as well as the definitions of self-talk just reviewed, already point to different ways of viewing the nature of self-talk. Strikingly, this debate is not unlike the one regarding the function of language itself. Specifically, much of the discussion evolves around the question of whether self-talk should be viewed as a vehicle of thought or whether it is better analyzed as a way of communicating with oneself.

The first explicit mention of self-talk in Plato’s Theaetetus unites these views, suggesting that “the soul when thinking appears to me to be just talking – asking questions of herself and answering them, affirming and denying” (Plato 1970 (1892): 252). Accordingly, thinking is equated with self-talk, which, in turn, is a form of communication within oneself.

In contrast, the two pioneers of contemporary explorations of self-talk, Jean Piaget and Lev Vygotsky, held two opposing views regarding the relation between self-talk, thought, and communication. What they have in common is that both view self-talk as an important milestone in children’s cognitive development and, thus, as a window into the nature of cognition more generally.

Jean Piaget is generally considered the first psychologist who took an interest in self-talk (which he termed egocentric speech) and its role in cognitive development. He hypothesized that egocentric speech is a developmental step towards social speech (used in interaction with others). Accordingly, children start with their own thoughts, externalized in egocentric speech (labelled self-talk in Figure 1) and are only able to communicate them to others when they can take the perspective of others (Piaget 1923/1962).

Figure 1:

The role of self-talk in cognitive development according to Piaget.

A decade later, another psychologist, Lev Vygotsky – the first to conduct systematic experiments on children’s self-talk – develops a view diametrically opposed to that of Piaget. Specifically, according to Vygotsky (1934/1986), self-talk is rooted in communication with others (i.e., the social world of the child), which often serves the purpose of regulating the child’s behaviour. Over time, this is internalized, and the child uses (overt) self-talk for self-regulation. Eventually, this leads to complete internalization, i.e., covert self-talk and thought. Thus, for Vygotsky, overt self-talk bridges social speech and covert self-talk, which he viewed as cognitively more sophisticated than overt self-talk or communication with others. This is illustrated in Figure 2.

Figure 2:

The role of self-talk in cognitive development according to Vygotsky.

While in both views, self-talk constitutes an important developmental milestone, they differ as to whether it is viewed as an aid to externalizing thoughts to be communicated (Piaget) or as an aid to internalizing communicated content to arrive at the capacity of thought (Vygotsky). Even though both accounts focus on the developmental role of self-talk in children, these opposing views are also found in explorations of self-talk (overt and covert) in adults, who are known to use overt self-talk while engaging in problem-solving and other activities (e.g., Duncan and Cheyne 2001; Duncan and Tarulli 2009; Winsler 2009). Some view self-talk (covert and overt) as a way to support the thinking process (Sokolov 1975), especially in demanding situations (Hardy 2006: 93), or even as a form of thinking itself. For example, according to Bunker et al. (1993]: 226), “Anytime you think about something, you are in a sense talking to yourself.” This emphasizes the role of self-talk as a thinking tool (Geurts 2018; Lupyan and Swingley 2012). In contrast, others view self-talk as an intrinsically dialogical phenomenon through which “the individual interprets feelings and perceptions, regulates and changes evaluations and convictions, and gives him/herself instructions and reinforcement” (Diaz 1992; Hackfort and Schwenkmezger 1993: 355). This emphasizes the communicative character of self-talk (see Deamer 2021).

No matter whether self-talk is primarily viewed as a way of thinking or as a way of communicating with oneself, the fact remains that it seems to serve several functions in adults, including self-control, self-attention, self-regulation, self-motivation, self-critique (even self-denigration), developing metacognition and self-awareness, processing social situations, problem solving, among others (Alderson-Day and Fernyhough 2015; Brinthaupt et al. 2015; Chella and Pipitone 2020; Kompa 2024). In the context of the present paper, the functions of self-talk are significant since there are studies which seek to correlate specific functions of self-talk with specific linguistic properties. For example, spontaneous and motivational self-talk have been argued to be characterized by the use of 1st person (e.g. ‘I’m actually good at this’, ‘I got this.’) whereas goal-directed and instructional self-talk by the use of 2nd person (e.g. ‘You are actually good at this’, ‘You need to mow the lawn this weekend!) (Latinjak et al. 2023). Moreover, it has been shown that goal-directed self-talk is more efficient when formulated as a question (e.g., ‘Will I make it?’) or in the 1st person plural (e.g. ‘We will make it.’) (Senay et al. 2010; Son et al. 2011; Van Raalte et al. 2018).

There are three main messages to take away from this brief review:

The contribution of this paper is two-fold. First, by adding a linguistic perspective, we can shed light on some of the questions regarding the nature and function of self-talk. Second, by adding self-talk as a linguistic object of study, we can shed light on questions regarding the nature and function of language itself.

2.2 Methodologies

Having introduced the phenomenon of self-talk (and its kin), a few words are in order about methodological issues that arise when exploring (the linguistics of) self-talk.

As mentioned above, the first studies of self-talk involved children, who frequently talk to themselves and hence can be observed doing so. This is what both Piaget and Vygotsky did. However, when it comes to self-talk in adults, observation of self-talk in natural settings (and hence corpus data) is unavailable. This is because talking to oneself is associated with a social stigma – so much so that according to Goffman (1981]: 81), “There are no circumstances in which we can say, ‘I’m sorry, I can’t come right now, I’m busy talking to myself.’” The reason for this stigma, according to Goffman (1981]: 78), is that it violates the social agreement about the communicative function of speech. Accordingly, the apparent decrease of self-talk with age may be related to this social norm (Duncan and Tarulli 2009: 177). For all these reasons, the exploration of self-talk comes with methodological problems (Ariel 2022).

