The acquisition of compositional meaning

Abstract

How do people create meaning from a string of sounds or pattern of dots? Insights into this process can be obtained from the way children acquire sentence meanings. According to the well-known principle of compositionality, the meaning of an expression is a function of the meanings of its parts and the way they are syntactically combined. However, children frequently seem to ignore syntactic structure in their sentence interpretations, suggesting that syntax is merely one of the sources of information constraining meaning and does not have a special status. A fundamental assumption in the argument in favour of compositionality is that speakers and listeners generally agree upon the meanings of sentences. Remarkably, however, children as listeners do not always understand what they are able to produce as speakers, and vice versa. For example, children's production of word order appears to develop ahead of their comprehension of word order in the acquisition of languages like English and Dutch. Such production–comprehension asymmetries are not uncommon in child language and motivate a view of compositionality as a principle pertaining to the result of perspective taking, and of meaning composition as a process of speaker–listener coordination.

This article is part of the theme issue ‘Towards mechanistic models of meaning composition’.

1. Principle of compositionality

How do people create meaning from a string of sounds or pattern of dots? Skipping over non-trivial aspects of interpretation such as identifying the basic units of speech and language and determining their structural relationships, an important question is how listeners and readers arrive at the meaning of a sentence on the basis of the meanings of the individual words in the sentence. In natural language semantics, a fundamental principle is the principle of compositionality of meaning. According to this principle, ‘[t]he meaning of an expression is a function of the meanings of its parts and of the way they are syntactically combined’ [1, p. 153]. The principle of compositionality is controversial because of the presence of non-compositional linguistic structures such as idioms and the influence of linguistic context and conversational setting on sentence interpretation. Also, there is debate about whether the principle is an empirical claim, a methodological assumption, or a trivial truth or falsehood (e.g. [2]). Nevertheless, it serves to explain important properties of meaning construction such as its productivity and systematicity, as Dowty [3] shows.

Goldberg [2] summarizes the argument in favour of compositionality presented by Dowty [3] as follows:

• (a)Speakers produce and listeners parse sentences that they have never spoken or heard before.
• (b)Speakers and listeners generally agree upon the meanings of sentences.
• (c)Since there exists an infinite number of sentences, they cannot all be memorized.
• (d)There must be some procedure for determining meaning.
• (e)Sentences are generated by some grammar of the language.
• (f)The procedure for interpreting sentences must be determined, in some way or the other, by the syntactic structures generated by the grammar together with the words.

Steps a, c and d account for the productivity of meaning composition, whereas steps e and f account for the systematicity of meaning composition. Step b is a fundamental assumption underlying the argument for compositionality, but one that is problematic for some versions of compositionality, as I will explain below.

Although meanings must in some way be dependent on syntactic structure, as step f indicates, Dowty [3] points out that this does not imply that compositionality specifies how meanings depend on syntactic structure. That is, it is not necessarily true that syntactic structure fully determines meaning composition, and it is an empirical issue how exactly syntax contributes to meaning composition. This contrasts with the way the principle of compositionality is often used in linguistics, namely as the claim or assumption that semantic composition must follow syntactic composition in every step, from the combination of individual words into phrases to the combination of phrases into a sentence.

In this paper I discuss the process of meaning composition and the role of syntax in meaning composition by considering children's acquisition of simple transitive sentences such as The car is pushing the boy. The paper reviews the literature on this topic and concludes that children's production of subject–object word order in languages such as English appears to be ahead of their comprehension of subject–object word order. This posits a challenge for traditional views on meaning composition, but is compatible with an alternative view according to which meaning composition requires speakers and listeners to coordinate their choices.

