Revisiting the epistemic gap: It’s not the thought that counts

ABSTRACT

This paper revisits the longstanding observation that children produce modal verbs (e.g., must, could) with their root meanings (e.g., abilities, obligations) by age 2, typically a year or more earlier than with their epistemic meanings (e.g., inferences). Established explanations for this “Epistemic Gap” argue that epistemic language production is delayed because small children can only reason about root meanings. However, root and epistemic uses of modal verbs also differ syntactically and in input representation. We present a corpus study on 17 English-learning children and their input, exploring early productions with both epistemic modal verbs and grammatically simpler and more frequent epistemic adverbs (e.g., maybe, probably). Results show that children use remarkably adult-like epistemic adverb sentences from even before age 2, when they are only producing modal verbs with root meanings. The Epistemic Gap is not well explained by general conceptual advancements. Instead, our data suggest input attestation and ease of form-to-meaning mapping may influence early child epistemic language. These findings are furthermore consistent with cross-linguistic differences in the timing of first epistemic uses of modal verbs, and with recent advancements in our understanding of infant and toddler modal reasoning abilities.

Acknowledgments

I extend thanks to my amazing RAs at the Child Language Lab @ NYU, Kathryn Rafailov and Michael Marinaccio, who helped with coding the data. For valuable discussions and insights, I thank the members of the ModSquad at UMD and NYU, especially Anouk Dieuleveut, Annemarie van Dooren, Valentine Hacquard, Maxime Tulling, and Dunja Veselinović, all of whom I’ve done related corpus work on modality with. And Dan Harris, Brian Leahy, Ana Pérez-Leroux, Tom Roeper, Petra Schulz and her lab, and Sandrine Tailleur.

Notes

1 Within root modalities there are differing subclassifications of meanings by framework or author—for example: dispositional, deontic, teleological, desire-intentional, and circumstantial, among others.

2 We use the term polysemous descriptively.

3 Another factor may be that small children do not care to talk about thinking; we set this aside because it should affect all grammatical categories used to express epistemics uniformly (see Hickmann & Bassano 2016 for overview).

4 There are some minor exceptions to this claim, but they are not relevant to our sample.

5 Theory of Mind (ToM) components often come up in relation to epistemic reasoning. Note that while we use epistemic reasoning and language when talking about inner states (e.g., Dino must be sleepy), reference to inner states is not a necessary condition for epistemicity. For example, It must be raining is also epistemic despite no explicit inner state reference; the speaker expresses an inference based on what they know/perceive (e.g., seeing wet umbrellas).

6 These are standardly called “epistemic,” which doesn’t fully square with the linguistic approaches to modal classifications taken here. Deontic versus Epistemic in this literature is best described as being about “pure” descriptive reasoning (called epistemic) versus socially embedded prescriptive reasoning (deontic) (see Dack & Astington 2011). N.B. that deontic tasks often include modal language (e.g., must), but the epistemic ones are typically plain conditionals.

7 There are some exceptions here, not of central importance to us—for example, the adjectival to be possible.

8 A root interpretation is also possible but requires additional temporal operators to shift the temporal orientation of the prejacent to future-oriented (e.g., By tonight) or contextual support to the same effect.

9 Belief verbs like think, know, and believe have been studied largely independently from functional modals in acquisition (Perner & Wimmer 1985; Shatz, Wellman & Silber 1983; De Villiers 2007, for overview). Think and know are the most frequent epistemic modals of any category in English usage, including to children (van Dooren et al. 2017), however, a substantial proportion of their usage occurs in formulaic expressions like I don’t know or with otherwise epistemicity-obscuring speaker meanings (see Dudley 2017).

10 Note we will call modal adverbs “weak” and “strong” rather than “possibility” and “necessity.” This is mostly due to probably, which is not a clear necessity modal because it is a gradeable modal (see Lassiter 2010).

11 Moore, Pure & Furrow (1990) is primarily a modal force behavioral study, but they compare among different syntactic categories (adverbs, belief verbs, modal verbs) for belief tasks. They find correlations between the three categories.

