Learning dimensions of meaning: Children’s acquisition of but

Connectives such as but are critical for building coherent discourse. They also express meanings that do not fit neatly into the standard distinction between semantics and implicated pragmatics. How do children acquire them? Corpus analyses indicate that children use these words in a sophisticated way by the early pre-school years, but a small number of experimental studies also suggest that children do not understand that but has a contrastive meaning until they reach school age. In a series of eight experiments we tested children's understanding of contrastive but compared to the causal connective so, by using a word learning paradigm (e.g., It was a warm day but/so Katy put on a pagle). When the connective so was used, we found that even 2-year-olds inferred a novel word meaning that was associated with the sentence context (a t-shirt). However, for the connective but, children did not infer a non-associated contrastive meaning (a winter coat) until age 7. Before that, even 5-year-old children reliably inferred an associated referent, indicating that they failed to correctly assign but a contrastive meaning. Five control experiments ruled out explanations for this pattern based on basic task demands, sentence processing skills or difficulty making adult-like inferences. A sixth experiment reports one particular context in which five-year-olds do interpret but contrastively. However, that same context also leads children to interpret so contrastively. We conclude that children's sophisticated production of connectives like but and so masks a major difficulty learning their meanings. We suggest that discourse connectives incorporate a class of words whose usage is easy to mimic, but whose meanings are difficult to acquire from everyday conversations, with implications for theories of word learning and discourse processing.


Introduction
To construct coherent discourse, we rely critically on connectives: words such as but, so and because.Connectives cue how phrases are contingent on one another, and contribute importantly to the overall interpretations of their utterances, as illustrated in the following three examples: 1a.Samuel teased Andrew so John scolded him.1b.Samuel teased Andrew because John scolded him.1c.Samuel teased Andrew but John scolded him.
Here, the same two clauses relate to each other in different ways, with the different connectives cueing different discourse coherence relations between them (Asher & Lascarides, 2003;Hobbs, 1979;Kehler, 2002;Mann & Thompson, 1988;Prasad et al., 2008;Sanders, Spooren, & Noordman, 1992;Wolf & Gibson, 2005).In (1a), so highlights the role of the second clause as a causal consequence of the first clause, in (1b) because highlights the first clause as a consequence of the second, and in (1c) but is used to contrast the meanings of the clauses.Importantly, these coherence relations can have verifiable impacts on how adults interpret the sentences: intuitions about whether the pronoun him refers to the subject (Samuel) or the object (Andrew) of the first clause typically vary based on the discourse relation that the connective indicates (Kehler, Kertz, Rohde, & Elman, 2008;Koornneef & Sanders, 2013;Winograd, 1972).In (1a) and (1c), for example, English-speaking adults tend to infer that the pronoun refers to the subject, while in (1b) they are more likely to infer that the pronoun refers to the object.
To explore this in more depth, we turn to an example of a coherence relation whose production and comprehension behaviour appear particularly asymmetrical and contradictory: contrast, the adversative relation that is typically marked in English by the connective but.As reviewed below, children appear to produce but in an adult-like way by age 3, suggesting that it is acquired quickly and easily.However, another body of comprehension work suggests that children do not actually understand the meaning of but until perhaps five years later.We then present a series of experiments designed to resolve this mismatch, by accounting for theories of children's sentence processing and analyses of the meaning of but.To foreshadow our results, we find that, indeed, children do not reliably interpret but to convey a contrastive discourse relation until they reach school age.This difficulty is not explained by basic methodological considerations, by children's language processing skills, nor by their ability to generate pragmatic inferences.Instead, we suggest that discourse connectives represent a class of words whose usage is easy to mimic, but whose meanings are extremely difficult to acquire, with implications for theories of word learning and discourse processing.

The acquisition of but
A contrast between two discourse arguments is achieved when there is a conflict or contradiction between the propositions of the first and second segment.For example, in sentence (2) below, the first clause introduces a situation (hot weather) for which a typical association (wearing light clothes) is in conflict with the true circumstances (wearing a winter coat).
2. [It was a hot day] UTTERANCE 1 but [Katy was wearing a winter coat] UTTERANCE 2 .
To achieve a contrastive reading between two discourse arguments, English speakers rely on a range of explicit linguistic expressions (e.g., however, even though, yet) of which the connective but is the most basic (Fraser, 1988).Often, the precise way in which two utterances contrast is underdetermined, and must be inferred by the listener, but the presence of a word like but highlights that the interpretation of the second clause must be understood to contrast in some important way with the interpretation of the first, even if the precise sense of contrast is not clear.By contrast, if but were replaced with so in (2), then the two clauses would cohere such that the second clause describes the (anomalous) consequence of the first, while if because were used then the second clause would be (nonsensically) explaining the first clause.
From observation of their spontaneous production, it would appear that children acquire the ability to signal contrastive discourse relations with the word but by about the age of three.Corpus studies suggest that when children begin to combine utterances, around two years of age, they initially only mark discourse relations with the connective and (e.g., Bloom et al., 1980;Braunwald, 1985), and only use it to express additive relations in which the meaning of the conjoined sentences is a simple combination of the meanings of the individual sentences (e.g., They're taking a vacuum cleaner to wipe and puppy dog's running, Bloom et al., 1980, p. 244).Soon, however, and becomes ubiquitous and children use it to express a range of different meanings, which come in a fixed order, first temporal relations (e.g., Jocelyn's going home and take her sweater off), then causal relations (e.g.,She put a bandaid on her shoe and it maked it feel better) and finally adversative relations expressing contrast between two events (It can't… and this can go here) (Bloom et al., 1980).Then, by age 3, children begin to use multiple other connectives, including explanatory so, followed by contrastive but, which is also used in additive and causal contexts (Bloom et al., 1980;Braunwald, 1985).While children continue using both connectives but and and to mark adversative relations, the consensus from observation of spontaneous production is that they appear to use but accurately well before they enter school, in contexts that allow one statement to modify or contrast with a previous statement (French, 1981;Gallivan, 1986;French & Nelson, 2012).
In stark contrast to this, those studies that have examined children's comprehension of but indicate that even school-age children struggle to understand its contrastive meaning.For instance, Kail and Weissenborn (1984) tested French-and German-speaking children on their knowledge of the contrastive but (French mais/German aber) and found that, in both languages, children did not master the connective before age 9.Even older 7-year-olds, the study shows, often interpreted but as and: in a sentence completion task, they typically produced supportive rather than contrastive continuations after hearing a prompt that finished with but.Similarly, in a study looking at children's reading, Cain and Nash (2011) found that 8-year-olds had particular difficulty in judging whether adversative connectives like but and although were being used appropriately or inappropriately.And finally, using spoken language and a pronoun-resolution task, Spenader (2018) found that younger Dutch-speaking school children did not interpret but (Dutch maar) contrastively.When asked for the referent of he in the example Big Bird asked Ernie for money but he didn't have any money, 7-year-olds were at chance, as compared to adults who assumed that the pronoun referred to the object of the first clause (Ernie), and also as compared to children's understanding of sentences with the causal connective because (Dutch want).Thus, these and other comprehension studies, using a range of methods and languages, have consistently found that children struggle to interpret but contrastively well into school age (Florit, Cain, & Levorato, 2017;Champaud & Bassano, 1994;Janssens, Drooghmans & Schaeken, 2015).Taken together with the results from children's spontaneous speech, there appears to be a striking asymmetry between children's accurate early production of contrastive discourse relations and their lasting difficulty understanding those same contrastive relations.

