One of the most significant challenges for children when learning their first language is assigning words to their corresponding syntactic categories. For instance, how do English-learning children know that ‘table’ is a noun, ‘eat’ is a verb, and ‘kiss’ can be both a noun and a verb? Generativist approaches have put forward the so-called Semantic Bootstrapping Hypothesis (Pinker, 1984; Fodor, 1998; Laurence and Margolis, 2001), which predicts that children use semantic information to map words into their corresponding grammatical category. In particular, children are said to have innately specified information in terms of nouns referring to objects, and verbs referring to actions. The present paper undertakes a critical examination of these assumptions: on the one hand, it tests the reliability of semantic information by examining the amount of nouns in children’s input that refer to specific objects or people; on the other hand, it examines the accuracy with which words could be categorized on the basis of a combination of multiple cues (i.e., both semantic and distributional cues).

The strength of the semantic bootstrapping approach lies on the idea that mappings between semantics and grammatical classes are universal (e.g., nouns denote objects while verbs denote actions in any language). On the contrary, other types of cues like phonological or distributional cues are language-specific. In particular, the noun-object mapping seems easily attainable in language development, as many studies have highlighted a noun bias in children’s early vocabularies across different languages, mainly because of the conceptual simplicity that nouns exhibit (Jackson-Maldonado et al., 1993; Bates et al., 1994; Caselli et al., 1999; Bassano, 2000; Gleitman et al., 2005). Concreteness or imageability might be the underlying predictor of the identifiability of nouns from their observed extralinguistic contexts. Thus, learners’ noun bias might be based on the assumption that object-reference items are the best ones that fit in a word-to-world pairing procedure. With the meaning of nouns, and the intuition that nouns relate to real-world objects, children might then start building a rudimentary nominal grammatical category.

Then, the semantic bootstrapping proposal claims that one source of information about the meaning of words is available from the beginning of the language learning process, and it constitutes the basis from which learners start building their initial grammatical categories. This initial information source allows the learner to acquire a subset of lexical items (i.e., nouns that refer to specific objects) which requires little linguistic knowledge and is pragmatically supported. Thus, the grounding of grammatical categories would start from the identification of semantic categories first, and semantic features would later on bootstrap grammar.

Nevertheless, the main problem within the semantic bootstrapping approach is that it presupposes that correlations between semantics and grammatical categories are perfect mappings in any language. Furthermore, such an approach also assumes that children start the language learning process with the expectation that such mappings actually exist. However, this assumed straightforward concept-word pairing is somehow problematic, as some have pointed out (Ambridge and Lieven, 2011).

To start with, naive learners with no prior knowledge of grammatical categories in their language and who are exposed to fluent speech might even fail to perform the object referent-noun mapping. Even if child learners were ready to map external referents to particular words in the input, how does the child know which word does the “object” semantic component refer to, out of the possible words she hears? First of all, the words to which children are exposed might refer to objects which are absent from the child’s sight when they are spoken (Gleitman, 1990). In addition, when addressing children, parents may refer to the same object with different words in different contexts (Yurovsky et al., 2012). Furthermore, the mapping task becomes even less clear when facing multiple-word utterances (Yu and Ballard, 2007). However, such utterances are the ones which children are most likely to encounter in the course of their linguistic development, as only a small percentage of utterances in child-directed speech have been reported to contain words in isolation (Bernstein Ratner and Rooney, 2001; Monaghan and Christiansen, 2010; Monaghan and Mattock, 2012; Feijoo and Hilferty, 2013). Furthermore, research shows that mothers never or hardly ever use words in isolation, even in situations where they are explicitly teaching new vocabulary to their children (Aslin et al., 1996).

