Breaking the Language Barrier: From Pre-Linguistic Action Distinctions to Linguistic Structures ^†

Abstract

The widely accepted view in theoretical linguistics is that language is the main source of human unique capacities. It is often used to justify the assumption that language can be studied as an autonomous system. However, this view has been challenged by recent studies in animal cognition, reporting capacities in animals comparable to those of humans. Hence, the assumption that language can be studied as an autonomous system has to be re-evaluated. The goal of this paper is to examine what such a re-evaluation means for both linguistics and epistemology. I discuss a concrete example where this re-evaluation seems necessary. This example concerns the distinction between intentional and unintentional actions. This distinction guides the behaviour of animals and pre-linguistic babies, and it also determines the grammaticality of numerous linguistic constructions. I argue that the default hypothesis should be that the linguistic distinction between these two types of actions is based on the corresponding pre-linguistic distinction. I show that to formulate this hypothesis, theoretical concessions from both linguistics and epistemology are needed. I conclude that re-evaluation of the assumption that language can be studied as an autonomous system will require a careful, case-by-case re-examination of linguistic and non-linguistic distinctions and a parallel readjustment of their analyses.

1. Introduction

In theoretical linguistics, the prevailing view is that language is the root of humans’ unique capacities for counting, problem-solving, and complex social behaviour. This view is often used to justify the assumption that language can be studied as an autonomous system [1,2,3]. However, recent studies in animal and insect cognition have found that comparable capacities are manifested in non-human species [4,5,6]. In light of these findings, the assumption of the autonomy of language must be re-evaluated because there might be phenomena that are best captured by abandoning the autonomy-of-language assumption and emphasising the continuity between linguistic and non-linguistic capacities. My goal in this paper is to examine what such a re-evaluation means for both linguistics and epistemology by discussing a concrete example where it seems necessary to withdraw the autonomy-of-language assumption.

As an anonymous reviewer rightly notes, we should not confuse the methodological claim that language can be studied as an autonomous system with the substantial claim that language is an independent neural system (or module) in the brain. In this paper, I argue against the methodological claim; therefore, phrases such as “the assumption of the autonomy of language” and “the autonomy-of-language assumption” should be understood to refer to the methodological claim. My view is compatible with the view that language and thought are supported by separate neural systems in the brain (as argued in [7]), as long as the relevant mechanisms are homologous.

I will start in Section 2 by introducing my primary example, which concerns the distinction between intentional and unintentional actions. We will look at studies reporting that animals and pre-linguistic babies react differently to unwilling and unable conditions, suggesting that there is a non-linguistic distinction between intentional and unintentional actions. We will also look at purely linguistic phenomena that are sensitive to the intentional–unintentional distinction. The main claim of this section is that the default hypothesis is that the explanations for the linguistic and non-linguistic distinctions between intentional and unintentional actions ought to align.

In Section 3, I show that to develop a satisfactory analysis of linguistic phenomena sensitive to the intentional–unintentional distinction, a strong contextualist approach to language must be adopted. According to this approach, the meaning of a sentence depends on the context of utterance, not only in the case of essential indexicals or similar expressions, as standardly assumed [8], but also in the case of ambiguous event descriptions. The strong contextualist approach is a concession on the part of linguistics needed to account for the linguistic distinction between intentional and unintentional actions.

In Section 4, I show that the concession from linguistics is not enough for explaining the alignment between the linguistic and non-linguistic distinctions between intentional and unintentional actions. I argue that epistemology, in turn, must make a concession. In particular, epistemic contextualism, that is usually presented as a semantic view according to which the content of knowledge-ascriptions depends on the context of utterance, should be strengthened. In the strong version of epistemic contextualism, the knowledge itself, and not only the semantics of ‘know’ and related expressions, is shaped by contextual factors.

In Section 5, I discuss some preliminary conclusions that can be drawn from re-evaluating the assumption of the autonomy of language and from aligning the explanations of linguistic and non-linguistic distinctions.

