Explaining Language Acquisition

4.1 Learnability Theory

1. 
   A class of languages. One of them is the "target" language, to be - attained by the learner, but the learner does not, of course, know - which it is. In the case of children, the class of languages would - consist of the existing and possible human languages; the target - language is the one spoken in their community.
   
2. 
   An environment. This is the information in the world that the learner has to go on in trying to acquire the language. In the case of children, it might include the sentences parents utter, the context in which they utter them, feedback to the child (verbal or nonverbal) in response to the child's own speech, and so on. Parental utterances can be a random sample of the language, or they might have some special properties: they might be ordered in certain ways, sentences might be repeated or only uttered once, and so on.
   
3. 
   A learning strategy. The learner, using information in the environment, tries out "hypotheses" about the target language. The learning strategy is the algorithm that creates the hypotheses and determines whether they are consistent with the input information from the environment. For children, it is the "grammar-forming" mechanism in their brains; their "language acquisition device."
   
4. 
   A success criterion. If we want to say that "learning" occurs, presumably it is because the learners' hypotheses are not random, - but that by some time the hypotheses are related in some systematic - way to the target language. Learners may arrive at a hypothesis - identical to the target language after some fixed period of time; - they may arrive at an approximation to it; they may waiver among a - set of hypotheses one of which is correct.
   

A key factor is the role of negative evidence, or information about which strings of words are not sentences in the language to be acquired. Human children might get such information by being corrected every time they speak ungrammatically. If they aren't -- and as we shall see, they probably aren't -- the acquisition problem is all the harder. Consider Figure 1, where languages are depicted as circles corresponding to sets of word strings, and all the logical possibilities for how the child's language could differ from the adult language are depicted. There are four possibilities. (a) The child's hypothesis language (H) is disjoint from the language to be acquired (the "target language," T). That would correspond to the state of child learning English who cannot say a single well-formed English sentence. For example, the child might be able only to say things like we breaked it, and we goed, never we broke it or we went. (b) The child's hypothesis and the target language intersect. Here the child would be able to utter some English sentences, like he went. However, he or she also uses strings of words that are not English, such as we breaked it; and some sentences of English, such as we broke it, would still be outside their abilities. (c) The child's hypothesis language is a subset of the target language. That would mean that the child would have mastered some of English, but not all of it, but that everything the child had mastered would be part of English. The child might not be able to say we broke it, but he or she would be able to say some grammatical sentences, such as we went; no errors such as she breaked it or we goed would occur. The final logical possibility is (d), where The child's hypothesis language is a superset of the target language. That would occur, for example, if the child could say we broke it, we went, we breaked it and we goed.

In cases (a-c), the child can realize that the hypothesis is incorrect by hearing sentences from parental "positive evidence," (indicated by the "+" symbol) that are in the target language but not the hypothesized one: sentences such as we broke it. This is impossible in case (d); negative evidence (such as corrections of the child's ungrammatical sentences by his or her parents) would be needed. In other words, without negative evidence, if a child guesses too large a language, the world can never tell him he's wrong.

This has several consequences. For one thing, the most general learning algorithm one might conceive of -- one that is capable of hypothesizing any grammar, or any computer program capable of generating a language -- is in trouble. Without negative evidence (and even in many cases with it), there is no general-purpose, all-powerful learning machine; a machine must in some sense "know" something about the constraints in the domain in which it is learning.

More concretely, if children don't receive negative evidence (see Section ) we have a lot of explaining to do, because overly large hypotheses are very easy for the child to make. For example, children actually do go through stages in which they use two or more past tense forms for a given verb, such as broke and breaked -- this case is discussed in detail in my other chapter in this volume. They derive transitive verbs from intransitives too freely: where an adult might say both The ice melted and I melted the ice, children also can say The girl giggled and Don't giggle me! (Bowerman, 1982b; Pinker, 1989). In each case they are in situation (d) in Figure 1, and unless their parents slip them some signal in every case that lets them know they are not speaking properly, it is puzzling that they eventually stop. That is, we would need to explain how they grow into adults who are more restrictive in their speech -- or another way of putting is that it's puzzling that the English language doesn't allow don't giggle me and she eated given that children are tempted to grow up talking that way. If the world isn't telling children to stop, something in their brains is, and we have to find out who or what is causing the change.

Let's now examine language acquisition in the human species by breaking it down into the four elements that give a precise definition to learning: the target of learning, the input, the degree of success, and the learning strategy.