This type of error analysis suggests a novel view of language learning: children are born Oxford University Press, - Language Arts & Disciplines - pages Charles Yang has been teaching computational linguistics and language.
Table of contents
- In This Article
- Children’s Acquisition of Syntactic Knowledge
- Deep Learning for Natural Language Processing
It is a simple observation that children make mistakes when they learn a language. Yet, to the trained eye, these mistakes are far from random; in fact, they closely resemble perfectly grammatical utterances by adults--who speak other languages. This type of error analysis suggests a novel view of language learning: The book presents evidence for this perspective from the study of children's words and grammar, and how language changes over time. Read more Read less. Customers who viewed this item also viewed. Page 1 of 1 Start over Page 1 of 1. The Price of Linguistic Productivity: Customers who bought this item also bought.
How Children Learn and Unlearn the Languages of th. Review "According to Yang, many of today's irregular verbs are historical survivors of what were once systematic rules Oxford University Press; 1 edition March 27, Language: Related Video Shorts 0 Upload your video. Share your thoughts with other customers.
In This Article
Write a customer review. There was a problem filtering reviews right now. Please try again later. Kindle Edition Verified Purchase. There were some great insights into knowledge acquisition which I was able to relate to as a parent of two young children acquiring English and Spanish. His arguments and models made it clear why the Spanish conjugations are much easier for them to pick up than English, and why they struggle with English past tense.
Though Professor Yang likes to say that children come into the world with all grammars, and have to forget the ones they don't use, I feel this is somewhat distracting, as it applies to the models and not the children. His predictions and explanations for languate acquisition and language change are otherwise very enlightening. One person found this helpful. Amazon Giveaway allows you to run promotional giveaways in order to create buzz, reward your audience, and attract new followers and customers.
Learn more about Amazon Giveaway. This in turn allows faster and more error-free convergence on the adult grammar. It is worth considering an example. Suppose children knew from the positive input surrounding them, that pronouns often substitute for another noun phrase, often a name, that has already been introduced in the sentence. That is, they have come to realize that in a sentence like 1a , the pronoun he can refer to the troll. The pronoun can, of course, also refer to some person who is not mentioned in the sentence but is perhaps salient in the context, but this interpretation is not our concern here.
Linguistic input of this kind could lead the child to form the erroneous generalization that a pronoun can always refer to a name that is elsewhere in the sentence. This generalization would lead children to misinterpret a sentence like 1c. However, based on input sentences like 1a and 1b , logical children would assume that sentences like 1c can mean that the troll said he himself cleared the obstacles cleanly.
The principle, known as Principle C, requires them to pay attention to the position of the pronoun and the name in the hierarchical structure of the sentence, not just to the ordering of the pronoun and the name in the sentence. The particular position of the pronoun relative to the name in the sentence hierarchy is what prevents coreference in 1c. The perspective of the generative linguistic theory is outlined first, followed by the constructivist perspective on early child representations of syntactic knowledge.
Children’s Acquisition of Syntactic Knowledge
The language component, Universal Grammar, is ready to analyze the positive input available from speakers of the surrounding language and to start building the grammar of the local language English, Mandarin, Hindi, etc. In a sense, acquiring the syntax is easy, because UG contains a computational system that generates sentence structures. The computational system provides advance knowledge of the potential kinds of elements available in human languages such as Noun, Verb, etc. Therefore, once the child has figured out what syntactic category a particular sound in the soundstream maps on to, the computational system can use the lexical items to build representations for phrases and sentences.
The representations for the phrases and sentences that children build are hierarchical structures. For example, the sentence Daddy want white milk might be represented by the child as in 2a. The finer details of the tree structure are not important—what is important is that both child and adult representations are hierarchical structures.
The child has access to the range of syntactic categories. The sentence-level category is Inflection Phrase IP shown at the top of the tree. The adult sentence representation with the tense and agreement information complete is shown in 2b. The information for tense and agreement is represented in the Inflection node, and eventually is pronounced on the main verb wants. Since the theory of UG assumes that children are born with the capacity to represent structures using the same categories and phrase structure as adults, none of this has to be learned.
Early grammars have no abstract syntactic categories. Children have to learn the range of syntactic categories and possible constructions employed in their language from the caretaker input. A phrase like Whassat? Children gradually begin to produce multi-word utterances and after considerable exposure to frequently used constructions, start to form generalizations across similar utterances and form what are known as schemas or templates.