To overcome these difficulties, several methodologies have been employed. In psychological studies, self-talk is typically explored through self-reports. For example, Brinthaupt et al. (2009) develop the Self-Talk Scale, which measures the frequency of self-talk and its correlation with particular functions of self-talk. However, this does not allow us to test for specific linguistic properties of the kind I wish to explore here, as it does not target specific linguistic forms.

One source for self-talk data comes from fictional contexts, literary and film alike (see Banfield 1982 for the significance of fiction for linguistic analysis). That is, in the context of movies or plays, characters are sometimes talking to themselves. In fact, there is a whole scholarship surrounding soliloquy in Shakespeare, for example (Murphy 2015). Some movies revolve around the premise that one of the characters can “hear” the thoughts (thus inner speech) of others, such as Mel Gibson’s character in What Women Want and Sookie Stackhouse in True Blood.

In the context of novels, too, self-talk in the form of an inner monologue is a common literary strategy through which a story unfolds. An early example of this is found in Arthur Schnitzler’s Leutnant Gustl, whose inner speech has recently been analyzed in Krifka (2023b). While neither film, theatre, nor novels provide us with naturally occurring data, they at least give us some insights into the intuitions of their authors as to what constitutes well-formed self-talk.

This brings us to the main methodology used to collect data for the present paper: the use of native-speaker intuitions. Specifically, while data from fictional work gives us insight into what speakers might say in self-talk, they do not provide us with negative data or minimal pairs, as is the case for corpus data in general. Hence, linguists have long relied on native speaker judgements of constructed examples in the form of elicitation tasks, which have stood the test of time as a valid methodology (Arppe and Järvikivi 2007; Langsford et al. 2018; Schütze 2016, Sprouse and Almeida 2017). While classically, such elicitation tasks have been used to target language outside of interaction and, hence, without the specification of who is talking to who, it is safe to say that these tasks are typically not assumed to target language in the context of self-talk.

Using native speaker judgments to elicit the language used in self-talk thus relies on the premise that native speakers do indeed have judgements about what constitutes well-formed utterances in the context of self-talk. This assumption receives indirect support from the fact that speakers do have clear judgements about language in interaction, including differences pertaining to the identity of the interlocutor (Wiltschko 2021). For example, speakers have judgements about the use of formal and informal pronouns even in the absence of a relevant interlocutor. What makes the elicitation of language in interaction somewhat more involved than eliciting sentences in isolation is the fact that contexts must be explained, and consultants must pair a particular context with a particular utterance. For this reason, Wiltschko (2021) develops the conversation board methodology, which is based on Burton and Matthewson’s (2015) storyboard elicitation task. This consists of presenting the consultant with cartoon-like pictures that minimally include a panel depicting the context of the utterance and a speech bubble for the interlocutor whose utterance is under investigation. The utterance may be presented within this speech bubble for a well-formedness judgement, or else the consultant may be asked to provide an appropriate utterance given the relevant context. In Goddard et al. (2022), this methodology was used for eliciting self-talk data where the context of self-talk was varied between thinking to oneself (depicted via a thought bubble), talking aloud to oneself (indicated by a speech bubble), talking to oneself in the mirror, and typical conversations with another interlocutor.

If not otherwise indicated, the data reported in this paper comes from simple, well-formedness judgements with native speakers. Judgements have been very clear and consistent across consultants. I also report on some of the data collected by Goddard et al. (2022), which were consistent with the findings obtained from simple, well-formedness judgements.

For completeness, note that by necessity, linguistic judgements can only be obtained for overt self-talk. This is because to judge the language of self-talk, it must be externalized (i.e., overt). These judgements might extend to silent self-talk, which has been reported to sometimes mirror overt self-talk and which has been shown to even activate brain areas that involve motor planning (e.g., Barch et al. 1999; Pratts et al. 2023; Yetkin et al. 1995). However, there are some versions of silent self-talk (i.e., inner speech) which cannot be observable, and which are known to differ linguistically in that they appear to be less articulated, condensed, and abstract (Alderson-Day and Fernyhough 2015: 942–943). Though this form of inner speech might have the potential to shed light on the nature of language and its relation to thought, I have nothing to say about this phenomenon. What follows is a discussion of overt self-talk only.

2.3 The linguistics of two modes of self-talk: an analytical challenge

As we have seen, the psychological literature recognizes differences in form, function, and linguistic expression, i.e., whether the person engaged in self-talk is using a 1st or 2nd person pronoun to refer to themselves. This linguistic difference will be at the core of the current exploration of self-talk. As a terminological convention, I use the terms you-centered self-talk and I-centered self-talk to distinguish between these two modes.^[6]

A first question we may ask is if the difference between I-centered and you-centered self-talk correlates with any of the other dimensions of variation (i.e., form or function). The answer seems to be negative, as I now show. In terms of its form, both I-centered and you-centered self-talk can be used both overtly and covertly (Holmberg 2010). Next, consider the function of self-talk. Recall that it has been claimed that spontaneous and motivational self-talk are characterized by I-centered self-talk, whereas goal-directed and instructional self-talk are by you-centered self-talk (Latinjak et al. 2023). If this is indeed the case, this suggests a correlation between the function of self-talk and its linguistic expression. However, this correlation does not seem to be categorical. While I-centered self-talk is certainly possible in spontaneous and motivational self-talk, it is not restricted to these functions. For example, the statement in (9), based on Holmberg (2010]: 57) can hardly be classified as motivational, yet I-centered self-talk is possible, as in (9a). Moreover, while it might be classified as spontaneous (though it is unclear exactly what criteria one might apply), it still allows for you-centered self-talk, as in (9b).

Similarly, goal-oriented self-talk, too, allows for both I-centered and you-centered self-talk, as shown in (10).