The but asymmetry
Why is there such an asymmetry between children's facility producing connectives like but and their apparent difficulty understanding those same words?One possibility is that the comprehension studies discussed above are actually underestimating children's competence, because the tasks provided are too complex or pragmatically infelicitous (cf.French, 1981;French & Nelson, 2012;Gallivan, 1986).For example, Kail and Weissenborn (1984) tested children's knowledge of the contrastive connective using sentence completions and acceptability judgments, which were themselves embedded in a complex context story that had a number of potentially confusing topics and alternative answers.Examining the materials, it is plausible that understanding those stories and distinguishing between acceptable and unacceptable sentences may have been challenging as it required a significant degree of realworld reasoning and metalinguistic, not only linguistic, awareness.By contrast, while Spenader (2018)'s pronoun resolution method was far simpler, the interpretation of ambiguous pronouns is itself challenging, as indicated, for instance, by the weak performance of the 7-year-olds on a control condition of the task.Thus, it may be that children would show more adult-like comprehension of butcomprehension that matches their skill at productionwhen they are assessed with more naturalistic tasks or with materials that are syntactically simpler and more pragmatically felicitous.
However, there are also good reasons to think that children really may struggle to comprehend but until the school years, and that their apparent facility with production may not be the strongest indicator of their linguistic competence.Most pertinently, learning the meaning of but may be difficult because what that word conveys is both variable, with a high degree of context dependence (Seligman, 1985), and subtle, with no consensus description amongst semanticists (e.g., Bach, 1999;Potts, 2007;Lakoff, 1971).
The variability of but's meanings can be seen in Seligman's (1985) analysis, which proposes that but is used in at least seven distinct yet related ways, including marking contrast between lexical items, such as John is tall but Bill is short (also known as semantic opposition), and contrast with an expected inference, such as John is tall, but he is not good at basketball, where the listener must infer that when someone is tall, they are expected to be good at basketball (see also Lakoff, 1971).That but has a variety of meanings can be seen in the fact that some of its senses are expressed by separate lexical items in other languages (e.g., where German uses aber to express most of the senses of English but, it uses sondern to express correction, as in I don't want an apple but an orange).Thus, given the variety of contexts in which but is used, it may be difficult for children to acquire it.
Another argument for children's difficulty acquiring but may be the subtlety of its semantic content.Semantic theories of but have assumed that while it is contrastive, that contrast may not be explicitly encoded in the truth-conditional contribution of the word.Frege (1956), for example, proposes that the meanings of but and and are almost identical, except that the former suggests that 'what follows is in contrast with what would be expected from what preceded it' (p.296).Grice (1975), following this, labelled the contrast carried by but as a conventional implicature, i.e., an implicature triggered by the meaning of the word, rather than by its context of use.The idea that the contrast indicated by but is a conventional implicature is now quite widespread.However, many current semantic theories (Bach, 1999;Potts, 2007;Wilson, 2016) favour an alternative approach in which the contrast expressed by but is instead semantic, although carried along a different dimension from the primary truth-conditional meaning of the word.For instance, Potts (2007), following Bach (1999), argues that the contrast in but cannot be a conventional implicature, because it appears to be part of the truth-conditional content of a sentence (e.g., if Sami tells you that Bert is crazy but effective, then it would be wrong to report Sami said that Bert is crazy and effective suggesting that the contrast is a part of the at-issue meaning of the sentence).Relatedly, in the Relevance Theoretic tradition, it has been suggested that but's contrast is a type of so-called procedural meaning (e.g., Blakemore, 1989), a conventional meaning that does not express content, but that instead guides the comprehender in how they should draw inferences.
These theoretical perspectives differ, but from the standpoint of language learning they all provide the same basic insight: that the contrast expressed by but represents an additional dimension of meaning, beyond the primary meaning communicated by the speaker.And given that, it is reasonable to assume that but's contrast will be difficult for children to master because related phenomena, such as scalar implicatures over quantifiers like some, are also somewhat difficult for children to master (e.g., Noveck, 2001;Lidz & Musolino, 2002;Papafragou & Musolino, 2003;Horowitz, Schneider & Frank, 2018).If but's meaning is indeed secondary, complex, and difficult to learn, then this would explain why comprehension experiments on children's understanding of but do not typically reveal adult-like knowledge until well into the school years.
However, this perspective would still leave unexplained why children seem to produce but with facility.If children do not actually understand its meaning, then surely we should expect their production to be different from adults'.One possibility, of course, is that prior work on children's production has somehow overestimated their facility with the term, perhaps because adult speakerscaregivers and researchers alikemay charitably recognise at least some level of contrast when but is used to juxtapose two clauses.
As such, studies to date have not satisfactorily clarified children's acquisition of but.Studies of production suggest adult-like competence, but it could be the case that this is an over-interpretation.Studies of comprehension suggest that but is acquired only later, but methodological issues in those studies may have underestimated children's capacity.

The present studies
In this paper, we provide a systematic experimental evaluation of children's ability to understand the contrastive discourse connective but, which we compare to their understanding of the causal connective so.We focus on the comparison of but with so because so (unlike and) generates unambiguous predictions about how to interpret discourse relations with expectations that are met, in contrast to but which is used for generating expectations that are violated or contrasted.To provide this, we examine how, and at what age, children use these connectives to inform their interpretation of novel words.This methodological choice builds on recent demonstrations that even young children can successfully use discourse context for word learning (Sullivan & Barner, 2016;Sullivan, Boucher, Kiefer, Williams, & Barner, 2019).For example, when they hear I'm very thirsty, but cold things hurt my teeth!Look at what I want!There's a gazz on the table, even two-year-olds are able to reliably identify the correct referent of the unfamiliar word when offered a choice between a cup of coffee, an ice cream, and a cup of ice water (Sullivan & Barner, 2016).Moreover, 4-to 5-year-old children show a modest, but statistically significant, ability to use the connectives and and because to infer different referents for unfamiliar words, indicating that children's word learning builds upon their discourse comprehension (Sullivan et al., 2019).We thus reasoned that this word-learning method would be more sensitive than previous paradigms used to investigate children's comprehension of but, such as pronoun interpretation (Spenader, 2018) or metalinguistic judgments (Cain & Nash, 2011).Moreover, our stimuli relied on simpler linguistic materials than the sentence-completion paradigm used by Kail and Weissenborn (1984).
In Study 1, participants engaged in a picture-selection task, in which they had to guess which of two pictures were referred to by a novel word, based on the context provided by a two-sentence description.These descriptions varied in the discourse relations and connectives that they contained.We compared children's interpretation of the novel words in three discourse contexts: one context without a connective (3a), one with the causal connective so (3b) and one with the contrastive connective but (3c After hearing one of these descriptions, participants chose which of two pictures the novel word might refer to (see Fig. 1A).In the example, ploam could refer to a depicted sledge or to a bicycle.In (3a), with no connective, we reasoned that participants would choose the sledge based on lexical associations.In (3b), the coherence relation so demands that the second clause provide a consequence of the first, and so participants should again choose the semantically associated picture (i.e., the sledge).Finally, the connective but in (3c) conveys that the meaning of the second clause contrasts with the first clause, and so participants should infer that ploam refers to the non-associated, contrastive picture (i.e., the bicycle).
Studies 1 and 2 assessed children's comprehension of these causal and contrastive coherence relations from 2 to 8 years of age.To foreshadow the results, we found that children struggled to interpret but as contrastive even at age 5.The subsequent studies (Studies 3-5) were designed to explore this difficulty, testing whether 5-year-olds, who are known to use the connective but in their conversations, might succeed under different conditions that could potentially facilitate comprehension of but and explain the source of children's difficulties.In Study 3, we consider whether particular aspects of the task and its design could limit children's performance, and show that modifications to the task do not relieve children's difficulty.In Study 4, we then consider whether limitations in children's sentence processing abilities can explain their difficulty with but, and show that they cannot.Finally, in Study 5, we consider whether children's difficulty with but reflects a lack of pragmatic support.We show that, under some conditions, increased pragmatic support does allow children to interpret but contrastivelybut only at the cost of their now failing to interpret the connective so as B. Skarabela et al. having a causal meaning.