2. The Primary Example

To question the assumption that language can be studied as an autonomous system, I make use of the distinction between intentional and unintentional actions. Wild and domesticated animals, as well as prelinguistic babies, seem to distinguish between these two types of actions. The acceptability of a range of linguistic constructions also depends on this distinction. Therefore, the distinction is a compelling example of a capacity that is both linguistic and non-linguistic.

The nature of the distinction is a complex topic in both linguistics and philosophy. In this paper, I will not address the question of what intentional actions are or how they should be analysed. It will be enough for my purposes to say that attributions of intention are tied to the recognition of mental states of others, their goals, and their abilities to control the action. Major book-length works on intention and intentional action include [9,10,11,12,13,14,15].

Studies on animals and pre-linguistic babies employ the unwilling–unable paradigm to test whether animals and babies recognise intention in action. This paradigm consists of three experimental conditions. In the unwilling condition, the experimenter intentionally withholds a reward. In the unable-clumsy condition, the experimenter tries but fails to give the reward due to clumsiness. In the unable-blocked condition, a physical obstacle prevents the experimenter from delivering the reward. If a subject recognises the experimenter’s intention, their behaviour differs between the unwilling condition and the two unable conditions.

Schünemann et al. found that dogs were more hesitant and waited significantly longer before approaching the reward in the unwilling condition than in the unable conditions [16]. These results suggest that the dogs recognised a difference in the experimenter’s intention. They hesitated to approach a reward in the unwilling condition because they predicted that they would not receive it. In contrast, in the unable conditions, the dogs approached a reward right away because they still expected to receive it. Similar experiments demostrate that other animals, including apes, monkeys, parrots, and horses, distinguish between intentional and unintentional actions [17,18,19].

As noted by an anonymous reviewer, caution should be exercised when interpreting the results of cognitive experiments with animals. The main principle of comparative animal psychology (often referred to as Morgan’s Canon) advises interpreting animal behaviour in the simplest possible terms, avoiding attributing higher psychological processes unless absolutely necessary. For a useful discussion of Morgan’s Canon and its pros and cons, see, for example, [20].

Pre-linguistic babies start recognising intention-in-action at around 9 months of age. Behne et al. found that 9-month-old infants reacted more patiently when the experimenter could not give them a toy because of an obstacle or because the experimenter was distracted (the unable conditions) than when the experimenter teased them with the toy (the unwilling condition) [21]. Other studies with different methodologies corroborate these results [22,23].

The experiments discussed above suggest that wild and domesticated animals, as well as pre-linguistic babies, distinguish between intentional and unintentional actions. This distinction is crucial for social interactions and joint actions. So it should not come as a surprise that languages have nuanced ways to communicate this distinction.

As linguistic creatures, we intuitively recognise that the sentence ‘Alice baked a cake’ describes an intentional action, whereas the sentence ‘Alice lost her keys’ describes an unintentional action. We also recognise that the sentences ‘Bob took an illegal left turn’ and ‘Bob deleted the email’ are ambiguous between intentional and unintentional interpretations. However, the linguistic distinction between intentional and unintentional actions goes beyond surface-level interpretations. It also affects the grammaticality of sentences with complex structures.

Here is a concrete example in which the intentional–unintentional distinction affects grammaticality. This example concerns so-called subject obviation—a restriction on having coreferential subjects in some attitude reports. This restriction is widespread across Indo-European languages and is especially common in Romance languages. For example, [24,25,26,27,28]. Consider the Italian examples in (1): (1)a is a baseline example that shows a belief report with the subject of the main and the embedded clauses referring to different individuals. This sentence is grammatical. But when the two subjects refer to the same individual, as in (1)b, the sentence becomes unacceptable, which is marked by the asterisk ‘*’. The Italian examples are from [29], pp. 29–30. SBJV stands for the subjunctive mood.

(1)

a. Penso che parta domani. (Italian)

‘I think that s/he leave-SBJV tomorrow’

b. * Penso che io parta domani.