Thus, the use of I-centered and you-centered self-talk does not categorically correlate with either its form or its functions.^[7] Nevertheless, there are crucial differences between I-centered and you-centered self-talk, as first observed in Holmberg (2010). These differences manifest themselves in different linguistic profiles, which in turn suggest that we are dealing with two different modes of self-talk (Ritter and Wiltschko 2021). In the remainder of this section, I review some of these distinctions. For each of the linguistic properties, I start by illustrating it based on a typical conversation, and then I show that it distinguishes between you-centered and I-centered self-talk.

The first phenomenon I consider is the use of a vocative nominal. Consider first the examples in (11) where Alaka is talking to Thea. Here, the use of the vocative is well-formed, regardless of whether Alaka is talking about Thea (using the 2nd person pronoun) or about himself (using the 1st person pronoun).

Now, consider what happens when Alaka is engaged in self-talk. In this case, referring to himself with a 2nd person pronoun is possible in the presence of a vocative, as in (12a). In contrast, referring to himself with a 1st person pronoun is not possible, as in (12b).

Krifka (2023b) finds support for this restriction on vocatives in the literary work he explores (Schnitzler´s Leutnant Gustl). Specifically, he observes that passages that are characterized by I-centered self-talk switch to you-centered self-talk when a vocative is used. Significantly, there appears to be no instance of a vocative in I-centered self-talk.

The second phenomenon concerns imperatives. In a typical conversation, imperatives are characterized by a requirement for the subject to be equated with the addressee.^[8] This is illustrated in (13). While the subject of an imperative is often silent (indicated as pro _Adr in (13a)), it may remain overt. In this case, it may be realized as a 2nd person pronoun, as in (13b), or as a form of address (i.e., the addressee’s name, as in (13c)). Crucially, 1st person pronouns are ungrammatical in this context, as shown in (13d).^[9]

Note further that even when the subject of an imperative remains silent, there is still evidence that it must be syntactically represented in the form of a silent pronoun, which refers to the addressee (pro _Adr ). Specifically, when the direct object refers to the addressee, it has to be realized as a reflexive pronoun, as in (14a), whereas the use of a personal pronoun is ill-formed, as in (14b). This is because the silent pro _Adr binds the direct object, and hence, a reflexive pronoun is required as per binding theory. In contrast, when the direct object refers to the speaker, it is not bound by the silent pro _Adr subject. Hence, the reflexive 1st person pronoun is ungrammatical, as in (14c), whereas the personal pronoun is well-formed, as in (14d).^[10]

In a typical conversation, the prohibition of the use of 1st person pronouns follows, of course, from the requirement that an imperative is necessarily addressed to someone else. But what happens in self-talk where the speaker is addressing themselves? Crucially, imperatives must still be realized with addressee denoting (2nd person) subjects. The data in (15) illustrate that in the context of self-talk, the same pattern holds as in a typical conversation: the subject may remain silent, as in (15a), and when it is overt, it must be realized as a 2nd person pronoun, as in (15b), or as a form of address (in this case the speaker’s name), as in (15c). Crucially, the 1st person pronoun is ungrammatical, as shown in (15d).

Thus, even though in self-talk the speaker and the addressee are the same person, and hence the speaker appears to address themselves, the 1st person pronoun is ruled out. In other words, imperatives are restricted to you-centered self-talk (Ritter and Wiltschko 2021).

This is further confirmed by imperatives with direct objects, where we witness the same generalization as with imperatives in typical conversations. Only the 2nd person reflexive is licit, as in (16a). As in typical conversations, the non-reflexive 2nd person pronoun is ungrammatical, as in (16b), suggesting that there is an addressee denoting pro _Adr in subject position. In contrast to imperatives in typical conversations, however, in self-talk, a 1st person pronoun is ungrammatical, no matter whether it is realized as a reflexive pronoun, as in (16c), or as a personal pronoun, as in (16d).

Taken together, these facts demonstrate that imperatives are impossible in I-centered self-talk: the recipient of a command cannot be realized as a 1st person reflexive or personal pronoun.

The third property of self-talk I here discuss concerns verbs of cognition. Consider first the examples in (17), which involve a typical conversation. We can tell others something about our own mental state, as Alaka does in (17a). In contrast, we do not typically tell others about their current mental states. While there might be contexts in which this is possible, under normal circumstances, such utterances are deviant (hence, the example in (17b) is marked by a hashmark #).

The deviance of (17b) follows from the conditions of use for assertions. Asserting a proposition p is felicitous when two conditions hold: (i) the speaker is certain about the truth of p and (ii) the speaker assumes that the addressee does not know p. Since someone’s mental state is accessible to them and only to them, one cannot felicitously tell someone else about their mental state. Thus, even if Alaka is certain about Thea’s mental state (because Thea has a certain expression on her face), he could still not tell her about it because Thea would already know. Hence, (17b) is ill-formed.

Instead, what Alaka might felicitously say in contexts where he observes an expression of disbelief on Thea’s face is an assertion of having evidence for her mental state, as in (18a). This is well-formed because Alaka has direct access to how Thea looks, and Thea may not know that she does look like she can’t believe her luck. Thus, in this context, the utterance in (18a) follows the felicity conditions of assertion. And now the converse holds for (18b). Under normal circumstances, a speaker has access to their own mental state, and hence, they would not resort to talking about overt evidence for their mental state. Additionally, such evidence would be accessible to an addressee, and hence, telling them what the evidence suggests again violates the conditions for assertion. Hence, (18b) is ill-formed.

The same holds for other ways of stating one’s evidence for another person’s mental state, as shown in (19).

Now, consider what happens in self-talk. As Holmberg (2010) observes, verbs of cognition and certain experiencer verbs are not allowed in you-centered self-talk, whereas they are in I-centered self-talk (adapted from Holmberg 2010: 59f.).^[11]

About this contrast, Holmberg (2010]: 60) states that “you can’t refer to the self as an experiencer of feelings or holder of intentions or plans […].” He proposes that the referent of you in self-talk is a “mindless self” and hence cannot serve as the subject of a verb of cognition. While Holmberg recognizes that the same pattern holds in typical conversations, he suggests that the restrictions, though seemingly identical, have a different explanation. In a typical conversation, you refers to the addressee whose mind is inaccessible to the speaker. Clearly, in self-talk, the addressee’s mind is accessible, and hence Holmberg (2010]: 60f.) suggests that you in self-talk refers to a mindless self.