Study 1
The aim of Study 1 was to assess how children in a wide age range (from 2 to 8 years) use discourse relations indicated by connectives.In a picture-selection task, participants were introduced to a novel word in one of three different contexts and had to choose which of two pictures the word referred to.The experimenter explained that they spoke a different language and would use unfamiliar words for some things.The participants were presented with novel words in three different contexts: (1) a control condition with no connective (e.g., Mary brought a ploam on a summer day), (2) a causal context indicated by the connective so (e.g., It was a summer day so Mary brought a ploam) and (3) a critical contrastive context indicated by the connective but (e.g., It was a summer day but Mary brought a ploam).We measured whether children chose a semantically associated picture (here, a bicycle) over a contrastive picture (a sledge).
Recall that prior work testing comprehension found that children did not treat but as contrastive until around 8 years.At issue was therefore the age in this sample at which children would start to select the contrastive picture for the but condition (e.g., a sledge in It was a summer day but Mary brought a ploam) at above-chance levels, a pattern of behaviour indicating that participants not only understood that the meaning of but was different from the meaning of so, but also understood that the meaning of but was specifically contrastive.
The study used a two-alternative forced choice picture selection task with a within-subjects design.Each participant received 24 sentences with eight items in each of the three conditions (i.e., no connective control, connective so, connective but) counter-balanced in a Latin square design.
2.1.1.2.Materials.The study used three practice items and 24 test items.The first practice item tested the participants' understanding of a familiar word embedded in a simple discourse (i.e.,Here are two different things.One of these are glasses!Can you point to the glasses?, with a picture of glasses and headphones).The second and third practice items tested their ability to choose a referent for an unfamiliar label from a foreign language (i.e.,Here are two different things.In my language we use a milgo when it's dark outside.Can you point to the milgo?, with a picture choice between a glass of water and a torch).All materials can be accessed via the OSF.
Sets of six test items were created as in (4a-f) below.Novel words were embedded in one of three types of context: a control single sentence context without a connective illustrated in (4a) and (4d), a two clause context conjoined with the connective so in (4b) and (4e), and a two clause context conjoined with contrastive but in (4c) and (4f).For each set, two variants of each context were created with opposing meanings to control for complexity of inference, and to ensure that each picture was the intended referent half of the time.
4a. Mary brought a ploam on a summer day.(control) 4b.It was a summer day so Mary brought a ploam.(so) 4c.It was a summer day but Mary brought a ploam.(but) 4d.Mary brought a ploam on a winter day.(control) 4e.It was a winter day so Mary brought a ploam.(so) 4f.It was a winter day but Mary brought a ploam.(but) Each practice and test item was presented with a pair of pictures (see Fig. 1A, a complete set of stimuli and materials can be found via the OSF).Practice trials included three pairs of images with familiar objects (e.g., glasses and headphones, cup and torch, sunglasses and bread).All test trials showed a pair of images with opposed meanings (e.g., a bicycle and a sledge for item set 1 above), and each image was paired with an animal as the named character (e.g., a penguin for item set 1, see Fig. 1A).One item, due to experimenter error, presented non-identical animal characters with each image.The Latin Square design generated six lists of 24 test items, with eight items in each condition per list.
2.1.1.3.Procedure.The study was administered by approaching participants and their caregivers in public places (e.g., parks).Caregivers provided written consent on the child's behalf, and child participants gave their assent before proceeding with the study.Each participant was randomly assigned to one of six lists.Participants were instructed that in this game they should point to the picture that best matches the description they hear.The experimenter then presented a 'picture book' with one experimental item per page, starting with the three practice items and followed by 24 test items.The experimenter recorded the participant's response after each trial.The study took 10 min for most participants, and no longer than 20 min for the youngest participants.The participants were thanked and debriefed at the end of the study.
2.1.1.4.Analysis.We analysed these data in terms of the proportion of trials on which participants chose the picture that was semantically associated with the context (see Fig. 1A for an example).We expected that adults would choose the semantically B. Skarabela et al. associated picture on so and no connective trials, but would choose the non-associated, contrastive picture on trials where but indicated contrast.
Our critical analysis examined when children showed the predicted adult-like patterns of choosing the associated pictures at abovechance levels on so and no-connective trials, and the non-associated pictures at above-chance levels on but trials.That pattern of behaviour would clearly indicate that children had learned that but is contrastive.In addition, we carried out direct comparisons between so and but trials in order to assess whether children treated the meaning of but as being different from the meaning of so.Note, however, that these direct comparisons do not establish whether participants understood but as contrastive: In a situation in which participants chose the associated picture on 90% of so trials but only 70% of but trials, then this would show that the participants treated the connectives as different, but it would not indicate that they specifically treated but as having a contrastive meaning.
The data generated in this study, and our subsequent studies, typically followed extremely clear patterns, such that differences across conditions could be appreciated by the naked eye.Standard inferential statistical tests for clustered binomial data, like mixed effects logistic regression models, sometimes failed here because participants performed at ceiling in at least some conditions, which leads to separation in the data and poor, biased model fits.Therefore, in order to compare performance in individual conditions against chance, we relied on bootstrapped 95% confidence intervals (see e.g., the error bars in Fig. 1B).We used mixed effects models only to compare across ages and conditions.
In contrast, performance on the but trials (e.g., It was a winter day but Mary brought a ploam) showed a different developmental pattern.Between 2 and 6 years, children were increasingly more likely to choose the semantically associated picture, even though but signifies contrast 3 (β = 0.24 (0.08), z = 3.2, p = .002).It was not until around 7 years that this developmental pattern reversed, and even at 7 years performance was at chance (M = 0.52, 95% CI [0.38, 0.69]), with only adults choosing the contrastive picture more often than chance (M = 0.24, 95% CI [0.10, 0.40]).The pattern of development, therefore, was decidedly not monotonic.
Finally, we compared the proportion of trials on which participants selected the associated picture for the so and but conditions.Combining the child participants aged 2 through 6, 4 we found that they were significantly less likely to choose the associate in the but condition (β = -0.7 (0.14), z = 5.1, p < .001).However, this effect of connective was almost an order of magnitude smaller than the effect of connective for the participants aged 7 and older (β = -5.9(0.56), z = 10.6, p < .001).

Discussion
Study 1 generated two key findings.The control condition and the connective so condition both show that even very young children will use discourse contexts to guess new associated word meanings, consistent with earlier studies of discourse bootstrapping (Sullivan & Barner, 2016;Sullivan et al., 2019).However, the second key finding highlights the limits of children's understanding of discourse: even at age 7, children were not reliably able to use the contrastive connective but to infer that a word may have a contrastive meaning.Indeed, children younger than 7 consistently chose the associated meaning in these contrastive contexts, at rates that were only a little bit lower than in the two non-contrastive conditions.Since children's difficulty with but was so unexpected and the developmental trajectory so striking, we aimed to re-assess it with a slightly modified design and stimuli.First, in order to address the possibility that children were primed to select the associated picture in the but examples because this was the correct response in two of the three conditions, we removed the control (no connective) condition, so that now participants would be expected to choose associated versus contrastive meanings on an equal number of trials.We also developed a new set of items in which the intended contrastive readings never risked violating real-world knowledge (i.e., some examples in Study 1 involved actions that were either very implausible or impossible, like a cat barking, e.g., Alex barks, but it is a noffee with cat as the target, or sea travel by train, e.g., Max travelled across the ocean but he took the tham with train as the target).We reasoned that this factor may have caused adults to only choose a relatively low proportion of non-associated choices in the contrastive condition, and we thus predicted that these modifications would facilitate contrastive inferences on but trials in both children and adults.Furthermore, we simplified the context of the stimuli presentation and removed the foreign language aspect of the study, simply introducing the picture-selection task as a guessing game that involved matching made-up words to one of the depicted single objects.

Study 2: Confirming children's difficulty with but
Study 2 provided a further test of whether children really do struggle with but, using a new set of stimuli and a new set of participants aged three-, five-and seven-years, each age group established as distinct in their interpretation of but by the previous study, as 1 For all regressions, we report model structure in footnotes using lme4 syntax.This model had structure glmer(ChooseAssociate ~ scale(Age) + (1|Participant)), applied to all children tested (and not adults).
2 glmer(ChooseAssociate ~ scale(Age) + (1|Participant)), applied to all children tested (and not adults).The first practice item tested the participants' understanding of a familiar word embedded in a simple discourse (i.e.,Here are two different things.One of these is a football!Can you point to the football?, with a picture of a football and a bicycle).The second and third practice items tested their ability to choose a referent for an unfamiliar label (e.g.,Here are two different things.We use a milgo when it's dark outside.Can you point to the milgo?, with a picture choice between a cup and a torch).Visual stimuli were coloured drawings of objects, without the paired animals used in Study 1.All materials can be accessed via the OSF.
The sets of 4 test sentences followed a similar form to the so and but conditions of Study 1, as in sentences (5a-d), presented with a pair of images.For sentences (5a-d), the images were of a winter coat and a t-shirt.5a.It was a cold day so Katy put on a pagle.5b.It was a cold day but Katy put on a pagle.5c.It was a warm day so Katy put on a pagle.5d.It was a warm day but Katy put on a pagle.
We used a Latin Square design to generate four lists of 20 test items, with 10 items in each condition per list.Each list was counterbalanced for the presentation side of the target image, creating eight lists in total (see the OSF for the complete set of stimuli and lists).Pictures were presented as PowerPoint slides on a laptop computer.The stimuli were presented by the experimenter who carefully articulated each connective with strong emphasis.
2.2.1.3.Procedure.The study was administered in a quiet room in local nurseries or the developmental lab at the University of Edinburgh.Participants were presented with a picture-selection task, introduced as a guessing game with words and pictures, and they were instructed to listen carefully to a short description and guess the meaning of a made-up word and match it to one of two pictures that best describes its meaning.Unlike in Study 1 the images were presented as slides on a laptop or tablet (see the OSF for an example with instructions).

Analysis.
As in Study 1 we analysed these data in terms of the proportion of trials on which participants chose the picture that was semantically associated with the context.We used bootstrapped confidence intervals to compare behaviour against chance, and mixed effects models to compare behaviour in this study with that found in Study 1.

Results
As Fig. 2 shows, Study 2 replicated the key findings of Study 1.On so trials, older participants were increasingly more likely to choose the associated picture.By contrast on but trials, 3-and 5-year-old children chose the associated picture, just as in the prior study (Age 3: M = 0.60, 95% CI [0.49, 0.71]; Age 5: M = 0.83, 95% CI [0.74, 0.91]).It was only at seven years that children began to reliably choose the non-associated picture (M = 0.23, 95% CI [0.09, 0.39]).
Notably, the new stimuli in Study 2 did seem to affect participants' inferences.Compared to Study 1, participants within each age group chose the associated picture at a higher rate in the so condition (although these consistent differences between Study 1 and 2 were not statistically significant).Moreover, in this study's but condition, 7-year-olds and adults were less likely to choose the associated picture compared to those in Study 15 (β = -1.8(0.75), z = 2.5, p = .01),suggesting that the new stimuli led them to generate more robust contrastive inferences.But importantly, with these new but stimuli, 5-year-olds actually chose the associated picture more often in Study 2 than in Study 16 (β = 0.98 (0.47), z = 2.1, p = .04).Thus, while these stimuli seemingly made it easier for older children and adults to draw contrastive inferences, they simultaneously appeared to make it harder for younger children to draw contrastive inferences.