For example, children may have accumulated the knowledge that as well as Daddy want milk , other options such as Grandma want milk , or My baby want milk , and so on are also permitted.
Deep Learning for Natural Language Processing
This list eventually is generalized to a schema: Over time, the slots become identified with syntactic categories. A hypothetical development is shown in 3 , where 3f might represent the transitive construction in the adult grammar. Although the adult grammar incorporates syntactic categories like NP and VP in the schema, the schema are not shorthand for hierarchical representations. Schema are linear representations of constructions in the language.
- Rosalind Thornton!
- Download e-book for kindle: Knowledge and Learning in Natural Language (Oxford by Charles D. Yang.
- Science Fair Warm-Up, Grades 8–12: Learning the Practice of Scientists.
- Howie Goes Shopping (I Can Read! / Howie Series) (Spanish Edition);
- The Calico Tapestry (The Paranormal Collection Book 1)?
Arguments from both theoretical perspectives on whether or not children adopt hierarchical sentence representations will be reviewed. This topic has received considerable press in the literature. This Universal Grammar endows children with the computational system that is engaged when children represent sentences in their minds. Furthermore, Chomsky argued that in cases when children need to hypothesize a rule to represent a process in the language they are acquiring, that rule must be formulated by referring to positions in the hierarchical syntactic representations provided by the computational system.
The auxiliary verb or modal is moved in the hierarchical structure to a position higher than the subject NP. The tree in 4a shows the sentence before I to C movement applies and the tree in 4b shows that the auxiliary verb is has moved to the C position in the hierarchical structure. This linear rule would, nevertheless, still give the correct result: Is the baby eating a banana? When it comes to more complex structures, the hierarchical hypothesis and the linear hypothesis diverge.
When the subject NP is modified by a relative clause, the linear hypothesis yields the wrong result. The baby who is smiling is eating a banana , in which who is smiling is the relative clause modifying the subject NP. As the tree structures in 5 show, the structure-dependent rule works as before, moving the auxiliary verb in I to C to yield the question: Is the baby who is smiling eating a banana?
But if we were to apply the linear rule to the sentence, the first auxiliary verb encountered in the linear string of words would be the is in the relative clause.
If this were moved, the resulting question would be: Is the baby who smiling is eating a banana? Clearly, this is not a grammatical question. Thirty children between the ages of 3 and 6 years participated in the experiment. Overall, the results were taken to demonstrate adherence to the structure-dependence constraint. In auxiliary doubling questions, it is not possible to tell which position the fronted auxiliary verb originated in, given that is appeared both in the relative clause and the main clause.
In a follow-up experiment, Crain and Nakayama tested 10 children who had made the auxiliary doubling errors in the original experiment. This time, they asked children to form questions from statements such as The boy who can see Mickey Mouse is happy in which the relative clause contained can and the main clause contained is. Now it would be easy to tell if children were using a linear hypothesis as the can would be doubled, instead of is , as in Can the boy who can see Mickey Mouse is happy?
No child produced questions with can doubled, thereby supporting the proposal that children base their hypotheses on hierarchical structure. This renders the debate about whether movement rules are based on hierarchical structure or linear order irrelevant. In fact, these complex questions containing relative clauses are almost entirely absent in child-directed speech. Nevertheless, according to Ambridge et al. The next step is to simply substitute a complex NP, such as the baby who is smiling for simple NPs like the baby.
To do this, children need to notice that both simple and complex NPs have the same referent i. So far, there is no empirical data demonstrating that children can do this kind of distributional analysis, however. In order to make the argument that children are capable of this kind of distributional analysis, Ambridge et al.
Lewis and Elman trained a simple recurrent network to model question formation. As Ambridge and Lieven and Gualmini and Crain before them point out, it is possible that what the network learned was that local bi-grams like who smiling are unacceptable. This bi-gram is a sub-string of the ungrammatical structure-independent question Is the baby who smiling is eating a banana? On this account, the fact that the who smiling bigram is not predicted by the model would mean that children would not attempt the ungrammatical complex question form. But, as Gualmini and Crain note, if the relative clause is changed from a subject-gap relative clause to an object-gap relative clause, a sequence of words like who smiling can be grammatical.
In the object-gap relative clause smiling is the subject NP in the relative clause. This suggests that the computational model predicts that children would not be able to produce such object gap relative clauses either. For one thing, this relies on having learned the abstract template in 6 , but, in general, constructivist researchers claim that adult abstract schemas develop late.