This explanation runs into problems, however, when we consider the examples in (22). As in a typical conversation, the restriction against you being used as the subject of believe disappears when embedded under “It looks like…”, as shown in (22a). In contrast, in this context, I cannot be used as the subject of believe, as shown in (22b).

The conclusion we can draw from these data is that you in self-talk cannot be analyzed as referring to a mindless self, contrary to Holmberg (2010). From a theoretical perspective, this does not seem to be surprising. It is not clear why the addressee-denoting second person pronoun should have different referential properties in self-talk. Rather, you in self-talk behaves just like you in a typical conversation: it appears to refer to an inaccessible mind. But this is, of course, counterintuitive. A person engaged in self-talk has access to their own mind regardless of whether they use you-centered or I-centered self-talk.

Thus, the properties of utterances involving verbs of cognition highlight the analytical challenge the two modes of self-talk present us with. That is, we have now seen three linguistic properties that distinguish I-centered from you-centered self-talk. While I-centered self-talk is incompatible with vocatives and imperatives, you-centered self-talk does not support talking about one’s mental state. Thus, across all three properties, you-centered self-talk behaves no differently from typical conversations. This is summarized in Table 2.

Now, consider why the properties of self-talk constitute a conundrum that needs to be addressed. It may be tempting to assume that the difference between self-talk and typical conversations is pragmatic in nature. After all, these two modes of talking are defined by contextual variables, namely, who is talking to whom. At first glance, this invites a pragmatic solution. However, such a solution runs into non-trivial problems in the face of the two modes of self-talk.

Consider first what a pragmatic account for the ban on imperatives in I-centered self-talk might look like. I assume that speech acts realized as imperatives are associated with two conditions of use when used canonically: (i) the speaker wants the addressee to act in a certain way and (ii) the addressee would not act in this way if the speaker were not to utter the imperative. Arguably, imperatives are ill-formed in I-centered self-talk for the same reason as imperatives cannot have a 1st person singular subject. A speaker cannot command themselves to initiate an action, and this is true no matter if the speaker talks to themselves or to someone else (in which case they would indicate to their addressee that they are commanding themselves). While this might be a plausible account for the ill-formedness of imperatives in I-centered self-talk, it cannot account for their well-formedness in you-centered self-talk. This is because the context that characterizes self-talk is identical, no matter whether it is I-centered or you-centered. In both types of self-talk, one is giving a command to oneself, and hence, they should equally be ruled out under the pragmatic account sketched above.

Similar considerations hold for the use of vocatives in self-talk. That is, the function of a vocative is to either get the addressee’s attention or to maintain it (Zwicky 1974). On a pragmatic account, one might reasonably propose that one need not and, therefore, cannot get one’s own attention. This would explain the ill-formedness of vocatives in I-centered self-talk. However, this explanation cannot straightforwardly account for the well-formedness of vocatives in you-centered self-talk. Again, from a pragmatic point of view, both types of self-talk are characterized by the same context (a person talking to themselves) no matter which pronoun they use to refer to themselves. Thus, a pragmatic account might succeed in accounting for the ill-formedness of imperatives and vocatives in I-centered self-talk, but it cannot account for their well-formedness in you-centered self-talk.

As for the properties of cognitive predicates in the context of self-talk, here, a pragmatic account cannot straightforwardly explain the ill-formedness of you-centered self-talk. That is, if the reason one cannot use you as a subject of a cognitive predicate is that one does not have access to somebody else’s mind, then why does the same hold in self-talk? So why is you-centered self-talk ruled out when talking about one’s mental state? And why does it matter whether the person engaging in self-talk refers to themselves with a 1st or 2nd person pronoun?

In a nutshell, the analytical challenge surrounding self-talk presents itself differently for I-centered and you-centered self-talk. In both cases, the context is such that a speaker is talking to themselves, and thus, the speaker simultaneously serves as their addressee. Yet, the linguistic properties of I-centered self-talk suggest that there is no addressee, ignoring the real-world fact that there is, even if it is identical to the speaker. In contrast, the linguistic properties of you-centered self-talk suggest that there is an addressee, but what is ignored in this case is the real-world fact that the addressee is identical to the speaker; it has the same properties as in a typical conversation in which the addressee is different from the speaker. This is something a pragmatic account cannot explain because we would not expect that real-world knowledge be ignored. Rather, what the linguistic profile of self-talk suggests is that the properties are grammatical in nature. That is, it is the hallmark of grammar that it may remain insensitive to real-world knowledge. For example, while grammatical gender may be rooted in real-world properties, categorizing a noun as masculine or feminine may remain insensitive to real-world knowledge. What the linguistic properties of self-talk thus reveal is that the addressee role must be part of the grammatical representation of an utterance. As mentioned above, there is growing consensus that this is indeed the case, based on several linguistic phenomena that demonstrate a sensitivity to the presence of an addressee. In the next section, I shall briefly review some of these proposals, which will then allow us to evaluate them in a novel way based on how they fare in deriving the linguistic properties of self-talk.

4.1 The absence of an addressee-oriented layer in I-centered self-talk

Recall from Section 2.3 that I-centered self-talk is characterized by two restrictions that differentiate it from you-centered self-talk: it does not license the use of vocative nominals, nor does it allow imperatives. The relevant examples are repeated below for convenience.

According to the proposal in (31), I-centered self-talk is structurally deficient as compared to you-centered self-talk in that it lacks the addressee-oriented grounding layer. This derives the ban on vocatives and imperatives in I-centered self-talk as follows.