A further problem concerns the way the object-word mapping itself should proceed. It is known as Quine’s Gavagai problem, or the problem of referential indeterminacy (Quine, 1960). Imagine a mother-child interaction situation where the mother would point to a brown running dog and say “Look at the dog!”. Even if the visual image and the target word dog were immediately associated, how does the child know that the word dog actually refers to the dog itself and not, say, to a more general referent such as animal, or to a specific type of dog, or to a part of the dog (i.e., its legs, its tail…), or to a physical property of the dog (i.e., its colour, its fur…), or to the action of running itself? How do children know that they can equally use the word dog to refer to another type of dog (i.e., a sitting dog, a white dog, a dog from a different breed, etc.) and they cannot use it with other animals like a brown running cat? A mere world-to-word mapping assumption cannot account for children’s choice and learning of the word dog and its natural referents in the real world.

Recent empirical evidence has shown that cross-situational statistical learning might be the key to solve the referential ambiguity problem illustrated in the Gavagai problem (Smith and Yu, 2008; Scott and Fisher, 2012; Vlach and Johnson, 2013; Suanda et al., 2014; Benitez et al., 2016). This statistical learning approach suggests that adults as well as young children are able to map linguistic units to referents in the world by tracking co-occurrence probabilities across different learning situations. Thus, it seems that learners are sensitive to the statistical consistency with which a given word is used in front of a given referent and the mapping between heard word and seen referent can occur in this way.

However, the noun-object mapping proposal also assumes that the set of nouns which children hear from their input specifically refer to objects in the real world. What if children were exposed to superordinate terms? Or what if they were exposed to abstract words whose meaning does not relate to a specific object or referent? Traditional linguistic analyses postulate that the category “noun” is both a notional and a grammatical concept (Lyons, 1977). There is a central semantic concept of noun, which is present in all languages, and which includes words for persons, animals, and things. All the other more abstract ontological categories denoted by nouns appear to be generalizations from this core concept.

Acknowledging that this core concept includes only a subset of all possible nouns gives rise to the question of how large the proportion of nouns belonging to this subset actually is in English. In other words, can this subset account for all the examples of nouns that English-learning children are exposed to? Previous studies have already pointed out that semantic criteria alone do not provide a reliable basis to determine the category membership of many words in English, since there are many nouns which do not denote physical objects (e.g., an explanation) (Yu and Ballard, 2007; Tare et al., 2008), or there are many words in English which can be both classified as nouns or verbs (e.g., a kiss vs. to kiss, a walk vs. to walk, etc.) (Nelson, 1995; Maratsos, 1999; Tomasello, 2010; Conwell and Morgan, 2012).

In this line, Nelson et al. (1993) propose two distinguishable semantic classes of nouns: on the one hand, BLOCS neatly correspond to basic level object categories; on the other hand, XBLOCS refer to all those words which would naturally fall out of the cognitive basic level, either because their extralinguistic referent is too general or because they do not refer to a specific object referent at all. Nelson et al. (1993: p. 71) further distinguished different types of XBLOCS taking their meaning into account:

In their study, Nelson et al. (1993) showed that children learn and use many XBLOC words early in the language learning process. As multi-word combinations and productivity in noun morphology develops, noun roles are assigned to these words, showing that they are accurately categorized as nouns. Therefore, words which lack the expected semantic content that would yield their successful categorization (and which would therefore be left unclassified on the basis of semantics alone) are nonetheless accurately classified in their right word class from the early stages of the nominal category building process.

Other studies also point out that object nouns or action verbs do not necessarily dominate children’s earliest lexical productions, neither in English, nor in other languages (Gopnik and Choi, 1995; Bassano, 2000). And still, those words which do not neatly map into their corresponding semantic category are correctly classified by children and they are used in a grammatical way.

Given the evidence provided, it seems clear that there are other types of cues at work, alongside semantics, when words are being classified into their corresponding grammatical category. For instance, at least as far as English is concerned, several studies have provided evidence for the usefulness of phonological information as a key element for the access to grammatical properties of the language (e.g., Kelly, 1996; Monaghan et al., 2005, 2007; Fitneva et al., 2009). These findings suggest that, on the one hand, these phonological cues are reliably found in parents’ language. On the other hand, evidence has also been found that young language learners as well as adults are aware of such cues and their correlation to grammatical categories.