‘I think that I leave-SBJV tomorrow’

A notable property of subject obviation is that it disappears when the action in the embedded clause is interpreted as unintentional. In the example (2), the main and the embedded subjects co-refer, but because the embedded action—making mistakes—is unintentional, the sentence is acceptable.

(2)

Penso che io abbia fatto molti errori. (Italian)

‘I think that I have-SBJV made many mistakes’

The sentences in (1)b and (2) are similar in structure. The difference lies in whether the embedded action is interpreted as intentional or unintentional. This difference affects the acceptability of the sentence, which suggests that grammar is sensitive to the intentional–unintentional distinction. An anonymous reviewer points out that the examples in (1)b and (2) use different tenses in the embedded clause and asks whether tense plays a role in the intentional–unintentional distinction. This is an interesting observation, and tense might influence the (un)intentional interpretation. However, it cannot be the sole factor because a contrast similar to that in (1)b and (2) can be found with the same (progressive) tense in the embedded clause, as shown in (2′).

(2′)

a. * Penso che io stia leggendo il giornale. (Italian)

‘I think that I stay-SBJV reading the newspaper.’

b. Penso che io stia scrivendo alla persona sbagliata.

‘I think that I stay-SBJV writing to the wrong person.’

The intentional–unintentional distinction to which grammar is sensitive is not tied to the lexical meanings of verbs. The same lexical material, which can be interpreted as describing either an intentional or an unintentional action depending on the context, has a similar effect. For instance, ‘Bob took an illegal left turn’ can describe a situation where Bob intentionally broke the law or a situation where he turned left without realising that the turn was illegal. In (3), the embedded event description ‘took an illegal left turn’ is ambiguous, and the sentence is acceptable in contexts where the agent took an illegal left turn unintentionally, but it is unnatural in contexts where the agent knowingly broke the law. In (3), sensitivity to the context-dependent interpretation of the embedded action is marked by the hash sign ‘#’.

(3)

# Penso che io stia facendo una svolta a sinistra illegale. (Italian)

‘I think that I am-SBJV making an illegal left turn.’

Grammatical constructions that are sensitive to the intentional–unintentional distinction and show the same pattern as subject obviation in Italian are found across different languages and grammatical domains (see [30] for an overview). This widespread pattern of intention-sensitivity in grammar suggests that language has a mechanism to represent the distinction between intentional and unintentional actions. But is this mechanism related to the one that allows animals and pre-linguistic babies to distinguish between intentional and unintentional actions? I want to suggest that the default hypothesis is that the two mechanisms share the same origin. That is, language should not be regarded as a fully autonomous system; the representations used in communication are evolutionary and developmental extensions of distinctions already available without language. In my concrete example, this suggestion means that the linguistic distinction between intentional and unintentional actions is an extension of the non-linguistic distinction between these two types of actions found in animals and pre-linguistic babies.

To formulate the default hypothesis about the common origin, we need a framework that allows us to formalise and compare both linguistic and non-linguistic representations. A good candidate for such a framework is a Kripke-style knowledge model, which represents the knowledge of agents by means of possible worlds and the relations between them. This model has been successfully used in both linguistics and epistemology, and has also been applied to cognition (See [5] for a discussion of the common origin of perceptual and propositional representation). However, as we will see in the next two sections, to use the model as a general framework for formalising and comparing linguistic and non-linguistic representations, concessions are required from both linguistics and epistemology in how they interpret this model.

3. The Concession from Linguistics

In formal linguistics, the meaning of a sentence is standardly assumed to be a proposition. Propositions are functions from possible worlds to truth values; equivalently, one can say that the meaning of a sentence corresponds to a set of possible worlds in which the content of the proposition is true. For example, if w is a possible world, the meaning of the sentence ‘Alice baked a cake’ corresponds to the set {w: Alice baked a cake in w}. The content of a proposition is determined compositionally, by combining lexical expressions, such as ‘Alice’, ‘bake’, and ‘cake’, with non-lexical material, such as Tense and Determiners, according to a predefined set of rules. The resulting structure is called a logical form or LF. LFs are particularly useful for representing ambiguous sentences, such as ‘Everyone loves someone’, which can mean either that for every individual, there is someone that person loves, or that there is a particular person such that everyone loves him or her.