On independent grounds, Ritter and Wiltschko (2020) have argued that vocatives occupy the addressee-oriented grounding layer (i.e., SpecGround_AdrP). This analysis captures the fact that vocative nominals name the addressee and serve various functions, such as getting the addressee’s attention or alerting the addressee that the content of the utterance is particularly relevant to them (Zwicky 1974). It thus follows that vocatives are not licensed in I-centered self-talk as it lacks the relevant interactional position. Conversely, in you-centered self-talk, Ground_Adr is available, and hence vocatives are well-formed. This is schematized in (32), where the asterisk indicates that the vocative has no available position to be licensed.

As for the impossibility of imperatives in I-centered self-talk, again, it follows from the absence of Ground_AdrP. Suppose that the null subjects (pro_Adr) of imperatives must be coindexed with the addressee in SpecGround_AdrP (see Ritter 2024 for a suggestion along these lines). In the absence of SpecGround_AdrP, there is no antecedent for the subject of the imperative. It thus follows that imperatives are not licensed in I-centered self-talk as it lacks the relevant interactional position that would provide the antecedent for pro. Conversely, in you-centered self-talk, Ground_Adr is available, and hence imperatives are well-formed. This is schematized in (33), where the asterisk indicates that the addressee has no available position to be licensed.

Another difference that sets apart I-centered from you-centered self-talk concerns the use of discourse markers, i.e., units of language which serve to regulate the interaction and thus many of them are addressee-oriented. They are precisely the empirical domain that motivated the interactional spine hypothesis. That is, according to Wiltschko (2021), sentence-peripheral discourse markers are directly associated with the interactional spine. Hence, it is predicted that discourse markers which are licensed in Ground_Adr will not be available in I-centered self-talk. This prediction is borne out.

First, consider huh, which can be used as a confirmational. As shown in (34), in a typical conversation, huh can be used by the speaker to request confirmation for their assumption that the addressee holds the belief expressed in the proposition. In (34a), the belief to be confirmed is that the addressee should read Moby Dick, and in (34b), the belief to be confirmed is that the speaker should read it. In both cases, the speaker wants confirmation that the addressee believes the proposition to be true.

That huh is indeed addressee-oriented is supported by the fact that it cannot be used to confirm a belief the speaker firmly holds. This is shown by the contrast in (35). One does not have direct access to another person’s mental state, and hence, a speaker can use huh only to confirm their assumptions about someone else’s mental state, as in (35a). In contrast, if the propositional content is such that the speaker has direct epistemic access, as when it concerns their own mental state, then the use of huh is ruled out, as in (35b).

Contrasts of this kind led Wiltschko (2021) to assume that huh must be interpreted in Ground_AdrP – it is always about the addressee’s epistemic state. This is schematized in (36).

Now, consider what happens in self-talk. As shown in (37), huh is possible in you-centered self-talk but not in I-centered self-talk (cf. Ritter and Wiltschko 2021).

This follows from the analysis of self-talk. Since I-centered self-talk lacks Ground_AdrP, huh is not licensed (as indicated by the asterisk in (38a)). This contrasts with you-centered self-talk, which contains Ground_AdrP and hence, huh is well-formed, as in (38b).

Note that a purely pragmatic analysis would not be able to account for the contrast in (37). That is, one might argue that (37b) is ruled out because one cannot request confirmation from oneself. However, the same logic would equally apply to you-centered self-talk, as here, too, one is requesting confirmation from oneself. In the analysis in (38), however, the contrast in (37) is grammatically conditioned. Hence, real-world knowledge cannot interfere with the grammatical restriction on I-centered self-talk. In the absence of Ground_AdrP, there is no grammatically licensed position for an addressee, even if the speaker knows that they are addressing themself.

Next, consider the use of the discourse particles doch and nämlich, which are sensitive to the addressee’s epistemic state. As we have seen in Section 1, doch is used when the speaker assumes that the addressee knows (or should know) the propositional content of their utterance. The sensitivity of doch to the epistemic state of the addressee can be analyzed by assuming that it associates with Ground_Adr (either via agree or covert movement; see Thoma 2016; Wiltschko 2024b). Hence, we expect that doch is sensitive to the mode of self-talk. It should be possible to use doch in you-centered self-talk but not in I-centered self-talk.^[14] This prediction is borne out, as shown in (39).

Now consider nämlich, which is used when the speaker assumes that the addressee does not (or cannot) know the propositional content. Thus, nämlich is ruled out in self-talk, no matter whether it is you-centered, as in (40a) or I-centered, as in (40b).

The ungrammaticality of nämlich in self-talk is due to the simple fact that it cannot be the case that a person engaged in self-talk will assume that they do not (or cannot) know the propositional content that they are uttering. In this way, the relation of the person engaged in self-talk to themself is akin to the relation between two close acquaintances (see example (5), Section 1).

4.2 The addressee in you-centered self-talk

In this section, we turn to the restriction on you-centered self-talk discussed in Section 2.3, namely the ban on verbs of cognition. I show how it can be analyzed and argue that it provides additional evidence for a grammatical approach (Section 4.2.1). Moreover, I discuss a prediction of the present analysis, namely that we expect two different types of you-centered self-talk. That is, given that you-centered self-talk is analyzed as a conversation one has with oneself, we expect that the actual self may hold either of the two available interactional roles. That is, a person engaged in self-talk may take on the role of the speaker addressing a mirror image of themselves (for example), or else, they may take on the role of the addressee being addressed by an imaginary speaker. I present evidence that this prediction is indeed borne out (4.2.2).

4.3 A grammatical difference between self-talk and typical conversations

At the core of the present analysis lies the claim that self-talk comes in two guises: I-centered self-talk is essentially a form of thinking out loud, while you-centered self-talk corresponds to having a conversation with oneself. This difference, I argued, is structurally conditioned in that only you-centered self-talk contains a syntactic position for the addressee (SpecGround_AdrP). If you-centered self-talk is indeed a conversation with oneself, the question arises as to whether it differs from a typical conversation with another individual. This is the question I address in this subsection. I show that there is at least one difference involving the use of intonational tunes, and I argue that this difference is structurally conditioned (see Ritter and Wiltschko 2021).