12 Literature on attitude verbs (lexical modals) parallels the modal verb literature, with early observations showing root attitude verbs (e.g., desire-intentional modals like want) preceding the belief verbs (e.g., epistemics like think, know) in production (De Villiers 2007, for overview). Also like with modal verbs, explanation is standardly attributed to conceptual difficulties with belief reasoning despite additional grammatical and pragmatic differences (see Hacquard & Lidz 2016). Likewise, literature on direct versus indirect evidentials (knowledge-source marking, with indirect types requiring inference from evidence; see Aikhenvald 2004) parallels the modal verb literature for both observational asymmetry (direct > indirect knowledge sources) and conceptually grounded analyses (see Aksu-Koç & Slobin 1986; also Ünal & Papafragou 2016), and syntactic approaches (Rett & Hyams 2014).

13 The corpus begins at 1;09, but van Dooren et al. (2017) sampled from 2;00.

14 And possibly, but no children used this modal.

15 We rely on age as a between-child comparison to align with previous studies.

16 We did not use the first of repeated uses (“FRU”) metric because epistemic uses are rare and semantics is paramount (unlike for, e.g., productivity of regular past tense marking). Instead we rely on our “sureness” metric.

17 The MacWhinney corpus reports from newborn. To better align with other samples, we sample input from 1;00.

18 With one exception. We excluded futures (going to, shall, will) and hypothetical/counterfactual would to save efforts because these yielded no relevant data in the child samples.

19 Coders only disagreed on a few items (n = 8). This was a “find the first usage” search rather than a coding of a predefined sample, so a Kappa test was not run. Items coders disagreed on were discussed. In some cases a coder missed a clue to meaning. Others were genuine gray cases, which we treated as not eligible for “2”-level sureness.

20 Some children used no epistemic MVs that passed our criteria for a clear use, but did use MVs with low sureness ratings. These children were: Alex (might at 2;10,25), Emily (have to at 2;06,06, might at 2;09,02), and Trevor (have to at 3;03,12). All were either unclear in general (might) or were too likely intended as root (both have to uses).

21 All modal sentences were double-coded, and the interrater reliability for the raters was Kappa = 0.90.

22 All modal sentences were double-coded, and the interrater reliability for the raters was Kappa = 0.92.

23 Since epistemic MVs have both root and epistemic meanings, it was crucial to hand-code all items, but for the modal adverbs we only did this for one parent (Abe’s Father) to get a sense of the data.

24 We sampled can, but as only negated occurrences were epistemic, we list this modal more narrowly as can’t.

25 None of these were rated as “2.” Both coders agreed there was always a possible deontic flavor, but both coders also agreed uses like “That’s supposed to be a pig” in contexts where the speaker or addressee has no agency over the present-oriented prejacent event (i.e., not while drawing a pig, but when looking at a drawing) showed inferences.

26 The chi-square assumption of independence of observations is sometimes thought to be violated by corpus samples, as the same speaker supplies multiple uses per cell. However, this is not straightforwardly a true violation as we are comparing spontaneous utterances, not individuals, so each independent use is taken as a proxy for independence.

27 Sweetser’s (1990) theory of metaphorization is meant to apply to specific lexical items’ meanings, so this is consistent with root-before-epistemic applying only to individual modal verbs. However, her work is couched in the larger context of conceptual priority of physical over mental.

28 Similarly, while direct comparison is not possible given different sampling processes and time ranges, children’s first uses of EMVs (Table 5, Examples 17 and 18) largely reflect which EMVs are used most often in the input (Table 16). For adults, might and must are the most frequent EMVs in terms of raw counts. In the main study, including the might uses, children’s earliest EMVs include eight might, five must, one may, one could (Table 5).

29 A generous analysis is that children know the suggestion uses are possible precisely because maybe encodes epistemic possibility. Maybe serves the purpose of politely contributing a demand (e.g., going to the park) by hedging it as possibility that may or may not come true (e.g., maybe we can go to the park) (see also Dieuleveut et al. 2019).

30 N.B. that “deontic” is often used as a cover term for all root modalities, e.g., in Neill & Atance 2000.

31 There are some principled exceptions, like counterfactuals (see van Dooren et al., submitted).

Additional information

Funding

This work was supported by the National Science Foundation [BCS-1551628].