We now consider how the linguistic distinction between intentional and unintentional actions can be represented in the standard system sketched above, and whether this representation can also apply to the parallel non-linguistic distinction. One strategy to represent the difference between sentences like ‘Alice baked a cake’ and ‘Alice lost the keys’ would be to attribute this difference to the lexical meaning of the verbs ‘bake’ and ‘lose’. For example, we could say that ‘bake’ has an additional lexical component that indicates that baking is performed intentionally. The verb ‘lose’ lacks such a component, and thus the action it denotes is typically interpreted as unintentional. The precise content of this additional lexical component depends on how the lexical strategy is implemented. Some accounts represent it as a responsibility relation linking the agent and the event [31]. Others argue that verbs expressing intentional actions encode a decision component [32]. Still others analyse intentional interpretations as involving control or causation [33,34].

The lexical strategy, however, has an unwelcome consequence: if we wish to extend it to ambiguous sentences, such as ‘Bob took an illegal left turn’ or ‘Bob deleted the email’, we would have to assume that ‘take’ and ‘delete’ each require two distinct lexical entries. Moreover, the same lexical ambiguity would have to be postulated for most verbs in the language, since the majority of verbs allow both intentional and unintentional interpretations [35].

Another strategy would be to say that ‘Alice baked a cake’ and ‘Alice lost the keys’ have different LFs, and that this difference in LFs gives rise to the distinct interpretations. This strategy can be extended to ambiguous sentences, since LF is a suitable tool for analysing ambiguous constructions. However, the two-LF strategy has a drawback: it is difficult to justify. There are three main sources of LF ambiguity: (i) scope ambiguity, as illustrated earlier by ‘Everyone loves someone’; (ii) structural ambiguity, when the elements of a sentence are combined in different orders, as in ‘Clara saw the man with the telescope’, which can be analysed either as ‘[Clara saw [the man] with the telescope]’, where the act of seeing is carried out with the telescope, or as ‘[Clara saw [the man with the telescope]]’, where the man has the telescope; and (iii) the resolution of context-dependent elements such as the pronoun ‘he’ in ‘Bob told Carl that he was a fool’. In ambiguous examples such as ‘Bob deleted the email’, there are no scope-bearing elements that would justify the first type of ambiguity. Nor is there reason to suspect that these examples involve structural ambiguity of the second type. One might suggest that ‘delete’, like ‘he’, is context-dependent and thus appeals to the third type of ambiguity. In this case, ‘delete’ would be interpreted as intentional in some contexts and as unintentional in others. But this strategy conflicts with the standard view that context-dependent expressions form a closed class (essential indexicals, predicates of personal taste, and modal auxiliaries, among others), whereas linguistic expressions that form an open class (common nouns, adjectives, and verbs) are not context-dependent [8] (See [36,37] for proposals that challenge the standard view).

An anonymous reviewer suggests that there might be a moderate semantic alternative involving event semantics or “lexical mechanisms that encode responsibility/agency and are resolved in context”. As far as I can tell, such a semantic alternative would be subsumed under the lexical or the two-LF strategy discussed above, and similarly, it would not allow us to represent the non-linguistic distinction between intentional and unintentional actions.

Both the lexical and the two-LF strategy face significant difficulties. But even if we overlook their shortcomings, neither strategy can be used to represent the non-linguistic distinction between intentional and unintentional actions. This is because each strategy places the distinction into the linguistic material—either into lexical expressions themselves, which are unavailable to creatures without words, or into linguistic structures, which are inaccessible to creatures without language. Yet linguistic material is not the only means of explaining interpretive differences; an alternative is to locate the distinction in the context in which the interpretation is made.