First, consider the intonational tunes associated with wh-questions in a typical conversation. As shown in (51), in a typical conversation, wh-questions can be realized with either falling (↘) or rising (↗) intonation (Bartels 1999; Bolinger 1989).

Crucially, you-centered self-talk differs in this respect, as shown in (52). Only falling intonation is possible when talking to oneself (Ritter and Wiltschko 2021; see also Krifka 2023b).

Following Wiltschko and Heim (2016), I assume that a rising intonational tune signals a call on the addressee to respond. As such, it occupies the head of Resp(onse)P, the highest category in the interactional structure.^[16]

Falling intonation, in contrast, is the default and, hence, is not interpreted as a meaningful intonational tune. This is because it arises naturally due to the fact that during an utterance, pitch declines automatically with the decrease in subglottal air pressure (Cohen et al. 1982). As such, unlike rising intonation, falling intonation does not have a syntactic representation (Wiltschko 2024a).

Given this analysis, the impossibility of rising intonation in you-centered self-talk indicates that RespP is available in a typical conversation, as in (53a), whereas it is not in you-centered self-talk, as in (53b) (Ritter and Wiltschko 2021). Hence, there is no position available to host rising intonation, as illustrated in (53b) by means of the asterisk.

This analysis implies that in self-talk, the grammatical regulation of turn-taking is not available. While in typical conversations, the addressee is an active participant from whom the speaker can request a response, in you-centered self-talk, they are not. A person engaged in self-talk does not call on themself to respond. This analysis is consistent with Holmberg’s (2010]: 57) observation that self-talk is always a “one-way communication.”

The logic behind Ritter and Wiltschko’s analysis regarding the lack of RespP in self-talk holds for initiating moves only, where the speaker may signal a request for a response by their interlocutor. This is implemented by assuming that the Response-set is other-oriented, as in (30b), repeated below for convenience. The situation is, however, different in reaction moves, where the response set is self-oriented (30a).

Crucially, reaction moves can but need not be reactions to previous conversational moves. Rather, speakers can react to non-linguistic events as well, and if they do, RespP is still available (Wiltschko 2021, 2024b). In the present context, this raises the question as to whether RespP is available in self-talk when it serves to mark a reaction move. In other words, is the structural deficiency of you-centered self-talk purely grammatically conditioned, or does it reflect the logic of initiation versus reaction moves? If it is purely grammatically conditioned, we would expect a categorical absence of RespP, no matter whether it marks initiation or reaction. If, instead, it is pragmatically conditioned, we would not expect a categorical ban on RespP, but instead, RespP should be available when it serves to mark the move as a reaction move. In what follows, I present evidence suggesting that RespP is available when a person engaged in self-talk reacts to a non-linguistic event.

Upper Austrian German has a discourse marker (ma), which serves to mark a reaction move (Wiltschko 2024b). It can be used to mark a reaction to a non-linguistic event, as in (54) (Wiltschko 2024b: 183).

Wiltschko (2024b) analyses ma as a pro-form for a speaker-oriented GroundP, which occupies the specifier position of a self-oriented (reacting) RespP, as schematized in (55).

According to this analysis, with the use of ma, a speaker marks a reaction to their own epistemic state. This, in turn, captures the fact that ma may indicate surprise. However, ma is not a discourse marker dedicated to surprise. Rather, ma can also be used in a context where the speaker suddenly remembers something, as in (56) (Wiltschko 2024b: 193).

In this context, no surprise is involved as the propositional content is something that Reingard previously knew but has temporarily forgotten. What the two contexts in (54) and (56) have in common is that the speaker marks a reaction to their own epistemic state. Thus, ma constitutes an ideal test case for whether self-talk allows for the presence of RespP in reaction moves. Significantly, ma can be felicitously used in self-talk, no matter whether it is I-centered, as in (57a), or you-centered, as in (57b).

For completeness, note that English too, has discourse markers used in similar contexts. For example, Schourup (1982]: 14) analyses oh as marking the presence of unspoken thought (see also Aijmer 1987). He provides the minimal pair in (58) (Schourup 1982: 15), which illustrates the contribution of oh within a typical conversation.

According to Schourup (1982]: 15), oh in (58b) indicates that the thought expressed in the following sentence just entered the speaker’s mind. Thus, with the use of oh a speaker may implicate that their failure to make the call was due to forgetfulness and not malevolent intent.

This description suggests that oh might be amenable to an analysis akin to Wiltschko’s (2024b) analysis of Upper Austrian ma. Specifically, oh is used to mark the utterance as a reaction and hence it may be analyzed as occupying SpecResp. If so, the examples in (59) provide evidence from English too, that RespP is licensed in self-talk, at least in the context of reaction moves.

In sum, we have now seen that you-centered self-talk differs from typical conversations only in that the need to regulate turn-taking is obviated. Consequently, the means to do that are not available, as evidenced by the impossibility of rising intonation in questions. This suggests that initiation moves are structurally deficient in self-talk with RespP missing. However, we have also seen that there is no categorical restriction on the projection of RespP, as evidenced by the fact that reaction moves can be introduced by units of language associated with RespP. This supports the view that self-talk may, in fact, be fundamentally interactional, just as regular conversations are, even though there is no other interlocutor to interact with.

