Drawing rings around the world - The acquisition of meaning

Contents

1 The Sky‛s the Limit
1.1 Introduction
1.2 Approaches to Language Acquisition
1.3 The Quest for Semantic Universals

2 Rings Around the World
2.1 Prototype Theory
2.2 Semantic Feature Theory
2.3 Functional Core Concept Theory

3 Are You Talking to Me?
3.1 A Brief History of Machine

4 Discussion
4.1 Colourless Green Ideas

5 References

6 Confirmation of Authorship

1 The Sky‛s the Limit

Der Horizont des Denkens ist undehnbar vorgegeben, weil das Denken in dem Gedankenlosen, aus dem es sich erhebt, verankert ist...Völlige Freiheit des Denkens im Sinne eines gänzlich ungehinderten Erfassens beliebiger Objekte ist eine Utopie, denn ihr könnt nur so weit denken, wie das Organ eures Denkens es zuläßt.

GOLEM XIV, (Lem 1984: 79)

1.1 Introduction

This is the implication of the anthropic principle: a universe, that allows human existence can only be observed and conceptualised – linguistically structured, for our purpose – by human beings due to the constraints set by the internal structure of the universe. This is, any theory about the world and about the phenomena it consists of has to take into account that the available data possibly reflects only a sample of the topic to be theorised. Absolute objectivity is a myth. Theories should be regarded as points of view, in its very sense.

The fictional GOLEM XIV, artificial anthropologist at the MIT of the year 2029, encounters the problem from an opposite perspective: he knows well about man and the thematic role, man holds in universe. But he is not able to communicate about his knowledge without cutting down on his actual message because natural language does not provide the concepts required.

From this point of view, the study of semantic acquisition becomes a methodological one. When communicating linguistically about the world we live in and act on, how do we know that our words – those Saussurian signs with a concept on one side and an arbitrary form on another – actually cover the intended objects and ideas in their entirety? At least, the question is worth to be asked.

Bloom (2001) extends the question to at least two specifications. First, when discussing Williard Quine´s gavagai problem (describing the get-together of a linguist, a native speaker of some unexplored language, and a rabbit, at whose turn up the native utters the word gavagai), Bloom concludes that “there is an infinity of logically possible meanings for gavagai.“ (Bloom 2001:3). The utterance could refer to any component of this situation, the properties of the rabbit itself, such as its biological categorisation, its colour, texture, or acoustic appearance. Gavagai could even refer to the very spatio-temporal characteristics of the moment itself, to “time slices of rabbits“, as Bloom puts it.

Second, when discussing the phenomenon of fast mapping, Bloom points out that for certain dimensions, such as colour, possibly chiefly those properties of an object are stored linguistically, which escape from human perception and that an object´s perceivable properties are “stored in the world“ (Bloom 2001:34). This is, these properties would have little impact on the internal structure of a word´s meaning because, economically, the relevant information is ubiquitous, as long as people are able to perceive it.

An object´s name and where it comes from, however, are social and historical facts, accessible only through attending to what others say. If you forget this information, you might never encounter it again. (Bloom 2001:34)

At first sight, this point of view seems to heavily contradict a theory of semantic acquisition we will encounter in the course of this paper, namely the semantic feature theory. Again, we will have to keep in mind that we are just dealing with points of view based on data which is per se stored in the world.

The acquisition of word meaning and its internal structure are the topics of this paper. As a matter of methodology, we will focus on language novices, which typically are children up to the age of five, and additionally, we will include another type of language newcomers^[1]: artificial linguistic systems.

In order to put the discussion on a solid foundation, the following two chapters of the present paper will provide both a general and a more specific overview of the topic of semantic acquisition.

In chapter 1.2 we will consider well-established approaches to language acquisition in general. By understanding the nature – nurture controversy as carried out literarily between B.F. Skinner and Noam Chomsky in the late 1950ies, the notion of innateness will be introduced and established as a headterm for further distinctions related to the study of natural language processing. In chapter 1.3 we will investigate the question of whether or not semantic knowledge can be traced back to a set of universal semantic primitives and briefly discuss an approach of identifying them.

Part 2 sets the core of the present paper and introduces the most common views on natural semantic acquisition, namely Rosch´s (1978) prototype theory and Clark´s (1973) semantic feature theory. Additionally, we will consider the functional core concept theory as postulated by Nelson (1974).

Part 3 provides a brief introduction to several recent approaches to semantic acquisition in artificial linguistic systems. In chapter 3.1 we will consider The Talking Heads Experiment based on Luc Steel´s (1999, 2000) works on language evolution with multi-agent systems as well as Yang´s et al. (1998) approach to supervised machine learning and their postulate of case prototypicalities as semantic primitives.

Thus, this paper follows a top-down strategy. We will grope our way from the more general notions to the more specific, and from the more complex systems to the more simple ones. In a concluding discussion, we will summarize the concepts and methodologies we have encountered in the course of this paper with special respect to the linguistic management of colour.

1.2 Approaches to Language Acquisition

In 1957, B.F Skinner published his work Verbal Behavior and presented a theory on language acquisition which was strongly based on the school of behaviorism and the mechanisms of stimulus-response theories. He postulated a working model that suggested the process of operant conditioning to be the driving force of language learning.

From the behaviorists point of view, behavior results from a positive or negative reinforcement of an operant. The operant is defined as a voluntary action carried out by an organism - for example accidently banging one´s head against a doorframe as well as pressing a bar when a light is flashing, or uttering the word ,mama´. Positive reinforcement - a pell of food, or the mother´s glowing eyes - increases the probability of the operant´s reoccurrence, while negative reinforcement – a considerable headache, or even the mother's absence at the very moment of utterance – is likely to decrease this probability.