Discussion
Studies 1 and 2 document and then replicate a striking limit to younger children's ability to comprehend discourse: a failure to understand that the connective but implies a contrast between two clauses, and to use that contrast for word learning.While children as young as 2-or 3-years could use discourse-supported associations to determine a word's meaning, it was not until age 7 that children were able to reliably override those associations when but conveyed a contrastive discourse relationship between two clauses.
Importantly, children's difficulty with but up to age 5 years was robust across both our initial stimuli and a second stimuli set that made it easier for adults and older children to infer contrast.The fact that the second set did not also make it easier for younger children to infer a contrastive relation, but instead appeared to make it harder, provides a strong cue that children of this age have a genuine difficulty interpreting but, at least in this experimental context.Younger children's apparent difficulty with but is both unsurprising and surprising.On the one hand, these data are unsurprising because they are consistent with prior reports that children have difficulty comprehending but (e.g., Spenader, 2018;Kail & Weissenborn, 1984).Even though our study used a simpler task than prior work, the only slight difference that we found compared to those reports is a slightly lower age at which but is first well-comprehendedwhere those studies suggested that children as old as 8 still struggle, we found robust success at age 7 in Study 2.
On the other hand, these data are surprising because it does not make intuitive sense that children as old as five years would struggle with but in this simple two-clause context with no other syntactic complexities or unusual scenarios with several charactersboth corpus studies and the lived experiences of parents suggest that children use but frequently and do not appear to often use it in unlicensed contexts.Indeed, caregivers who observed the study often expressed their disbelief about their child's behaviour.In our remaining studies, we thus tested a number of potential explanations for why children may struggle to understand but contrastively.

Part 2: Testing three factors behind five-year-olds' difficulties with but
We tested three potential explanations for why children under 7 years did not appear to interpret but contrastively in Studies 1 and 2. Study 3 examined explanations based on the design of our task, Study 4 examined explanations based on limits to children's language processing skills, and Study 5 examined explanations based on limits to children's ability to generate pragmatic inferences.Fig. 3 provides an overview of these three explanations and the tasks that we designed to evaluate them.
All of these studies focused on 5-year-olds, chosen because they had robustly misinterpreted but in both Studies 1 and 2 (i.e., consistently choosing the associated picture above chance) and because it was particularly surprising that children of this age would be unable to interpret but.Our running hypothesis was that these children do in fact possess a sophisticated understanding of the meaning of but, but that they failed to use it in our previous two experiments, perhaps because of task design or another aspect of their psycholinguistic competence, such as language processing skills or inferencing ability.However, as we show, that running hypothesis was not supported.

Study 3: Task design
While our task was relatively simple, it had features that could conceivably limit children's ability to draw contrastive inferences from the word but and use those inferences for word learning.We tested two that we considered most likely.First, it is possible that children were not expecting contrastive discourse in a task that required guessing the meaning of made-up words.Previous studies showed that a discourse bootstrapping paradigm is suitable for young participants and that even two-year-olds can infer meanings of unfamiliar words from the surrounding linguistic discourse (Sullivan & Barner, 2016).However, contrastive contexts are rarer than non-contrastive contexts, and require non-standard inferences on the part of the listener as to how discourse segments connect (see Sanders, 2005).Thus, it is possible that children struggled with our task simply because we posed them the unusual task of inferring new word meanings under contrastive contexts.Study 3A therefore tested whether evidence for contrastive interpretation of but might emerge in a task that does not build on unfamiliar words, specifically a sentence-completion task.A second potentially confounding aspect of our task is the relation between the linguistic and visual stimuli.It has previously been suggested that children may generate non-adult interpretations of a sentence if a) it requires reanalysis, and b) it is heard alongside a visual stimulus that is consistent with an initial (not reanalysed) interpretation.For instance, Hamburger and Crain (1984) found that children generate non-adult-like interpretations of phrases like the second green ball (where they interpret second to modify only ball, not green ball) if they can concurrently see a line of balls where the second-positioned one is green, but do not do so if the line of balls is only revealed after the phrase is heard.They argued that the concurrent visual stimulus caused children to initially fixate on the second-positioned ball (rather than the second of two green balls, see also Meroni & Crain, 2003), and they then struggled to move beyond this initial interpretation.For the present context, we reasoned that hearing the initial clause of each stimulus item (e.g., The day was cold…) might cause children to focus on the associated picture, either because they were implicitly predicting that it would be named (Mani & Huettig, 2012) or simply because of priming based on semantic associations (Jerger, Martin & Damian, 2002;Arias-Trejo & Plunkett, 2009).Children would then struggle to move beyond that associated interpretation on hearing the contrastive connective.This reasoning implies that children may provide better evidence for understanding contrastive but if they cannot see an associated picture while listening to a sentence.We tested this in Study 3B.

Study 3A: Sentence completion
Study 3A tested whether 5-year-old children would show more evidence for contrastive interpretation of but when the task design did not require guessing the meaning of an unfamiliar word.Instead, children heard stimuli as before but with the final unfamiliar word removed (e.g., It was a cold day so Katy put on a…) and then chose which of the two pictures completed the sentence, using either a verbal report or a point.Otherwise, the methods here were identical to Study 2 when children participated in a picture-matching task, except that only 16 5-year-olds (8 female) were assessed (the list of stimuli is available on the OSF).
However, this change to the method did not provide evidence that 5-year-olds would contrastively interpret but.As before, and as Fig. 4 shows, they chose the associated picture above chance on both so trials and on but trials (So: M = 0.86, 95% CI [0.73, 0.96], But: M = 0.77, 95% CI [0.59, 0.92]), indicating that they did not treat but as contrastive.Additionally, a comparison between so and but

Study 3B: Preventing prepotent response
Next, we assessed whether performance would change if 5-year-olds could not see the images while listening to the sentences.Here, we followed the method and protocol from Study 2, using the same verbal and visual stimuli, except that only 16 5-year-olds (8 female) participated and that children were presented with a blank screen while listening to the sentences, and were only shown the pair of visual stimuli once they had listened to the entire stimulus sentence.
Again, however, this change to the procedure did not affect children's behaviour.As before, they robustly chose the associated picture on both so and but trials (Fig. 4).When children had to process the sentence before seeing the images, they still chose the associated meanings at above chance rates for both connectives (So: M = 0.95, 95% CI [0.82, 1], But: M = 0.72, 95% CI [0.60, 0.85]), indicating that participants did not treat but as contrastive.In addition, a comparison between conditions indicated that participants were somewhat less likely to choose the associate on but trials (β = -7.4(3.5), z = 2.1, p = .03).9

Study 4: Language processing
We next tested whether children's difficulty with but might be driven by limitations in their language processing ability.Again, we considered two possibilities.First, that difficulty with but may result from a general difficulty with reanalysis.There have been a number of demonstrations that children struggle with reanalysis during sentence processing (e.g., Trueswell, Sekerina, Hill, & Logrip, 1999;Huang et al., 2013; and see Study 3B above), and interpreting but contrastively in It was a cold day but Katy put on a pagle would, to some degree, require the participant to revise their expectations about what is likely to be said following It was a cold day….To test this possibility, in Study 4A we examined how 5-year-olds generated contrastive inferences when the contrast was indicated at the start of the sentence, rather than after the first clause.
To do this, we needed to use a connective other than but, because but cannot be used sentence-initially to introduce a conflict or contrast between two subsequent propositions.English has a small set of items that allow this, including even though, although, despite/ in spite of the fact, although we are not aware of any studies that examine their acquisition in the preschool years.These items are of course much more rare in parent-child conversations than the highly-frequent connectives but and so, yet are still also clearly part of children's input.For example, an analysis of the Gelman corpora indicates that children hear even though approximately as often as they hear the verb kick (Gelman et al., 1998;Gelman, Taylor & Nguyen, 2004;Gelman, Ware, Kleinberg, Manczak, & Stilwell, 2014;Jipson, Gülgöz & Gelman, 2016).In addition, corpus-based English language descriptions argue that even though is more colloquial than the other connectives and makes a stronger contrast than although (Carter et al., 2011;Dušková et al., 1988).For these reasons, even though was selected as the best candidate for the study (e.g.,Even though it was a cold day, Katy put on a pagle).If difficulty with but is actually a difficulty with revision, then we expected that children should be more capable of inferring contrastive meanings when even though is used, as it allows them to infer contrast before they build up expectations about the rest of the sentence.
Second, we considered the possibility that failure to understand but could be a consequence of so-called noisy channel processing (Gibson, Bergen & Piantadosi, 2013;Yurovsky, Case & Frank, 2017;de Carvalho, Dautriche, Fievet, & Christophe, 2021), in which children interpret sentences not just on the basis of what they hear, but on the basis of what they expect people to communicate as well.If children do not expect speakers to communicate contrastive meanings, especially meanings that describe less conventional scenarios, then they may ignore evidence for those meanings, and interpret a sentence as having contained a different connective besides but.To test this, Study 4B examines whether children would generate more contrastive interpretations of but if they were first provided with greater evidence that the speaker intended to communicate a contrastive meaning.Specifically, children received feedback in the first ten trials offering evidence that but was being used contrastively, and so was being used non-contrastively.