2. Children's acquisition of compositional meaning

An important aspect of the acquisition of sentence meaning is the acquisition of argument structure, that is, the ability to determine who is doing what to whom. In transitive sentences in English, argument structure is expressed by word order: the first noun phrase in the sentence is the subject and the second noun phrase is the direct object. Several studies have shown that English-learning children between the ages of 1 and 2 are already sensitive to the word order of English when interpreting transitive sentences (e.g. [4–6]). For example, Hirsh-Pasek & Golinkoff [6] carried out an intermodal preferential looking task with English-learning children from seventeen to nineteen months, in which two videos were presented simultaneously while a sentence was played. The study used simple transitive sentences such as Big Bird is tickling Cookie Monster. Because all sentences contained two animate nouns and a reversible verb, word order was the only cue for interpretation. One video showed the event with the first noun as the actor (e.g. Big Bird tickling Cookie Monster), and the other video showed the event with the second noun as the actor (e.g. Cookie Monster tickling Big Bird). Hirsh-Pasek and Golinkoff found that the children were guided by word order and looked at the video matching the word order of the sentence more than at the non-matching video. Later studies (e.g. [4,5]) confirmed these results. This suggests that, in English, word order is available as a cue to sentence interpretation from roughly 1.5 years.

However, comprehension studies using transitive sentences in which word order is not the only cue for interpretation found that 2-year-old English-learning children still make a large number of errors in their interpretation of simple transitive sentences [7–10]. These studies, all using an act-out methodology in which children had to act out their interpretations with toys, found that 2-year-old English-learning children show variability in their interpretation of simple transitive sentences, depending on the presence of other cues. For example, Chapman and Miller investigated the comprehension and production of word order in children between 1;8 (1 year, eight months) and 2;8 old, with a mean age of 2;2, and found that, when the subject was animate and the object was inanimate, as in the sentence The girl is pulling the boat, children's interpretations were largely correct (93.8% correct). In contrast, when the subject was inanimate and the object was animate, as in the sentence The car is pushing the boy, their interpretations were correct in only about half of the cases (50.1% correct). The incorrect responses revealed that children interpreted the animate noun phrase (here: the boy) as the subject, rather than the first-mentioned noun phrase (here: the car). Thus, the children in Chapman and Miller's study appeared to ignore word order in favour of interpretations in which an animate entity acts upon an inanimate entity. Chan et al. [7] obtained similar results testing English-learning children on novel rather than familiar verbs. Children aged 2;6 performed lowest on sentences with an inanimate subject and an animate object (58% correct), and best on sentences with an animate subject and an inanimate object (86% correct). In contrast, children aged 3;6 in their study interpreted all sentence types based on word order (all at least 95% correct). These findings suggest that English-learning children show an adult preference for word order over animacy in comprehension from age 3 (cf. [10]).

Interpreting the animate noun phrase as the subject is consistent with the observation that, cross-linguistically, subjects tend to be animate whereas objects tend to be inanimate [11]. In some languages, such as the Papuan language Fore, interpreting the animate noun phrase rather than the first noun phrase as the subject would in fact be the grammatical choice for adults [12]. These observations suggest that animacy is not merely a cognitive heuristic or bias that is used when syntactic information is unavailable, but rather must be treated on a par with word order and other syntactic cues.

A widely accepted view in language acquisition research is that correct comprehension of words and grammatical structures precedes their correct production. Some have even called this a logical necessity: ‘Logically, comprehension must precede production. How else can speakers know which words to use to convey a particular meaning?’ [13, p. 246]. The idea is that children first learn to comprehend a particular form, and only later learn to use this form correctly. Children's acquisition of subject–object word order seems to deviate from this general pattern. As described above, 2-year-old English-learning children show variability in their comprehension of subject–object word order and make many comprehension errors. However, the same 2-year-old children's use of subject–object word order in their productions appears to be adult-like, even when the subject and the object differ in animacy ([8,9]; see also [14] for a study examining production only in children between 2;4 and 2;11 old). For example, in Chapman and Miller's study the mean correct responses by the 2-year-olds in the production task were between 80% and 90% for all sentence types. Thus, 2-year-olds' production of word order is closer to adult performance than their comprehension of word order.