4.4 Interim conclusion: an interaction-based typology of self-talk

In this section, we have seen empirical evidence for a three way-distinction in self-talk: I-centered self-talk differs from you-centered self-talk in that it does not license any linguistic phenomena that require the grammatical representation of an addressee (vocatives, imperatives, and addressee-oriented discourse markers). On the other hand, I-centered but not you-centered self-talk allows for the use of verbs of cognition. In addition, we have seen evidence for two types of you-centered self-talk. Specifically, the person engaged in self-talk can take on the role of the listener, with the speaker being construed as a disembodied (inner) voice. Alternatively, the person engaged in self-talk can take on the role of the speaker with an externalized picture or mirror image serving as the addressee. These two types of you-centered self-talk are empirically distinguished in the availability of social deixis (e.g., formal addressee-denoting pronouns): a disembodied voice is not a social being, and hence social deixis is not allowed when the person engaged in self-talk takes on the role of a listener. Finally, self-talk differs from typical conversations in that there is no need for the regulation of turn-taking, and hence, rising intonation is not available, though reactions can be marked as such.

These empirical differences (summarized in Table 5) receive a straightforward analysis within an interaction-based approach toward the topmost structure of an utterance, as summarized in Table 6. The difference between I-centered and you-centered self-talk resides in the presence versus absence of the addressee-oriented grounding phrase. The difference between typical conversations and self-talk resides in the presence versus absence of the other-oriented response phrase, the layer of structure that hosts the request for a response in initiating moves.

On this approach, the differences in modes of talking are structurally conditioned. The difference between the two types of you-centered self-talk derives from the assumption that the person engaged in self-talk may identify with either the listener or the speaker and hence be assigned the role of holder of Ground-Spkr or Ground-Adr, which are both available in you-centered self-talk. In this way, you-centered self-talk is no different from typical conversations, in which an individual too, may identify as the speaker or as the listener.

Significantly, what the empirical properties of self-talk reveal and what the structural analysis captures is that human language does not seem to have a dedicated means for self-talk. According to the analysis developed here, the different modes of self-talk reflect structural differences. But crucially, there is no dedicated structure for self-talk.^[17] If there were a dedicated structure for self-talk, we might expect it to take on the form of a reflexive construction of sorts. That is, languages have means to mark that event roles are assigned to one individual only and hence that the referent holding the agent role is identical to the referent holding the patient role. This type of reflexive marking in the event domain can be realized on the predicate or the internal argument in the form of a reflexive pronoun (Reinhart and Reuland 1993). To the best of my knowledge, there are no equivalent reflexive markers available for interactional roles. In other words, I know of no markers that would indicate that the individual holding the speaker role is identical to the individual holding the addressee role. On this view, then, self-talk itself is not special. Rather, the differences in grammatical representation are independently available. The structure associated with I-centered self-talk simply corresponds to a structure that represents a thought an individual may hold and express. Thus, it can be conceptualized as thinking out loud. In contrast, the structure associated with you-centered self-talk corresponds to a structure that represents a thought packaged for conversational interaction (i.e., embedded in a structure that makes the utterance sensitive to the presence of an addressee). In this way, you-centered self-talk can be conceptualized as having a conversation with oneself. And like in a typical conversation, the person engaged in this conversation may identify as the speaker or the listener.

Suggestive evidence for this typology of self-talk and for the assumption that the different modes of talking are associated with different grammatical and, thus, mental representations comes from the fact that the same distinctions are observed in covert self-talk (i.e., the phenomenon often referred to as inner speech, see Section 2.1).

Specifically, it is known that covert self-talk is not a unified phenomenon but instead that it is subject to substantial variability (Hurlburt et al. 2013). On the one hand, there is the often-discussed difference between condensed and expanded inner speech (Fernyhough 2004), which has no direct correlate in overt self-talk as the latter cannot be condensed. However, the differences of the type discussed here for overt self-talk have also been observed for covert self-talk. In the case of covert self-talk, the evidence does not come from linguistic considerations, however, but instead is based on neuroimaging data.

First, based on a neuroimaging study, Alderson-Day et al. (2016) establish a difference between monologic and dialogic covert self-talk. They show that in dialogical versions of covert self-talk, a wider network is involved than the classical regions associated with language (production and comprehension). Specifically, dialogic covert self-talk appears to involve regions responsible for Theory of Mind and social cognition. This appears to correlate with the distinction between I-centered and you-centered self-talk we have established for overt self-talk based on linguistic evidence.

In addition, Hurlburt et al. (2016) point out that in the literature, the term inner speech may refer to two phenomenologically and psychologically distinct phenomena, which they refer to as inner speaking and inner hearing, respectively. From the present point of view, this seems to correspond to the difference between the person engaged in overt self-talk identifying with the speaker or with the listener. This is corroborated by neuroimaging results that show evidence for two types of inner speech, one where production-related regions are activated, whereas another where perception-based regions are activated (Pratts et al. 2023). According to Kompa (2024]: 647), “[t]here is inner speaking and signing, which is accompanied by a sense of agency, and inner hearing or auditory imagery […], where one experiences oneself as being passive.” Fernyhough (2009) suggests that the latter type of self-talk is akin to a dialogue between the person engaged in self-talk, who identifies as a listener, and an inner interlocutor who is conceptualized as an ‘open slot’ fillable by any imaginary interlocutor.

I conclude that the typology of self-talk we have established here based on linguistic evidence has psychological correlates. These psychological correlates point towards the significance of viewing (some forms of) self-talk as intrinsically dialogical. I thus submit that an interaction-based analysis of the type developed here is well suited to capture not only the linguistic profiles of the different types of self-talk but also their psychological characteristics.

5.1 A commitment-based analysis of self-talk

To the best of my knowledge, Krifka (2023b) presents the only other explicit linguistic analysis of self-talk within a framework that assumes that some pragmatic properties are syntactically encoded at the very top of the tree. Drawing on the insights of Ritter and Wiltschko (2021) and using data from Schnitzler’s Leutnant Gustl, he also analyses the difference between I-centered and you-centered self-talk in terms of structural deficiency. At the core of his proposal is the claim that I-centered self-talk lacks CommitP, i.e., the phrase dedicated to encoding a speaker’s (public) commitment to the propositional content. Rather, I-centered self-talk consists of only JudgeP (encoding the speaker’s private perspective) and ActP (encoding the illocutionary force). This contrasts with you-centered self-talk, where CommitP is present, just as in typical conversations. This is schematized in (60) (adapted from Krifka 2023b).