Study 4A: Reanalysis and even though
Study 4A tested whether children would show stronger evidence for contrastive interpretations when but was substituted with the connective even though.The method follows Study 2 when children had to guess the meaning of an unfamiliar word, but again only using 16 5-year-olds (8 female), and replacing sentences using but with sentences beginning with the contrastive connective even though (e.g.,Even though it was a cold day, Katy put on a pagle) (see the OSF for verbal stimuli).
Again, however, children did not show strong evidence of generating contrastive interpretations (Fig. 4).In both the so and even though conditions, they consistently chose the associated picture (So: M = 0.95, 95% CI [0.91, 0.98]; Even though: M = 0.70, 95% CI [0.54, 0.83]).This suggests that children's difficulty with but is not specific to reanalysis and extends to multiple connectives that imply contrastive discourse relations.Additionally, a comparison between conditions indicated that participants were marginally less likely to choose the associated picture on even though trials than so trials (β = -1.6 (0.8), z = 1.9, p = .052).10

Study 4B: Feedback
Study 4B assessed whether increasing children's beliefs that the speaker will use contrast would then lead to more contrastive, rather than associated, interpretations.16 5-year-olds (8 female) took part in both a feedback and a test phase of the study.In the first part of the experiment (Trials 1-10), children participated in the Feedback phase when they were invited to complete sentences in a similar manner to Study 3A, either by saying a familiar word or by pointing at one of the pictures.During this phase, they received feedback on their choices (e.g., if they chose the associated picture on a but trial or non-associated picture on a so trial, then the experimenter would say 'not quite right' and invite the child to listen carefully as they provided the correct target form embedded in the stimulus sentence as they pointed to the target image; no further explanation was provided).At the end of this phase, children were told that the game was changing and they now had to guess the meaning of a made-up word given.("Now, we will change the game.I will finish the sentence with a funny word.Listen carefully and then guess what the word means and point to the picture that best describes the new word.Ready?").In the subsequent Test phase (Trials 11 through 20), no feedback was provided.
In the Feedback phase, children chose the associated meaning on the large majority of so trials (M = 0.94, 95% CI [0.83, 1]) and also on the majority of but trials (M = 0.8, 95% CI [0.69, 0.91]), meaning that children rarely received feedback on so trials but received feedback on the majority of but trials.Importantly, in the critical Test phase, we found that this feedback had not caused children to interpret but contrastively.Again, on the majority of trials, children chose the associated meaning even when the connective but was used (Fig. 4), although the 95% confidence interval now overlapped chance (So: M = 0.84, 95% CI [0.71, 0.94], But: M = 0.65, 95% CI [0.48, 0.84]).Moreover, in this study, a comparison between so and but trials during the Test phase indicated no significant difference (β = -0.87(0.9), z = 0.95, p = .34). 11inally, we ran an analysis comparing children's responses during the Feedback and Test phases, examining whether feedback caused children to interpret but, but not so, more contrastivelya result that one might expect given that children were often given feedback that but trials were contrastive but never given feedback that so trials were contrastive.Overall, children were more likely to choose the associate on so rather than but trials (β = 0.85 (0.21), z = 4.1, p < .001).12However, the effect of the feedback was not particular to but trials.Across the two connectives, children were less likely to choose the associate in the Test phase rather than the Feedback phase (β = -0.64(0.2), z = 3.2, p = .002),and the interaction term showed that the effect of phase did not differ between the so and but trials (β = -0.09(0.19), z = 0.47, p = .64).Interestingly, an analysis of the individual datapoints suggested that these results were mostly driven by a small number of children (3/16) for whom the training led to consistently generating contrastive meanings for both but and so, rather than all children showing a small effect of feedback.
The major implication of Study 4B is that children's difficulty with but is unlikely to be due to contrastive meanings being rare and thus downweighted as a possible interpretation.When given explicit feedback that the speaker wanted them to interpret but trials contrastively, children still did not robustly interpret subsequent but trials contrastively.Therefore, it does not appear that noisy channel principles can explain children's unexpected behaviour here.
Moreover, a striking additional finding is that the feedback actually caused some children to start treating both but and so contrastively.Even though children were never given feedback that so should be interpreted contrastively, a small number of children inferred that feedback on but trials should cause them to interpret both but and so sentences contrastively.This pattern of data is quite interesting because it is consistent with the possibility that children do not understand the distinct discourse functions of both connectives, but as well as so.It thus raises the question of whether 5-year-olds actually understand the meaning of so, despite their accuracy with it on our task.We return to this in Study 5B.

Study 5: Pragmatic inferencing
We next considered the possibility that children's difficulty with but may be driven by difficulties in the pragmatics of deriving contrastive readings.Specifically, in order to interpret but contrastively, children need to understand what information is being contrasted, and inferring this information may be either difficult for them (cf. the analysis of children's scalar implicature difficulties in Barner, Brooks & Bale, 2011), or it may be contextually underdetermined in the verbal stimuli used so far.As an example, consider the item It was a cold day but Katy put on a pagle.To get a contrastive interpretation in which pagle means t-shirt, the child needs to first infer that on a cold day we wear heavier clothing like a winter coat, and then contrast that inference.Thus, children may struggle to generate the contrastive interpretation because they either do not generate an inference with which to create a contrast, or they generate an inference that, when contrasted, does not lead to the interpretation that pagle refers to a t-shirt (e.g., generating the inference that when it is a cold day one is likely to get chilly, and so it is contrastive for a character to be able to overcome such a problem by wearing a warm clothes like a coat that will stop them from getting chilly).
We approached this possibility in two ways.In Study 5A we specified the relevant inference that the participants could then contrast (e.g., The day was cold.When the day is cold, you should put your winter clothes on, but Katy put on a pagle).In Study 5B we went beyond this by specifying not only the inference, but a specific lexical item that the participants could contrast (e.g., The day was cold.When the day is cold, we wear a coat, but Katy put on a …).In both of these studies, the key logic was the same: if difficulty with but is due to a failure to generate the right inferences, then children's interpretation of but contrastively should be facilitated by making the inference or lexical item to be contrasted explicit.

Study 5A: Explicit inference
Study 5A followed the same protocol from Study 2 with 16 5-year-olds (8 female) who guessed the meaning of an unfamiliar word, using the same visual stimuli, while the verbal stimuli from Study 2 were modified to make the relevant inference explicit (see 6a-d) (stimuli available via the OSF).If children's difficulty lies in determining the inference to contrast, then we expect them to treat but as contrastive under these conditions.6a.The day was cold.When the day is cold, you should put your winter clothes on, so Katy put on a pagle.6b.The day was cold.When the day is cold, you should put your winter clothes on, but Katy put on a pagle.6c.The day was hot.When the day is hot, you should put your summer clothes on, so Katy put on a pagle.6d.The day was hot.When the day is hot, you should put your summer clothes on, but Katy put on a pagle.However, even when the inference was made explicit, we still found no evidence that 5-year-olds would interpret but contrastively (Fig. 4).Again, participants were more likely to choose the associated picture in both the so condition and in the but condition (So: M = 0.91, 95% CI [0.84, 0.98], But: M = 0.74, 95% CI [0.53, 0.91]).Thus, making an inference explicitat least at this level of granularitydid not cause children to reliably treat but as contrastive.Moreover, in this study, the comparison between so and but trials indicated no significant difference (β = -0.5 (1.5), z = 0.3, p = .75).13