The finding of poor comprehension of word order in studies such as Chapman & Miller's [8] yields a challenge for principles and parameters theories of grammar as well as usage-based approaches to language and has been dismissed as a confound of the complex task demands (e.g. [15]). After all, act-out tasks are cognitively demanding and are known to be prone to certain response biases. Therefore, Cannizzaro [16] carried out a picture selection task and a preferential looking task, as well as a parallel sentence elicitation task, with Dutch- and English-learning children between 2;3 and 4;1 old and adult controls. In Dutch, like in English, main clauses generally have subject–verb–object word order, although Dutch word order is more flexible despite the absence of overt case marking on full noun phrases and also allows for object–verb–subject word order in specific discourse contexts (that were not part of Cannizzaro's study). The picture selection task used simple transitive sentences out of context such as The car is pushing the cow, in which the animacy of subject and object were manipulated. Upon hearing the sentence, children had to select the matching picture from two animated pictures that were presented simultaneously. In the elicitation task, children were asked to describe the event depicted in an animated picture. Note that in controlled elicitation tasks, in contrast to corpus studies, one can be reasonably sure about the intended meaning of a produced form and whether the first noun is intended to be the subject or the object, as the meaning to be expressed is provided visually. For the Dutch-learning children with a mean age of 2;9 (younger group) and 3;8 (older group), correct performance in comprehension was 54% (younger) and 70% (older), and in production 81% (younger) and 92% (older). For the English-learning children with a mean age of 2;9 (younger group) and 3;9 (older group), correct performance in comprehension was 60% (younger) and 80% (older), and in production 79% (younger) and 94% (older). So both English- and Dutch-learning children were found to be better at correctly producing subject–object word order in the elicitation task than they were at correctly comprehending subject–object word order in the picture selection task, and older children were better than younger children. The cognitively less demanding preferential looking task did not make it easier for children to comprehend transitive sentences. Whereas Dutch- and English-speaking adults showed a clear preference for looking at the matching picture within the first 1000 ms of hearing the sentence (with a mean proportion of looks to this picture close to 0.80), neither the Dutch- nor the English-learning younger or older children showed a preference for the matching picture in the analysed time frame of 3600 ms (with their mean proportion of looks to this picture staying below 0.60 for Dutch and reaching just above 0.60 for English). Dutch-learning children's comprehension was affected by animacy, in line with earlier studies on English using an act-out methodology [7,8,10]. No animacy effects were found for the English-learning children's comprehension, likely because the children were older than in earlier studies. This difference in animacy effects between the two languages suggests that word order comprehension is acquired at a slightly later age in Dutch than in English, possibly because Dutch-learning children receive less evidence for subject–object word order due to the more flexible word order of Dutch compared to English.

In all, these findings confirm the earlier suggested pattern that, in the acquisition of a first language in which argument structure is marked by subject–object word order, production of word order precedes comprehension of word order.

3. Asymmetries between production and comprehension

In recent years, it has become apparent that, beyond subject–object word order, several other aspects of language also show this developmental pattern of production preceding comprehension. Based on evidence from different types of psycholinguistic experiments, this puzzling developmental pattern has been suggested to arise for various linguistic forms in various languages (see [17] for an overview and discussion). For these linguistic forms, children were found to consistently produce the form correctly (i.e. in the appropriate sentence context and discourse context), but fail to consistently interpret the same form correctly. These asymmetries seem to occur in phonology, morphosyntax, semantics as well as pragmatics, indicating that production–comprehension asymmetries cover the entire spectrum of language and are not restricted to particular areas of language.

To give another example from the domain of morphosyntax, the English third person singular inflectional morpheme -s emerges early in children's spontaneous production and is produced largely correctly before age 4. In children's comprehension, however, this inflectional morpheme is used reliably as a cue for sentence interpretation only from age 5 on, according to a study by Johnson et al. [18]. Since the finite verb in English must agree with the subject, the presence or absence of the morpheme -s on the verb indicates whether the subject is a singular or plural entity. In their study, Johnson and colleagues presented children with sentence pairs such as The duck swims in the water versus The ducks swim in the water. In these sentences, the onset of the verb swim masks the plural morpheme -s on the subject noun, as a result of which listeners cannot tell whether the subject is singular or plural based on the subject form alone, but need to consider the inflectional morpheme -s on the verb. When children in a picture selection task had to select the correct picture for one of these sentences from two presented pictures (a picture showing one duck and another picture showing two ducks), 3- and 4-year-old children failed to use the inflectional morpheme -s in their comprehension and performed at chance level. Only 5- and 6-year-olds reliably used the inflectional morpheme -s as a cue in comprehension. So whereas English-learning children use verbal number inflection reliably in their production from the age of 3 on, they do not use it in comprehension when it is the sole cue to subject number until the age of 5.