Besides, syntactic or distributional information might be a very powerful cue that assists young language learners in language development as well. Regarding word categorization, the context of a word with respect to other words in the same sentence might provide indications about the category of that word in English. For example, English nouns are typically preceded by determiners and followed by nominal morphology (e.g., the babies), while verbs are typically preceded by auxiliaries or strong subject pronouns and followed by verbal morphology (e.g., she walked).

Studies on computer simulations have provided evidence for the usefulness of distributional and positional information for an initial categorization of words in the absence of semantic or referential information (Cartwright and Brent, 1997; Redington et al., 1998). Such distributional information appears to be available not only to adult speakers but also to young language learners as well (Mintz, 2003; Monaghan et al., 2007; Feijoo et al., 2015).

Furthermore, empirical evidence from artificial language studies seems to suggest that children’s learning of grammatical structure as well as their word-reference associations improve when words are coherently marked by a combination of different types of cues, either phonological, distributional or semantic cues (Gomez and Lakusta, 2004; Gerken et al., 2005; Lany and Gomez, 2008; Lany, 2014). However, when words are not reliably marked by these cues, infants fail to learn their semantic or grammatical properties, since young language learners are more likely to learn from deterministic rather than probabilistic cues, and they would only rely on relatively robust correlations between word-forms and their corresponding grammatical category (Lany and Saffran, 2010; Yurovsky et al., 2012; Lany, 2014).

In particular, Lany and Saffran (2010) found that experience with reliable distributional cues in the input is a key factor that predicts children’s learning of word meanings: when words’ distributional properties correctly indicated the grammatical category to which words belonged, infants successfully learned word-referent mappings. In contrast, infants failed at the word-referent pairing task when distributional cues were not reliably correlated with grammatical category membership.

The main goal of the present study was to test the likelihood with which children could classify all the nouns they hear in their corresponding grammatical category using semantic cues derived from the input. To this end, two studies were carried out: in the first one, a corpus-based analysis of child-directed speech explores the potential strength of semantic information alone in children’s input. The second study examines the benefits that a combination of semantic and distributional information could provide to language learning children when facing the task of word categorization.

As seen earlier, the objective of the first study described in the present paper was to analyze the potential usefulness of semantic information as far as the categorization of English nouns is concerned. Traditional accounts on noun categorization based on semantic information have put forward the idea that young language learners might group all nouns together under the semantic label of “object” and all verbs together under the semantic label of “action” (Pinker, 1984). The fact that most nouns refer to common objects and their subsequent imageability based on notional grounds has been argued to be the reason why nouns are learned before verbs or before words encoding actions and relations in language development (Caselli et al., 1999; Bassano, 2000; Gleitman et al., 2005).

However, more recent findings show that nominal elements denoting common objects dominate children’s early vocabularies only as far as types are concerned (Nelson et al., 1993; Gopnik and Choi, 1995; Nelson, 1995). However, reports on early vocabulary production show that there are more tokens of non-nominal expressions (i.e., verbs and relational words such as there, up or no) than of nouns, and that this tendency is not only true of the English language (Gopnik and Choi, 1995).

Thus, as the authors suggest, the makeup of children’s early lexicons might not be the result of there being a more “learnable” or more “imaginable” category in semantic terms, but it might be a reflection of children’s actual linguistic experience, in such a way that children’s first words might be instantiations of the kind of words that their parents used with them. This is also coherent with the results obtained from the input analysis undertaken in the present studies. As seen earlier (see Table 2), the descriptive data that was obtained from the corpus preparation reveal that the kind of linguistic experience to which the four children under consideration are exposed contains far more nominal types than non-nominal types (i.e., there were overall 5,388 noun types and 4,233 other open class word types). However, when it comes to tokens, other open class words exceed nominal tokens by far (i.e., the four corpora together contained 51,577 nominal tokens, but 88,047 other open class word tokens).