Study 5B: Lexical contrast
Finally, we asked whether 5-year-olds would interpret but contrastively if we provided a precise lexical item to contrast against.The study followed the protocol from Study 2 with two modifications.First, we modified the verbal stimuli to specify the relevant inference that participants needed to draw and also, importantly, mention a specific lexical item that would be part of that inference.As a result, the modified sentences referred directly to one of the pictures (see 7 below).Second, we used a sentence completion task rather than a task that requires guessing unfamiliar labels.This was to avoid a confound in which children might use mutual exclusivity to infer that the unfamiliar label could not refer to the picture that had just been previously mentioned.
7. The day was cold.When the day is cold, we wear a coat, so Katy put on… An important aspect of the new contexts in Experiment 5B's stimuli is thatunlike all the previous studiesthey intuitively provide a better fit for a contrastive continuation than a non-contrastive one, as the latter would involve the repetition of the object label.For example, it is more marked to say When the day is cold, we wear a coat, so Katy put on a coat than to say When the day is cold, we wear a coat, but Katy put on a t-shirt.This raises the intriguing possibility that, if five-year-olds have only fragmentary knowledge of the two connectives but and so, we might expect globally more contrastive interpretations in this study for both connectives as a result of children's sensitivity to discourse features and strategies related to referential communication, including the expectation that objects of transitive clauses tend to introduce new referents (see Allen, Hughes & Skarabela, 2015;Du Bois, 1987, 2003).The possibility that children may have fragmentary knowledge of so was raised in Study 4B that showed that some children started giving contrastive readings to both so and but in response to feedback that their but responses were incorrect.Moreover, while children across our studies were highly accurate in the so conditions, it is not clear that so was actively driving their interpretation, as they behaved very similarly in the control (no connective) condition of Study 1.
Because of the coronavirus pandemic, data for this study was collected online and participants were assessed on a video call.We used a sentence completion task, as in Study 3A, with the same visual stimuli as Study 2 delivered as a slide presentation on a shared screen, and the linguistic stimuli spoken by the experimenter.The two pictures appeared on different colour backgrounds (blue or green).Participants typically completed the sentences by naming one of the two pictures, and some also named the colour of the background the picture appeared on, or pointed to the picture.Fifteen 5-year-olds participated (7 female), along with one 6-year-old (female, 77 months).
Unlike the prior seven studies, the 5-year-olds in Study 5B did not show a robust tendency to choose the associated meaning on but trials.Instead, as Fig. 4 shows, they were now more likely to choose the non-associated, contrastive meaning (But: M = 0.30, 95% CI [0.12, 0.51]).While the confidence interval around this proportion just overlapped chance, the rate of choosing the associated picture was significantly lower than in Study 5A14 (β= 4.16 (1.5), z = 2.7, p = .007).However, this apparent new success on but trials was complemented by a striking failure on so trials.Unlike all our prior studies, the 5-year-olds were now at chance on whether they chose the associated or non-associated meaning (So: M = 0.48, 95% CI [0.28, 0.7]), and the comparison between so and but trials indicated only a marginal difference (β = -2.6 (1.5), z = 1.7, p = .09). 15 Thus, in this context where children apparently could derive a contrastive meaning for but, they now no longer derived a reliably non-contrastive meaning for so.
To understand these importantly distinct data better, we broke the group-level results down at an individual child level (Fig. 5).Eight children clearly interpreted both but and so contrastively, four clearly interpreted both but and so non-contrastively (as in our prior studies), two children showed an adult-like pattern of interpretation, and the remaining children were either at chance for but or at chance for so.Thus, while children as a group were at chance when interpreting so, few individuals showed this result.Instead, most children either interpreted so contrastively or interpreted so non-contrastively.Importantly, the individual-level data provides little evidence that children distinguished the meanings of but and so in these contexts.If children interpreted but contrastively, then they tended to also interpret so contrastively, and if they interpreted but non-contrastively, then they interpreted so non-contrastively.

Exploring individual differences in children's understanding of but
Finally, building on the individual child analysis from Study 5B, we examined the degree to which children across our studies showed variation in their understanding of but and so.We focused on Studies 3A through 5A, not incorporating Study 5B, where children's responses importantly differed.Fig. 6A shows individual childrens' performance across these tasks.The data are notable because they suggest two clusters.The large majority of children chose associated meanings at roughly similar rates in both the so and but conditions, but a significant minority of children, about 25%, showed an adult-like pattern of choosing the associated meaning for so and the non-associated, contrastive meaning for but. 16Thus, some of the children in this study did in fact interpret but contrastively.
We then conducted a set of regression analyses to test what could predict whether children would interpret but contrastively.Initial comparisons suggested that children's behaviour was not predicted by their gender.However, a regression analysis suggested that children who showed the more adult-like pattern tended to be older.We regressed proportion choosing the associate in the dataset above against the connective use and age, and found that the two interacted so that the effect of connective was larger in older 5-yearolds (interaction between connective and age, β = -0.66(0.28), z = -2.4,p = .019,see Fig. 5B). 17Thus, it seems that while most 5-yearolds in our sample did not understand the contrastive meaning of but, a small number of children, typically the older children in our sample, had already learned it.

Discussion
Studies 3 through 5 were designed to examine a number of factors that could explain away children's misinterpretation of contrastive but that we documented in Studies 1 and 2. Focusing on five-year-olds, we tested the contributions of task design, language processing, and pragmatic inferencing.The results provided striking evidence for a genuine difficulty interpreting but.Despite our manipulations, children in five out of six of the studies continued to interpret but non-contrastively, and in the one study where children did succeed in interpreting but contrastively, they simultaneously failed to reliably interpret the connective so noncontrastively.
Specifically, children failed to interpret but contrastively in both manipulations of Study 3, where we removed the emphasis on guessing the meaning of unfamiliar words (Study 3A), and where we made it harder for children to generate a prepotent response by removing the concurrent visual stimuli (Study 3B, cf.Hamburger & Crain, 1984).Children also failed to recover a contrastive interpretation in both manipulations of Study 4. First, in Study 4A, where we minimised the need to perform semantic reanalysis by fact three clusters in our data.The Elbow plot can be found on the OSF repository for this paper. 17glmer(ChooseAssociated ~ scale(Age) * Connective + (1|Participant)), data from studies 3A through 5A.using the sentence-initial connective even though, suggesting that the difficulty with but is not a difficulty with revision and reanalysis (cf.Trueswell et al., 1999).Second, in Study 4B, where we gave children explicit feedback that the experimenter was using but contrastively, suggesting that difficulty with but is not because of a prior about which meanings are likely (cf.de Carvalho et al., 2021).Finally, children's difficulty with but could not be explained by difficulty generating the pragmatic inferences necessary to understand the intended contrast (cf.Barner et al., 2011).When the relevant inferences were explicitly specified, children still failed to interpret but contrastively (Study 5A).
Only in Study 5B did children show evidence of interpreting but contrastively.In that case, the stimuli not only made the relevant inferences explicit, but provided a lexical item to contrast against (e.g., The day was cold.When the day is cold we wear a coat, but Katy put on…).Unlike in our prior studies, children now chose the non-associated picture for sentences using the connective but, indicating that  they had derived a contrastive interpretation.However, also unlike the prior studies, children now no longer chose the associated picture for sentences using so, indicating that they did not derive a causal interpretation for so.This pattern suggests that in Study 5B children were generating lexically driven contrastive interpretations from the global context of the stimuli sentences, rather than because they understood that but was being used contrastively.Consistent with this, an individual-level analysis indicated that the majority of children in this study did not interpret sentences containing but and so in different ways: they either interpreted both of the connectives contrastively or interpreted neither connective contrastively.Moreover, an individual-level analysis of Studies 3A to 5A also showed that most children did not distinguish the meanings of so and but, and that the relatively small proportion of children who did distinguish their meanings tended to be older.
Thus, none of our follow-up studies provided evidence that 5-year-olds, as a group, understand that the connective but highlights a contrastive discourse relation, and some of our data even suggest that children of this age do not know that the connective so highlights a causal discourse relation.

General discussion
How do children learn the meanings and discourse relations that are expressed by connectives like but? Prior corpus analyses suggested that children use but in a sophisticated way by their third birthday, but prior experimental work had suggested that children's comprehension may be far more fragile, perhaps not emerging until the school years.Our eight experiments evaluated this claim.Using a discourse bootstrapping paradigm, we tested whether children could use the connectives but and so to infer different discourse relations (contrast versus causality, respectively) and inform their interpretation of an unfamiliar label or production of a familiar word.The findings provided a striking confirmation of the prior experimental work.While even 2-and 3-year-old children selected meanings in keeping with a causal interpretation of so (Studies 1 and 2), by choosing associated meanings for novel words, we did not find evidence that children reliably interpreted but contrastivelyby choosing contrastive, non-associated meanings for novel wordsuntil age 7. Instead, younger children reliably chose associated meanings when the connective but was used (albeit at a lower rate than for so) indicating that they had interpreted but non-contrastively.
Studies 3 through 5 confirmed and clarified this finding.Five-year-olds still did not reliably interpret but contrastively when: 1) The task was simplified and did not involve guessing unfamiliar labels (Study 3A), 2) When participants could not see the response pictures while listening to the sentences, thus preventing premature associations between the sentences and the pictures (Study 3B), 3) When the need for semantic reanalysis was removed, by indicating that the sentence would be contrastive from the start by placing the connective even though sentence-initially (Study 4A), 4) When children were explicitly told that their associative responses on but trials were incorrect (Study 4B), and 5) When children were explicitly provided with the inference that but was supposed to contrast against (Study 5A).
The only occasion on which children consistently interpreted but contrastively was in Study 5B, when the context explicitly provided a lexical item that but could potentially contrast against.However, these contexts also made non-contrastive continuations pragmatically more marked (Givón, 1983;Du Bois, 1987, 2003), and indeed children in Study 5B also often interpreted the connective so as having a contrastive meaning.Given this, we suggest that children's contrastive interpretations in that study were mainly driven by the particular context used and, at least for some of the children, by their sensitivity to referential strategies related to contrast and newness (Allen et al., 2015), rather than by the meaning conveyed by the connective but itself.Moreover, the pattern of data in this study raise the possibility that, like with but, children may actually not fully understand the discourse contribution made by the connective so: with these stimuli, many participants interpreted sentences with so to be contrastive rather than causal, and this thus raises the possibility that, for many children, their accuracy with so trials in our prior studies was also driven mainly by the contexts that they were hearing, rather than by the presence of that connective.
Thus, our eight studies provided evidence that, as they enter formal education, children still have only a limited understanding of the meanings of discourse connectives, particularly the contrastive meaning of the connective but.In this Discussion we evaluate the strength of that evidence and whether alternative accounts might explain our findings, discuss why children may be able to fluently use these words without fully understanding them, and suggest why children may have difficulty learning these words, and how they eventually do come to understand them.