It is questionable whether this level of abstract schema would be in place by three to four years of age, when Crain and Nakayama show children can produce complex questions. There are other experimental data in the literature that show children manipulate hierarchical structure, rather than the wellformedness of local strings.
A study by Gualmini and Crain presented children with sentences that contained an object gap in the relative clause, ones like 8. For example, if we take just the locally well-formed piece He cannot lift the honey or the doughnut , the sentence would mean that he cannot lift the honey and he cannot lift the doughnut. In other words, he can lift neither one. Children have been shown in multiple studies in English and across language to access the conjunctive entailment Crain, The conjunctive entailment would emerge if children did not pay attention to the hierarchical structure of the entire sentence in 8 and were to attend just to the restricted part He cannot lift the honey or the doughnut.
However, suppose that children are carrying out distributional analysis and looking at locally well-formed units of words, as claimed by Ambridge et al. In this case, children could easily interpret the sentence as meaning The Karate man will give the Pooh Bear he cannot lift neither the honey nor the doughnut. The impossibility of combining the meanings of negation and disjunction in sentences like 8 is another example of structure-dependence.
In their experimental study with 3- to 6-year-old children, Gualmini and Crain showed that children analyzed disjunction correctly in sentences like 8. They never assigned the meaning that is consistent with the locally well-formed string He cannot lift the honey or the doughnut. They do not support the idea that children attend to local distributional properties of sentences.
The kinds of errors children make with wh-questions, and how the generative and constructivist theories explain them, is the next topic of discussion. At that time, wh-questions were derived from a base-structure in which an indefinite such as something , someone , etc. The first rule moved the question word to the appropriate position in the hierarchical tree structure, and the second rule accomplished subject-aux inversion or Infl to C movement , as discussed previously. Brown anticipated that children might produce wh-questions that mirrored a partial syntactic derivation in which one or both of the transformation rules failed to be carried out due to linguistic complexity.
As it turned out, children do not produce erroneous wh-questions with the wh-phrase unmoved e. These were ones such as What he can ride in? This also occurs in wh-questions. These auxiliary doubling wh-questions are ones like What can he can ride in? Children simply fail to suppress the pronunciation of the modal or auxiliary verb in the unmoved position.
Recall that usage-based accounts do not assume there is any movement, with statements and wh-questions having no derivational relationship to each other. Declaratives and wh-questions are separate constructions that children learn from the input. There is no proposal about the way in which constructions are built up that would expect children to produce the subject NP and auxiliary verb or modal in the reverse order i.
Instead, usage-based researchers propose that these nonadult wh-questions that are absent in the adult input stem from frequency effects. According to Rowland and Pine , a frame i. Each of these frames e. Lack of sufficient exposure to a specific wh-question frame causes children to cobble together a wh-question by drawing on existing constructions already in their grammar. The proposal that children overlay schema provides a neat account of the nonadult wh-questions children have been observed to produce in both spontaneous and experimental contexts.
The challenge is to prevent the overlaying of schema from occurring any time the child is unsure of how to produce a construction. From a usage-based perspective, the generative proposal fails to account for differences in inversion rates across auxiliary verbs and modals. Proponents of the constructivist theory point out that differences in inversion rates are not expected if children have acquired a general subject-aux inversion rule Rowland, A generative researcher may claim that such differences are simply due to the fact that the meaning of individual auxiliary verbs must be learned separately.
Equipped with this knowledge, the child should compute hierarchical sentence representations and have little difficulty acquiring the syntactic structures of the local language. On this account, children acquire the grammar quickly and in a relatively error-free manner, partly because their hypotheses are constrained by universal principles. On the other hand, the usage-based constructivist theory assumes that the child has no specialized knowledge of language or syntax, and must learn this, on the basis of positive input alone.
This is a slow process, because children must gradually build up knowledge of the constructions permitted in the language. The constructions are initially lexically specific schema that become more abstract over time. These are linear representations of permissible constructions. The challenge is to demonstrate how children do develop the local language without over-generating and producing sentences that are not part of the adult grammar. Proponents of the constructivist language acquisition research program have been tackling this problem in recent research cf.
Language acquisition in the absence of experience. Behavioral and Brain Sciences , 4 , — Handbook of generative approaches to language acquisition. The growth of grammar. The Oxford handbook of developmental linguistics.