The predominance of noun types over other open class word types has been explained by the fact that many very complex and abstract entities are realized as nouns in adult language (Nelson, 1995). This raises the question of how many of these nominal types are actually “easily learnable.” If children engage in word categorization tasks guided by the fact that all words that refer to common objects are grouped together under the noun category, then how many of the overall 5,388 noun types considered in the present study can actually be described by these semantic features and how salient is that proportion in statistical terms?

Previous analyses of linguistic input addressed to young language learners have shown that common object nouns represent only a very small proportion of all the noun repertory that children hear (Nelson et al., 1993; Nelson, 1995; Monaghan and Mattock, 2012). The initial prediction as far as the first study was concerned was that many of the nouns to which very young English-learning children are exposed refer to basic-level common objects, and will be subsumed by variable Sem2. Other nouns were expected to either refer to proper names of people (i.e., Sem1 nouns) or to non-discrete mass entities (i.e., Sem3 nouns). Neither of them would pose any learning problems either. However, an important number of nominal elements were also predicted to lack any of the above-mentioned semantic characteristics (i.e., Sem4 nouns). Without the necessary semantic components, those nouns would not guide children in their categorization tasks, provided children perform such tasks on the basis of semantic information.

These predictions are confirmed by the data obtained in the first analysis. The descriptive data from the first study show that only approximately half of the noun types as well as half of the nominal tokens refer to basic-level object nouns and were described by variable Sem2, while the other half belonged to the other three semantic subsets. Within those, about a third of the noun types and a quarter of the noun tokens were included in variable Sem4, which was the one which grouped together all nouns which did not have any of the semantic features that would foster correct grammatical categorization on the basis of semantic information alone. The results obtained from the discriminant analyses performed with types as well as with tokens confirm this weak correlation between semantic information and grammatical category assignment as far as nouns are concerned. Thus, for the type analysis, only 67.9% of the nominal types were correctly classified, and these completeness scores dropped dramatically in the token analysis, with only 39.9% of correctly classified nominal tokens.

A further objective in this first study was to see whether there was an overlap between nouns and other open class words as far as semantic information is concerned, and to test whether the overlap was significant enough so as to bring about a considerable misclassification of elements. Provided that, according to Pinker (1984), children engage in a semantic analysis of the input and make the hypothesis that all words that denote actions belong to the grammatical category of verbs, what do children do when they encounter action words which are not verbs? And what do they do with verbs that do not denote actions? Does the input offer a high proportion of contradictory information of this kind?

The present analysis shows that, indeed, there is a slight degree of semantic overlap between nouns and other open class words, since some of the XBLOC nouns have certain semantic features which are typical of verbs (i.e., mainly nouns that denote actions). At the same time, obviously not all of the other open class words denoted actions since, besides non-action verbs, adjectives, and adverbs were also included in this group. Thus, in the case of semantic information, the kind of overlap between nouns and other open class words seems to be one way, that is, some of the nouns might lack the corresponding nominal semantic features, or might have verbal semantic features, and might therefore be misclassified as other open class words (i.e., mainly misclassified as verbs). However, the same risk does not seem to hold for any of the other open class words, since they are very unlikely to contain any of the semantic features associated to nouns, and thus be misclassified as such on the basis of semantic information. Empirical analyses of children’s early vocabularies also suggest that, even when the very same word can be both a noun and a verb (e.g., English kiss, hug, call, help) most children assign those action words exclusively to the verb category, regardless of their parents’ frequent use of them as nouns (Nelson, 1995). As mentioned earlier, the discriminant analyses using semantic variables that were performed on all types and tokens confirm this, since the accuracy scores that were obtained in all cases were very close to 100% of correctly classified other open class words.