Do 5-year-olds really not understand the meaning of but?
The claim that children struggle to acquire the meaning of but is striking.After all, even 3-year-olds use but frequently, and in a variety of licensed ways, including to mark protest and contrast (e.g., Bloom et al., 1980;French, 1981;Gallivan, 1986).And as mentioned before, caregivers who observed our studies frequently expressed their disbelief that their child had been unable to reliably infer contrast for sentences containing but.Given this, it is worth considering the possibility that our studies misestimate children's competence with connectives.
One possibility is that our data patterns are unreliable, but this does not appear likely.Instead, across our eight studies, we observed a remarkable consistency in how children interpreted the connectives but and so, across a range of experimental features.Only in Study 5B was the pattern of responses different.Thus, children's misinterpretation of but in these contexts appears to be a reliable developmental pattern.
A second possibility is that our tasks underestimated children's competence with but, and that a different methodological approach would elicit more adult-like behaviour.Our experiments aimed to reveal an adult-like competence through a range of different means, including different sets of stimuli, different types of response measures (sentence completion, word learning), explicit interventions such as feedback, and a variety of contexts designed to make inferences about contrast easier.Few appeared to make a significant difference to children's behaviour.For example, one suggestion is that children were simply making their choices based on associations between the pictures and the words that they were hearing, rather than processing the sentences more deeply for their discourse structure.However, several of our findings show that this is not likely.When children could not see the pictures while hearing the sentences (and thus could not generate associations online, Study 3B) they still tended to choose the associated picture on but trials.Moreover, in Study 5B, when associations between the pictures and the sentences were maximised, with one picture even being directly mentioned, children now did not choose the associated image.Thus, associations do not seem a parsimonious explanation of children's choices.
Another potential methodological explanation of children's difficulties, based on a common feature across our studies, is the use of a within-subjects design to assess comprehension of both but and so.Within-subjects designs are typically preferred as they reduce measurement error.However, as suggested by a Reviewer, there are also cases where within-subjects designs can increase a chance of a spillover effect, in which responses in one condition may impact responses in the other (e.g., Snedeker & Yuan, 2008).It could therefore be that a between-subjects design in which children only respond to contrastive but trials, rather than but trials mixed with so trials, would better highlight children's competence with the connective.This needs explicit testing in the future.However, a reanalysis of our current data provides some potential initial evidence against the idea that children's interpretation of but in our studies was influenced by exposure to so examples.In particular, a spillover explanation predicts that children should become less able to interpret but as the study proceeds, which was not the case in our dataset. 18Moreover, spillover also predicts that when children answer but trials before hearing any so trials, then they should show contrastive behaviour, but that was again not the case in our dataset.When children's very first trial was a but trial, 19 they still chose the associated meaning on more than half of the trials (M = 0.58, 95% CI [0.43,0.73]).Thus, we are confident that our methods did not radically underestimate children's competence, and we suggest that our variety of studies provide a reliable and broad test of children's lack of ability to infer contrastive discourse relations in these contexts.
Still, while our data generally make the case that children, especially 5-year-olds, have a surprisingly limited understanding of but, there are also some qualifications to that conclusion.For example, our individual child analysis of Studies 3A through 5A (Section 3.4) indicated that, while most of our participants did not distinguish the meanings of so and but, a proportion (about 25% of participants) were able to do so, choosing the associated picture on so trials and the non-associated picture on but trials.Importantly, those 5-yearolds who did interpret but contrastively also tended to be older, a point that we return to below, but which indicates that children's understanding of but has begun to emerge at this age.Further support for this comes from the group-level performance of 5-year-olds who were less likely to choose associated meanings on but trials than on so trials.Thus, our data provide some evidence that an understanding of but is present amongst some children at age 5, even if most children failed to understand it.
Finally, an important further limitation to our data is that it did not provide a detailed characterisation of all of the different senses of meaning that connectives such as but or so can take.Namely, the focus in our study was on but's use to indicate contrast through a violation of expectations, a sense that is often discussed in theories of discourse structure (e.g., Sanders, Spooren, & Noordman, 1992) and that has often been studied in prior work with children (Kail & Weissenborn, 1984;Champaud & Bassano, 1994;Florit et al., 2017).However but is also used with a variety of other senses of meaning, such as marking protest (e.g., but I don't want to) or lexical contrast (e.g., this dog is big but that dog is little, and see also the stimuli in Experiment 5B).It may be that while children struggle with the violation of expectations sense of but, they may be able to interpret but contrastively when it is used in alternative contexts where its contrastive meaning can be supported by other aspects of discourse, sentence structure or lexical choices.That possibility remains to be tested, and is likely to be an important component for building strong theories of how children acquire the semantics of terms like but.
To summarise, given all the considerations above, we suggest that the most parsimonious interpretation of our experiments is that, even at age 5, most (but not all) children fail to understand the contrastive meaning of but, at least when it is used to indicate a violation of expectations.We do not think it is likely that our experiments provide an unreliable measure of children's competence, nor that our experiments are an invalid measure of children's competence.Detailed examination of our data suggest that a small percentage of children have acquired but's contrastive meaning by age 5, however 5-year-olds as a group did not provide good evidence that they understand but to be contrastive.

Production without comprehension
If most children cannot understand but by 5 years, then why do studies of production suggest that even younger children use it in competent and sophisticated ways?We suggest two reasons for this, one uncontroversial and one potentially controversial.The first, and uncontroversial, possibility is that children may often use but in the context of common multiword constructions (e.g., but I don't want to), without a full understanding of all individual lexical items (see Morris, 2008 for evidence of item-based learning of the connectives and and or).This is a fairly-well established phenomenon (e.g., Hurewitz, Papafragou, Gleitman & Gelman, 2006 for similar asymmetry in quantifiers like some), and could give the false impression that children know a wide range of but's senses.
The second, and more controversial, possibility is that children's competence may have been overestimated because but's meaning 18 Specifically, for the but data from 5-year-olds in Study 2 and from Studies 3A through 5A, we conducted the regression glmer(SelectAssoc ~ scale(ItemOrder) + (1|Participant)), both for the full dataset and for the first five trials and the first ten trials.In no case was there a significant effect of item order. 19With data taken from the same set described in Footnote 18.
B. Skarabela et al. is not only subtle but also interpretable in many different contexts.This means that even if a child is using but erroneously, then listeners (including researchers) may not recognise that the child has committed an error, because they are flexible and accommodating in the way that they interpret the child's use of but.As a result, children may use but in ways that adults take to be competent, without the child fully understanding but's contrastive meaning.To illustrate, consider this example of how one 5-year-old used but in the Gelman corpus (Gelman, Taylor & Nguyen, 2004): Mother: What's this?Child: Yeah, it's inside but I don't know what it is.
Mother: Well, it's kinda hard to tell.
The child appears to competently use but to express contrast.But consider the possibility that this child actually misunderstands but to be synonymous with so, as with our experimental data.In this scenario, the child is expressing a meaning that is not adult-like, but that is also coherent in this context (equivalent to Yeah, it's inside so I don't know what it is).However, that non-adult-like meaning will not be at all obvious to the mother listening to the child, or to an annotator reading the transcript, who both would instead interpret the child as producing a sentence with a contrastive discourse relation.The key point is that a child can misunderstand the meaning of the word like but, and use it to express a non-contrastive discourse relation, yet because but can be interpreted in many different ways, then an adult hearing (or reading transcripts of) these uses may still assume the child's use to be correct and contrastive.
These two points in combination mean that there is plausible reason to believe that prior work on production may have overestimated children's knowledge of but.First, children may often use but in the context of frequent sentence frames, without understanding the meaning.Second, even if children do use but with an incorrect meaning, then this misuse may be difficult for an adult to detect (see also Clark, 1971 on the difficulty of detecting these types of errors in children's speech).