A third example of an asymmetry between production and comprehension in English is the mismatch between 4- and 5-year-old children's poor comprehension of object pronouns such as him and her and their adult-like production of these forms. Children aged 4 and 5 frequently interpret her in questions such as Is Mama Bear touching her? as herself (e.g. [19]), but do not confuse her and herself in production [20]. This pattern of errors is known as the Delay of Principle B Effect, since it pertains to the syntactic binding principle B, and pronoun interpretation is delayed compared to children's interpretation of reflexives such as himself and herself. This pattern is attested in various languages and various experimental tasks, including act-out tasks, picture selection tasks and truth-value judgement tasks, but does not seem to occur in production.

Summarizing, when acquiring morphosyntactic knowledge, such as knowledge about word order, inflection or pronouns, children's production sometimes precedes their comprehension.

4. Explaining and resolving asymmetries

Having established that production may sometimes precede comprehension in children's linguistic development, thus giving rise to production–comprehension asymmetries, let us now turn to possible explanations for such asymmetries.

Asymmetries between production and comprehension are puzzling from the perspective of principles and parameters approaches to language acquisition (e.g. [21]). If a child displays mature production of word order, according to these approaches this serves as proof that the child possesses knowledge of the relevant morphosyntactic principles, and that the relevant parameters with respect to constituent order have been set (e.g. complement-head parameter, specifier-head parameter). So if the same child is then found to make errors with word order in comprehension, the source of these errors must lie outside the grammar. It is not possible to explain observed differences between production and comprehension of word order as the outcome of the same set of morphosyntactic principles and parameters. For example, if an English-learning child shows knowledge of the grammatical principles and parameters that result in subjects preceding objects, as evidenced by the child's mature production of word order, the same principles and parameters must also be obeyed in comprehension, resulting in the first noun phrase in the sentence being interpreted as the subject and the second noun phrase as the object. Thus, asymmetries between production and comprehension require an extra-grammatical explanation in a principles and parameters approach. However, in many cases it turns out to be challenging to explain a particular asymmetry by extra-grammatical factors such as task effects, immature pragmatic skills or processing limitations. Task effects do not explain why, using the same task, an asymmetry may be observed in one language but not in another, immature pragmatic skills are expected to hamper production as well, and processing limitations alone do not explain why asymmetries occur with some forms but not with related forms (see [17] for discussion).

A similar challenge arises for usage-based approaches to language acquisition (e.g. [22]). According to these approaches, children learn grammatical patterns like word order on a verb-by-verb basis. Initially, their knowledge of grammar is limited to certain lexical items (e.g. certain verbs), and only later do they generalize this grammatical knowledge. If a child shows knowledge of the order of subject and object with particular verbs (for example, push and pull) in their production of transitive sentences, this suggests that they have learned the transitive frames for at least these verbs. They are then expected not to make mistakes in their comprehension of transitive sentences containing these same verbs either. If they do make mistakes in comprehension, this requires an extra-grammatical explanation in a usage-based approach as well. However, the same concerns about extra-grammatical explanations apply here.

In contrast, asymmetries between production and comprehension follow naturally from constraint-based grammars such as optimality theory, in which comprehension and production are viewed as different directions of use of the same grammar (e.g. [17,23,24]). According to optimality theory, the grammar is a set of constraints mapping an input to the optimal output. The optimal output is the output that satisfies the set of hierarchically ranked and potentially conflicting constraints best. In comprehension, the form is already given as the input (e.g. the sentence The car is pushing the boy) and the optimal meaning must be selected for this form. In contrast, in production the meaning is already given as the input (for example as a predicate-argument structure) and the optimal form must be selected for this meaning. Smolensky [23] argues that this type of direction-sensitive grammar is able to explain why comprehension usually precedes production in child language. The explanation is based on the hypothesis that, in children's grammar, constraints preferring unmarked forms (i.e. typologically frequent and simple forms such as syllables ending on a vowel) are initially ranked too high compared to the adult grammar. As a consequence, young children are predicted to frequently produce reduced forms (such as ta to refer to a cat, omitting the final consonant and simplifying the k-sound to a t-sound), while the same constraints under the same ranking give rise to the adult meanings for the full forms (e.g. the word cat).