When sketching the standard linguistic system at the beginning of this section, according to which the sentence ‘Alice baked a cake’ is represented as the set of possible worlds {w: Alice baked a cake in w}, I left unspecified the set to which w belongs. Traditionally, it is assumed that w is a member of the set of all possibilities W. Thus, a more precise representation of our example in the standard system is {w ∈ W: Alice baked a cake in w}. We can think of the set W as a logical space that we, as attributors, interpreters, or theoreticians, divide into fully specified possibilities. In this system, sentences with different lexical expressions or different LFs result in different propositions because they pick out different subsets of W, and within this system, there is no way, independent of the linguistic material, to represent a difference in interpretation.

An alternative to the standard system is the contextualist system, according to which the content of a sentence depends on the way the logical space is divided. To be concrete, let me outline a version of the contextualist system developed by Stalnaker for explaining various linguistic phenomena [37]. According to this view, context is not used merely to determine unspecified meanings, as, for example, in the case of reference resolution with pronouns, as we saw above. Rather, context is the material out of which propositions are constructed. For example, the proposition expressed by the sentence ‘Alice baked a cake’, according to this view, does not correspond to a subset of all possibilities W; rather, it corresponds to a subset of those possibilities that the conversational participants use to represent information relevant to their communication. Let CS (for a ‘context set’) be a set of possibilities into which the conversational participants divide the logical space for the purposes of their conversation. Then, the sentence ‘Alice baked a cake’ expresses the proposition corresponding to the set {w ∈ CS: Alice baked a cake in w}. The partitioning of the logical space in CS has to take into consideration both how the participants of a conversation view possibilities and how we, as attributers or interpreters, view them.

A useful (for us) feature of the contextualist system is that the same LF can result in different meanings when it is evaluated with respect to different contexts. This feature allows us to represent the distinction between intentional and unintentional actions in ambiguous sentences without appealing to the difference in linguistic material. We can say that sentences like ‘Bob deleted the email’ have one LF and unambiguous lexical meanings; the difference in interpretation arises from the difference in how CS is determined. That is, under the intentional interpretation, ‘Bob deleted the email’ corresponds to the subset of possible worlds in which Bob deleted the email relative to some context CS₁, {w ∈ CS₁: Bob deleted the email in w}. The same sentence under the unintentional interpretation expresses the proposition that corresponds to a subset of possibilities relative to a different context, CS₂, {w ∈ CS₂: Bob deleted the email in w}. In other words, we shift the difference in meaning from the sentences to the contexts.

How to explain the difference between CS₁ and CS₂ depends on an analysis of the nature of intentions and intentional actions. This question is beyond the scope of this paper since the contextual strategy is independent of the exact nature of intentional actions. However, let me list a number of options that have been discussed in the literature. One option would be to say that in CS₁, where an ambiguous sentence is interpreted as intentional, but not in CS₂, where it is unintentional, the conversational participants commonly believe that the agent is aware of their action or is engaged in practical reasoning [9,38,39]. Another option would be to say that in CS₁, but not in CS₂, the conversational participants take into consideration the agent’s goals or their plan of action [11,13,40]. Finally, to distinguish between CS₁ and CS₂, we could use the notions proposed in linguistics, such as responsibility, decision problem, and control [31,32,33,34], but apply them to the context rather than the lexical material itself.

Whichever explanation turns out to be more satisfactory, the important feature of the contextualist strategy is that the common knowledge of conversational participants affects the partitioning of the space of possibilities in such a way that the same linguistic material (that is, a sentence with the same lexical expressions and an unambiguous LF) is interpreted as intentional or unintentional depending on the context. This strategy disassociates the intentional–unintentional distinction from linguistic material and provides a formalisation that can, in principle, be used in a framework that unifies linguistic and non-linguistic representations. This step forward, however, comes at a cost that sentences do not have absolute meanings because the propositions they express are context-dependent. In the next section, we are going to see that a unifying framework requires a similar concession from epistemology.