How do children acquire discourse connectives?
Finally, we turn to the major open question: if children do not understand the meaning of but at age 5, how do they learn it by age 7? Or put another way, if children can understand but by age 7, then why are they not able to understand it at age 5?
We suggest the following hypothesis: the full range of adult meanings of words like but are extremely hard to acquire by attending to informal conversational discourse alone.That is partly because those meanings are subtle (and, as suggested above, their misuse is difficult to detect in conversations), and partly because the distribution of meanings used in informal conversations may differ from the distribution used in formal discourse and writing (e.g., Chafe & Danielewicz, 1987;Taboada & Gómez-González, 2012).As a result, most children have not acquired the meaning of but as an explicit marker of contrast at the point at which they enter formal education.However, once children enter formal education and begin to systematically engage with written language and formal discourse, their understanding of these terms becomes explicit and transformed.
This hypothesis breaks down into three components: a claim that the adult meaning of but is hard to acquire from informal conversations alone, a claim that the wider range of adult meanings of but is learned through literacy and exposure to written language, and a claim that the major effect of literacy on learning begins when most children start school.We discuss these claims in reverse order.
The data in Fig. 6B provide evidence that children's understanding of words like but transforms suddenly, and so offers circumstantial evidence that formal educationwith new systematic exposure to print and formal discourse as well as instructioncontributes to this transformation.Specifically, Fig. 6B shows 5-year-olds' performance on but and so trials in Experiments 3A through 5A, plotted by age in months, along with logistic best fit lines for that data.The best fit lines show gradually improving performance on but and so within the age range of 60 to 72 months; as argued earlier, we suggest that this is because an increasing proportion of children have learned the meaning of but, rather than all children getting slightly better.However, when the best fit lines are extrapolated to older ages, we see thatfor butthe improvements between ages 5 and 6 do not actually predict behaviour at age 7 (while performance at so is at ceiling).Instead, a much higher proportion of children can understand but by age 7 than we would expect given the observed improvement from ages 5 to 6 (see Janssens, Drooghmans, & Schaeken, 2015 for further evidence that age is not a good predictor of children's interpretation of but).One way to interpret this non-linearity is the potential addition of a new factor that causes children to rapidly develop their knowledge of connectives between ages 5 and 7. Since the major environmental change for children who participated in our studies during this age period is the beginning of formal education, and with it an introduction to reading and writing, we suggest that it is these activities that provide a key trigger for children's learning of the wider meaning of connectives.
Is it plausible that literacy and schooling would change children's understanding of connectives?Yes.First, literacy opens routes to wider discourse contexts.The language of schooling, formal discourse, and literacy activities, including story reading, differs from informal conversations (Chafe & Danielewicz, 1987;Halliday, 1994;Schleppegrell, 2001;Crain-Thoreson et al., 2001).Moreover, the way that but is used in text and in school may be quite different from the way that but is used in conversations with caregivers.These educational contexts, therefore, are likely to push children to engage with the meaning of but in quite a different way.Second, exposure to text offers learning opportunities that are not available in informal conversations.Readers, but not listeners, can focus on individual words, including terms like but, and explore and discuss their functions and meanings.Third, elementary and primary school curricula specifically focus on building children's knowledge of words like but through explicit instruction and practice.For example, the National Curriculum for England mandates that children be able to link sentences using conjunctions such as but during Year 3 (ages 7 to 8) (English programmes of study: key stages 1 and 2, 2013), while Scotland's Curriculum for Excellence mandates that children be taught to write sentences linked by connectives such as but during the first level of primary school (ages 6 to 9) (Education Scotland Benchmarks: Literacy and English, 2017), a stage at which our participants begin to interpret but contrastively.In textbooks for this age range, knowledge of connectives is built through exercises in which children choose the right connectives for different situations, or offer completions for partial sentences that finish with different connectives.These exercises appear to be effective at changing children's understanding of but: a 15-minute formal instruction, for only seven days, raised children's understanding of connectives (Soltis & Pflaum, 1979), both in terms of metalinguistic awareness and reading ability.Thus, there is ample reason to believe that even a short formal instruction and targeted practice could trigger children's attention to these terms and affect how they understand connectives such as but.Once children become aware of their function, they can begin to detect the terms in a wider range of settings and meanings, some of which may be restricted to the formal register and school settings (e.g, Schleppegrell, 2001).
This relates to the third claim of our hypothesis, that the contrastive meaning of but is difficult to acquire from spoken input alone.First, but's contrastive meaning may be hard to acquire because it does not predict strong differences in word use.For example, a parent may say They can see us through that mirror but we can't see them, an example found in the Gelman corpus (Gelman, Ware, Kleinberg, Manczak, & Stilwell, 2014).However, it would not be anomalous to say They can see us through that mirror but we can also see them (i.e., the antithesis).But signals contrast in both examples, yet they are semantically direct opposites.If but's meaning does not restrict word use, as these examples illustrate, it may be hard for a child to learn.
Second, as suggested above, but's usage may be systematically different in informal and formal discourse.For example, an informal analysis of the conversational usage of but in the CHILDES database (MacWhinney, 2000) indicates that it very often occurs turninitially.This type of usage may be far more rare in children's (and adults') literary material.Thus, the relative frequency of the different senses of but that children are exposed to in their conversational inputsuch as but as a marker of protest, or as an interjection may be quite different from the relative frequency of senses that they are exposed to in written stories and school settings (see Crain-Thoreson et al., 2001 for style differences between conversations and during story reading).This, along with the ambiguity of but's meaning discussed in the previous paragraphs, may make it particularly hard for children to learn its contrastive usage from conversational input alone.For the minority of children who have learned the contrastive meaning of but by age 5 (see above), we hypothesise that they may have received a particularly rich and advanced literary input, presumably from being read by their caregivers (e.g., Volodina & Weinert, 2020;de Ruiter, Lieven, Brandt & Theakston, 2020).
To summarise, we propose that children have not mastered the contrastive meaning of but before school, both because their spoken input provides only ambiguous evidence for that meaning, and because the input may be dominated by other senses of but, such as protest or opposition.At school, children quickly acquire this contrastive meaning because of increased exposure to print and formal discourse, reinforced by formal instruction.This causes the step-change that we documented in children's understanding of discourse connectives in the early years of formal education.We would also highlight that this step-change may not only occur for connectives like but, but for a range of other discourse devices whose use in written language differs from spoken language.
From one perspective, we can see that this proposal may be considered deflationaryit is quite rare in studies of language development to admit that children acquire higher order linguistic knowledge from explicit instruction.But from another perspective we think that this proposal offers important explanatory value and opens new research questions.For example, the present perspective explains why children, before schooling, struggle to interpret discourse relations in a range of complex sentence types beyond the ones studied here, including sentences conjoined by adverbs like before and after (Clark, 1971;French & Brown, 1977;Keller-Cohen, 1987;Johnson, 1975;Trosborg, 1982;de Ruiter, Lieven, Brandt & Theakston, 2020).This perspective also highlights the potential for literacy to change children's higher-order knowledge of language, not just in its written form but also in face-to-face interactions.While the role of reading in raising phonemic awareness and manipulating speech sounds is well-established (e.g., Morais, Cary, Alegria & Bertelson, 1979), the role of reading in building new pragmatic and discourse skills for spoken language has been less explored.Together, this perspective suggests a route to supporting children's literacy instruction, by providing a clear documentation of their pre-literate discourse skills, and a pathway from written to spoken discourse.

Conclusion
These eight experiments highlight a striking limit to children's discourse comprehension that lasts until the early school years.At age 5, children systematically fail to interpret the contrastive connective but, despite a range of different facilitating conditions (Studies 3 through 5A), and they may also not understand the causal connective so (Study 5B).These limits to interpretation occur even though children use these terms in their spontaneous speech.We speculate that these limits can be explained in terms of an increasing exposure to print and the formal discourse of schooling.

Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

B
.Skarabela et al.

Fig. 1 .
Fig. 1. A. Example Stimuli for Study 1. B. Proportion of participants selecting the associated meaning by age and condition in Study 1.Large dots show group means, small dots show individual means.Error bars show bootstrapped 95% confidence intervals.

Fig. 2 .
Fig. 2. A. Example Stimuli for Study 2. B. Proportion of participants selecting the associated meaning by age and condition in Study 2.

Fig. 5 .
Fig. 5. Individual child data from Experiment 5B.Each line connects the two means of the two conditions from each individual child.

Fig. 6 .
Fig. 6. A. Individual-level behaviour of the 80 5-year-olds from Studies 3A to 5A.Each line connects data points from an individual child B. Age differences in 5-year-olds' performance across Studies 3A to 5A (small points indicate individual children) alongside logistic best fit lines to the 5year-olds' data, and mean performance at age 7 in Study 2 (large points).
Participants.16 adults and 48 children, 16 (8 female) in each age group (3-, 5-, 7-years of age) were recruited via a lab database, local nurseries and lab Facebook page.All participants were native speakers of English.Ethics approval was obtained before recruiting participants.Participants or their guardians provided consent before the study.2.2.1.2.Materials.The study included three practice items and 20 new test items.