Based on the same direction-sensitivity of the grammar, an optimality-theoretic grammar also predicts reverse asymmetries where production precedes comprehension, depending on the particular configuration of constraints [24]. Two relevant constraints for explaining word order variation are an animacy constraint requiring the subject to be higher in animacy than the object and a word order constraint requiring the subject to linearly precede the object [25,26]. If the animacy constraint is stronger than the word order constraint, animacy will overrule word order when selecting the optimal interpretation. This would explain children's non-adult interpretations, that seem to be based on features such as the animacy properties of the noun phrases rather than the order of the noun phrases in the sentence. In contrast, the adult interpretations would be explained by the inverse ranking of the two constraints and satisfy the word order constraint. In production, whether the subject and the object of a transitive relation are animate or inanimate is already given in the meaning to be expressed and is therefore irrelevant for the selection of the optimal order of subject and object. Hence, the relative strength of the two constraints is also irrelevant and the optimal form will always satisfy the word order constraint. As a consequence, children are expected to consistently produce sentences with mature subject–object word order. So young children's performance in word order production is predicted to exceed their performance in comprehension.

Most asymmetries between production and comprehension are resolved during the normal course of language development. In optimality theory, children are hypothesized to learn the grammar of their language through two distinct learning mechanisms: constraint reranking and bidirectional optimization (e.g. [17,27]). Both mechanisms require that the child combines the direction of production with the opposite direction of comprehension (called robust interpretive parsing in constraint reranking algorithms). This allows them to reduce the number of errors and converge to the adult pattern. Constraints are reranked if the optimal output according to the child's grammar is different from the actually encountered form or meaning. Through this mechanism, children are expected to rerank the animacy constraint in a step-wise manner (or gradual manner, if a stochastic version of optimality theory is assumed) until it ends up below the word order constraint, upon encountering suitable evidence such as sentences with an inanimate subject and an animate object in a conversational context that makes clear the meaning of the sentences. Thus, asymmetries such as the word order asymmetry will disappear as a result of constraint reranking. For other asymmetries, such as the pronoun asymmetry, constraint reranking does not result in a symmetric pattern due to the configuration of the constraints involved [24]. These asymmetries can however be resolved if children take into account the opposite communicative perspective in their own choices, a process that can be modelled formally as bidirectional optimization. This allows the child to block a non-adult form or meaning if this form or meaning is better associated with another input. So a mature symmetric pattern of production and comprehension, according to which speakers and listeners in principle agree upon the meanings of sentences, only arises when speakers consider the listener's perspective and listeners consider the speaker's perspective.

5. Relating perspective taking and theory of mind

Resolving asymmetries between production and comprehension thus relies on the ability of perspective taking in language. This raises the question whether and how perspective taking in language relates to the cognitive ability of theory of mind, that is, the ability to attribute beliefs, intentions and desires to oneself and others and to understand that others can have beliefs that are different from one's own [28]. Children's theory of mind abilities develop gradually over time, starting with the ability to make implicit inferences about other people's intentions and desires around age 1.5, then passing explicit first-order false-belief tasks at around age 4 or 5, followed by the development of more complex and flexible mindreading abilities such as the comprehension of faux pas after age 6 (e.g. [29]). A precursor of this cognitive development may be the emergence of joint attention abilities in infants at around 1 year of age [30].