The fact that correct classification scores were far better with types than with tokens confirms the tendency described above in connection to early vocabulary production, where higher productivity of nominal types is observed, while tokens from other categories outnumber nominal tokens. Thus, children’s early word production can be seen as a direct reflection of the kind of linguistic environment that they have experienced. However, nouns cannot be claimed to be “easily learnable” on the basis of their semantic association to basic-level objects alone, since statistical analyses where the diagnosticity of such semantic classification was tested provided a considerable proportion of misclassified nouns.

These findings suggest that semantic information may not always be the only factor which is used to determine the assignment of words to their grammatical category. Furthermore, as other studies have also suggested (Monaghan et al., 2007; Monaghan and Mattock, 2012; Yurovsky et al., 2012), learners never hear speech with just a single kind of cue to word categorization in isolation. In children’s natural linguistic environment, there are multiple redundant language-specific cues to word category membership.

When distributional and semantic cues were put to interact with one another in the second study, overall successful categorization scores were expected to improve when compared to the results obtained with semantic cues in isolation. This prediction was born out by the results obtained from the second study. The interaction between semantic and distributional cues gave more successful results than semantic cues in isolation. The results obtained regarding successful categorization with both kinds of cues in combination were also higher than the results obtained in previous studies using distributional cues alone (Mintz, 2003; Monaghan et al., 2007; Feijoo et al., 2015). In this sense, the results from both our first and our second study suggest that semantic cues contributed in providing an accurate grammatical category of nouns, since very few other open class words are at risk of being misclassified as nouns. On the other hand, distributional cues might have contributed in providing greater completeness scores, with a larger number of nouns being correctly classified as nouns, since the low completeness scores obtained from semantic cues alone were improved.

Monaghan et al. (2007) have already proposed the Phonological-Distributional Coherence Hypothesis in their analysis with distributional and phonological cues. According to them, both sources of information contribute differently toward word classification. Other studies have also highlighted the benefits of multiple types of cue for word categorization as well as for other language learning tasks (Monaghan et al., 2007; Smith et al., 2010; Riordan and Jones, 2011; Monaghan and Mattock, 2012; Yurovsky et al., 2012). Thus, having a combined number of different variables seems to increase the likelihood with which a given element will be successfully encoded and learned by children.

On the other hand, having to attend to several types of cues does not seem to imply an increase in terms of difficulty or cognitive processing demand on the part of very young language learners, at least as far as the combination of semantic and distributional information is concerned. The empirical evidence available to date seems to suggest that children are able to attend to multiple cues and use them for language learning tasks from a very early age (Thiessen and Saffran, 2003). In particular, by 14 months, infants have been described to be able to use determiners to identify nouns (Höhle et al., 2004; Shi and Melançon, 2010). Furthermore, hearing a word in a predictable distributional context promotes its identification for adults as well as for young infants (Lany, 2014). Thus, young language learners may also more readily encode a novel noun when it is preceded by a determiner. Facility with encoding would also make it easier for infants to determine the referent of the novel noun, and to form a robust mapping between the noun and the object it refers to, provided such mapping exists.

Therefore, on the basis of the evidence provided by the data obtained in our studies, we can claim that the combination of semantic and distributional information found in natural child-directed speech could significantly contribute to the correct categorization of most of the nouns to which English-learning infants are exposed. However, the results reported in our two studies are restricted to the kind of English speakers considered in our analyses. There might be important differences -both in terms of quantity and quality- in the input used with infants exposed to languages other than English, or infants from cultures other than western cultures and who are conventionally spoken to in a different way. Thus, for example, other languages with less distributional regularities might exploit phonological or prosodic information that might assist young language learners in their word categorization tasks. Further research on child-directed speech by non-English speakers should shed some light on how the results obtained from the present studies would generalize cross-linguistically and cross-culturally.

SF, CM, AA, and ES: Contributions to the conception and design; analysis and interpretation of data; drafting the work and revising it critically; final approval of the version to be published; agreement to be accountable for all aspects of the work in ensuring that questions related to its accuracy or integrity are appropriately resolved.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.