4. The Concession from Epistemology

Let us now look at the situation from the other end. How can we formalise the distinction between intentional and unintentional actions in animals and pre-linguistic babies? Let us assume, for simplicity, that we can use propositional representation for this distinction. As we saw above, animals like dogs, monkeys, and parrots, as well as pre-linguistic infants, react differently in unwilling and unable conditions. So our task is to determine how their representation of an action in the unwilling condition differs from their representation of the same action in the unable condition. In what follows, I suggest that we can accomplish this task by adopting the relatively recent view in philosophy called epistemic contextualism (EC).

EC is the view that knowledge-attributions can change their contents depending on the context of attribution. In its most frequent form, EC is a semantic view according to which a knowledge sentence such as ‘S knows that p’ expresses a complete proposition (and thus can be assigned a truth value) only relative to a contextually determined standard. In this sense, the predicate ‘know’ and related expressions are similar to essential indexicals—expressions such as ‘now’ and ‘that’—which require the context for determining their content. EC has developed as a response to a sceptical paradox. Here is one version of that paradox: the argument in (4) appears to be valid with seemingly plausible premises, yet the conclusion is counterintuitive.

(4)

P1: I don’t know that I am not a brain in a vat.

P2: If I don’t know that I am not a brain in a vat, I don’t know that I have hands.

C: I don’t know that I have hands.

According to EC, the conclusion in (4) is counterintuitive because of the shift in the context-sensitive standard for knowledge attribution. When the sceptical possibility in P1 is considered, we are in a context that requires a high standard for knowledge attribution. In this context, both premises and the conclusion are true. However, in ordinary contexts, in which sceptical possibilities are ignored, the standard for knowledge-attribution is relatively low. In these everyday contexts, ‘I know that I have hands’ is true. The conclusion in (4) and the everyday statement that ‘I know that I have hands’ do not contradict each other because the truth value of knowledge sentences is context dependent, and the two statements are evaluated relative to different contexts.

However, EC in its semantic form does not bring us closer to representing the intentional–unintentional distinction in non-linguistic creatures, because it concerns the use of lexical predicates like ‘know’ and its kin, which non-linguistic creatures do not have. We need to adopt a stronger strand of EC, according to which not only the assignment of a truth value to knowledge sentences but also the content of a knowledge-attribution depends on the context of attribution. Here, the context of attribution is understood as a partitioning of a logical space that aligns with the attributor’s and the subject’s view on which possibilities are available in the context.

With the stronger strand of EC, we can say that in the unwilling and unable conditions, animals and pre-linguistic babies know different propositions, although these propositions are about the same action. The propositions are different because their content depends on the context of attribution, which is different in the unwilling and unable conditions.

As discussed above for different linguistic contexts, exactly which feature or features distinguish intentional and unintentional contexts depends on a concrete analysis of the nature of intentions and intentional actions. What is important here is that with the concession to a stronger version of EC and the linguistic concession discussed in the previous section, we can develop a framework that will allow us to formulate and evaluate the default hypothesis that the distinctions between intentional and unintentional actions in language and in non-linguistic behaviour share a common origin. In this common framework, the investigation can proceed by comparing contexts in which intentional and unintentional event descriptions are used with contexts in which different reactions in unwilling and unable conditions are observed.

5. Conclusions

This paper started with the remark that the assumption that language can be studied as an autonomous system needs re-evaluation, given recent research on animal and insect cognition. My main goal was to examine what such a re-evaluation means for linguistics and epistemology, using a concrete example. My example concerned the distinction between intentional and unintentional actions, which has a non-linguistic origin since animals and pre-linguistic babies recognise this distinction. But it is also relevant to linguistic representations since it affects the grammaticality of numerous linguistic constructions. I suggested that the default hypothesis should be that the linguistic and non-linguistic distinctions come from a common origin. I argued that to formulate and assess this hypothesis, a unifying framework is needed and sketched a way to develop such a framework by appealing to contextualist views in both linguistics and epistemology.