The proposed account for explaining asymmetries in child language could suggest a general two-stage developmental process in which children first acquire the adult constraint ranking and next develop a general ability for perspective taking in language, possibly supported by emerging theory of mind abilities. However, the developmental trajectories of the various asymmetries do not align. In English, mature comprehension of subject–object word order is acquired before mature comprehension of verbal number inflection (the former at age 3, the latter at age 5), which is acquired before mature comprehension of object pronouns (at age 6). This indicates that each of the asymmetries is resolved separately and requires the development of its own perspective-taking step, which builds upon its own previous step of constraint reranking. The resolution of these different asymmetries may be related to different stages of theory of mind development, but how exactly they relate is still largely an open issue.

6. Mechanistic models of meaning composition

Returning to the issue of compositionality, the observed asymmetries between production and comprehension in child language appear to go against the fundamental assumption underlying the principle of compositionality that speakers and listeners generally agree upon the meanings of sentences (step b in Goldberg's summary of Dowty's argumentation). Children as listeners do not always understand what they are able to produce as speakers, and vice versa. Thus, the sentences produced by child speakers are not always compositional in the sense described by Dowty [3] and Goldberg [2]. For speakers, sentences are only guaranteed to be compositional when they consider the opposite perspective of the listener. Thus, compositional meaning does not necessarily arise as soon as two words are combined, as it is not guaranteed that the listener retrieves the same meaning from this word combination as the speaker intended to express. Rather, meaning composition must involve speaker–listener coordination, and hence compositionality should not be viewed as a principle describing the result of word combination, but rather as a principle describing the result of perspective taking (see also [31]). Note that perspective taking in this sense is not restricted to complete sentences. Under an optimality-theoretic account, perspective taking has been argued to take place during word-by-word sentence processing and hence applies to sentence fragments as well [32].

Under the optimality-theoretic account discussed above, syntactic structure does not fully determine meaning composition. Instead, syntax is merely one of the sources of information constraining meaning and does not have a special status. For example, the word order constraint is only one of the constraints playing a role in the interpretation of transitive sentences, and is sometimes overruled by other constraints, such as the animacy constraint in child language. Similarly, contextual constraints such as those pertaining to discourse prominence may play a role and could even overrule syntactic constraints. Despite the less central role of syntax here compared to traditional views on compositionality, meaning is computed in a compositional manner, since the form produced by the speaker (be it a complete sentence, a sentence fragment, or a word) for a given input meaning on the basis of the constraints of the grammar should lead to the same output meaning again when the listener interprets the produced form on the basis of the same constraints under the same ranking in the same (sentence, discourse, visual and knowledge) context. Therefore, this type of meaning composition is different from the simple symbolic and modular type of composition that is rejected by Baggio et al. [33] on the basis of behavioural and neurophysiological evidence.

Because optimality theory is a formally defined grammar and the learning mechanisms of constraint reranking and bidirectional optimization are also formally defined, the mapping from an input form or input meaning to its optimal output corresponds to Marr's computational level [34]. By integrating the constraints of the optimality-theoretic grammar in a cognitive architecture, the processes involved in meaning composition are also specified at Marr's algorithmic level. See for example van Rij et al. [32], who integrated optimality-theoretic constraints in the hybrid computational cognitive architecture ACT-R, which combines symbolic with subsymbolic processes. At this level, cognitive factors such as working memory limitations and insufficient processing speed can affect the input–output mapping of the grammar, resulting in occasionally unsuccessful perspective taking and individual differences in linguistic abilities [17]. Thus, the model of meaning composition presented here is a mechanistic model that specifies the linguistic and cognitive mechanisms involved in meaning composition. Indirectly, through a cognitive architecture that supports the mapping from cognitive model to brain (such as ACT-R, see [35]), the processes involved in meaning composition may also ultimately be specified at the implementational level. Note that, while cognitive modelling is one way to relate an optimality-theoretic account to Marr's levels of analysis, Smolensky & Legendre [36] propose three levels of representation within optimality theory that relate in a different way to Marr's three levels of analysis.

To conclude, this paper argued that production–comprehension asymmetries in child language motivate a view of compositionality as a principle pertaining to the result of perspective taking, rather than to the mere combination of words. Only after perspective taking is it guaranteed that speakers and listeners agree on the meaning of linguistic forms.

Data accessibility

This article has no additional data.

Competing interests

I declare I have no competing interests.

Funding

I received no funding for this study.