Note that the core insight behind the interaction-based analysis had to do with the absence of the structural position responsible for introducing the addressee role (Ground_Adr). Crucially, Krifka’s commitment-based structure does not incorporate such an addressee-oriented phrase. Rather, in this framework, the corresponding structural positions distinguish between a private perspective and a public commitment, where the public nature of commitment arguably requires an addressee (otherwise, the commitment would not be public). On this view, then, the difference between I- and you-centered self-talk amounts to a difference between expressing one’s private thought (i.e., thinking out loud) and publicly committing to it (i.e., having a conversation with oneself). While this captures the core insight of the interaction-based analysis, in the absence of reference to an addressee in CommitP, the ungrammaticality of vocatives, imperatives, and other addressee-oriented units of language does not straightforwardly follow. This is not to say that the analysis could not be amended to capture these empirical differences, but it does not follow from the structural analysis alone. To be sure, the additional contextual variables that are incorporated in Krifka’s analysis will not help in this respect. Recall that Krifka’s commitment-based analysis is amended through contextual variables that introduce the relevant roles (speaker, addressee, judge, and committer) and stipulations of identity. Typically, the judge is identified as the committer. However, the identity of the committer differs depending on the illocutionary force. In assertions, the committer is identified as the speaker, while in questions, it is identified as the addressee. The relevant representations for assertions and questions in typical conversations are repeated below for convenience (Krifka 2023b: 350f.).

To capture the essence of self-talk, Krifka (2023b) amends the contextual variables. Specifically, I-centered self-talk is characterized by the absence of an addressee (hence, the contextual variable for the addressee is marked with strike-through in (61a)). In addition, Krifka proposes that the participants in you-centered self-talk can be identified as an inner self (‘Inneres Selbst’ IS) and an outer self (‘Äußeres Selbst’ ÄS). The inner self can have thoughts and feelings (expressed in I-centered self-talk), and the outer self can reflect upon these thoughts and is thus identified as the locus of self-awareness (Krifka 2023b: 2). For you-centered self-talk, Krifka assumes that it is the outer self which addresses the inner self. Hence, in assertions, the outer self is identified as the committer, as in (61bi), whereas in questions, the inner self (which functions as the addressee) is asked to be the committer, as in (61bii).

On this view, a dedicated addressee role is not part of the grammatical representation. Hence, it is not clear where the information that the outer self serves as the addressee comes from. But if there is no dedicated way to represent the addressee, we have no straightforward way to account for the ungrammaticality of addressee-oriented phenomena. It would have to boil down to a pragmatic explanation (e.g., the need for two participants), but it would have to be worked out how such a pragmatic account would incorporate a context with one individual that serves both roles via separating their self into an inner and outer self.

Another problem with this commitment-based analysis concerns the patterns of verbs of cognition. Recall that you-centered self-talk does not allow for the use of verbs of cognition. Roughly, this reflects the constraint on typical conversations that one cannot tell another what they are thinking. On the interaction-based analysis, it follows from the assumption that the addressee, when grammatically represented, is always treated as an inaccessible other mind. The problem with self-talk is that the person engaged in self-talk does have access to their own mind, and hence, the restriction on you-centered self-talk is difficult to understand on a purely pragmatic analysis. Under the commitment-based analysis, there is no grammatical representation of the addressee role, and hence, a grammatical analysis to capture this constraint is not available. Moreover, recall that Krifka explicitly postulates the contextual variables of an outer self addressing an inner self and that the former reflects on the latter. If so, we would expect that the inner self is transparent to the outer self, and hence, the restriction on the verbs of cognition remains a mystery.

Furthermore, the commitment-based analysis of you-centered self-talk in (61b) does not allow for a distinction between the two types of you-centered self-talk we have observed and their empirical correlates. In the interaction-based analysis, this distinction follows from the assumption that the person engaged in self-talk can identify either with the speaker or with the listener. This option is not available in the commitment-based analysis. If we were to speculate that, in this case, the person engaged in self-talk could identify with either the inner or the outer self, this would not explain why there should be a difference in the licensing of social deixis. Rather, we would expect that social deixis is ruled out in both types of you-centered self-talk since neither the inner nor the outer self can be viewed as social beings relative to each other.

A final empirical challenge the commitment-based analysis faces is that it fails to account for the difference in the availability of rising intonation. In the interaction-based analysis, this difference can be modelled through the availability of the topmost structure (RespP) in initiating moves. The commitment-based approach towards the very top of the tree has no equivalent structure. Thus, this difference must be accommodated differently.

Lastly, there is also a conceptual disadvantage the commitment-based analysis faces. It relies on the postulation of contextual variables that only seem to play a role in self-talk. That is, the variables for speaker and addressee available in typical conversations are replaced by inner and outer self, which are, by hypothesis, available in self-talk only. This amounts to saying that self-talk comes with its own grammatical ingredients. In contrast, the interaction-based analysis did not postulate any dedicated means for self-talk. I submit that this is a conceptual advantage of the interaction-based analysis over the commitment-based one.

5.2 A neo-performative analysis of self-talk

I now turn to a discussion of the neo-performative framework and its potential to analyze the properties of self-talk. To the best of my knowledge, there is no explicit analysis of self-talk available to draw upon. Hence, the discussion is restricted to what such an analysis might look like (subsection 5.2.1). There is, however, a neo-performative analysis available for self-directed questions. Given the apparent affinity between self-directed questions and self-talk, I shall include a discussion of this phenomenon as well (subsection 5.2.2).