Skinner did not discriminate between a rat´s and a human infant´s behavior because emphasis lay on external stimuli and outcome, while not observable conditions – such as the organism´s cognitive abilities – were about to be ignored due to methodological creeds. As Elliot (1985) puts it, “the organism was reduced to a black box,...so characterless as to be normally omitted from the equation.“ (Elliot 1985: 29)

Consequently, Skinner extended the observations of his work with rats and pigeons to the study of human language acquisition. He claimed a straight positive correlation between the controlling variables of a particular situation and the predictability of specific utterances resulting from this situation.

Noam Chomsky wrote a discerning and finally taking apart review of Verbal Behavior in 1959. His major criticism on Skinner´s approach was the incomparability of the observed, one-dimensional behavior of rats in boxes with the far more complex properties and less predictable outcomes of human language. Actually, Chomsky believed Skinner to have fundamentally misunderstood the nature of language.

Chomsky and other researchers provide numerous counterarguments to Skinner´s operant conditioning, as discussed in Aitchison (1998). First, the controlling variables are exclusively to be plotted ex post, since even a completely clear situation does not predict a specific utterance. Second, the idea of reinforcement “seems a very wolly notion“ (Aitchison 1998: 9), since Skinner commented on the observation of children frequently talking to themselves by stating that children are in the same way able to reinforce themselves due to their knowledge about their linguistic enviroment. Third, studies in mother-child interaction show that parents tend to negatively reinforce statements which are untrue rather than those which are not grammatical. It is a banality that linguistically competent speakers, however, tend to tell a lie rather than to speak ungrammatically.

Instead, Chomsky proposes a theoretical construct which he calls the language acquisition device, or LAD for short. He argues that although any child has ever been taught the complex rules of his or her mother tongue, he or she is fundamentally able to produce and understand an indefinite number of possible sentences. Chomsky believes the LAD to be an innate capacity for language acquisition which accepts as input the primary linguistic data and pursues the objective of constructing a valid grammar of the language in question. Full linguistic competence and creativity are the outcome of this acquisition process.

Problems arise with the notion of primary linguistic data. The input data of the LAD typically is the data of language performance, a degenerated, socially conditioned reality of the accumulated linguistic competence of the child´s linguistic enviroment, full of unfinished sentences, lemma exchanges and form errors. Nevertheless, children master the formidable task of language acquisition in a considerably short period and a remarkably uniform fashion. According to Chomsky, children know in advance what language looks like because they are genetically equipped with a registered copy of Universal Grammar, a blueprint of highly restrictive universal principles that underlie each human language.

In the previous paragraphs, we have reconstructed a linguistic version of the nature – nuture controversy which actually has not been started by neither Skinner nor Chomsky, but already had been discussed for centuries. But, at all, as Clark and Clark (1977) put it,

It is much more a matter of degree. At the very least, there must be some innate mechanisms that allow one to learn in the first place. At the same time, children will not acquire language unless they are exposed to it. The real question, then, concerns the nature of the innate mechanisms and the constraints they place on what is learned. (Clark 1977: 298)

In order to explain the nature of the innate mechanisms, various proposals have been made. Chomsky´s theory that children are born with language specific knowledge is often referred to as the content approach because it claims that a child´s brain contains innately information about Universal Grammar. Aitchison (1998) discusses two further possibilities.

As a counterpart to the content approach Slobin et al. (Aitchison 1998: 20;136ff) propose that the child's mind is configurated in a predetermined way to analyse the structures of its enviroment including those of human language. This is referred to as the process approach because instead of being equipped with knowledge of Universal Grammar, children are believed to have an innate problem-solving unit which enables them to construct a valid grammar of their mother tongue from the linguistic samples and clues they come across. The process approach comes in a strong and in a weak version. In the strong version, it is claimed that linguistic problem solving is just one task among structuring the external world and is based on the same cognitive abilities which are responsible for general development of intelligence. The weak version postulates that general intelligence may bootstrap a basic scaffolding that relies on semantic knowledge but that the acquisition of a valid grammar is due to a special linguistic problem-solving unit. As Aitchison (1998) points out, “syntax begins when children discover some discrepancy in their semantic scaffolding.“ (Aitchison 1998: 152)

The potential nature of this scaffolding and the quest for its universal components will be the topics of the following chapter.

1.3 The Quest for Semantic Universals

Knowledge about the world is not the same as knowledge about words. As Clark and Clark (1977) argue, people who have no possession of, or no access to their mental lexica through brain damage still have mental encyclopedias. But as a matter of fact, knowledge about the world becomes chiefly observable by the appropriate use of words. Consequently, the structure of word meanings displays very well the structure of world knowledge. The question is whether some concepts of the enyclopedia have to be conceptualised via cognition at all or emerge at some underlying level, for example as a matter of biochemical behavior (see Chapter 3.1).

The German philosopher Gottfried Wilhelm von Leibniz argued that human beings are born with a basic „catalogue of those concepts which can be understood by themselves and by whose combination all our other ideas are formed.“ (Aitchison 1994: 75) But an understanding of entities by themself equals a definition of entities by themself, and as Hofstadter (1991: 137) points out, such a definition would be circular and would lead to an infinite regression and finally to a paradox. A solution to this logical problem provides the concept of a recursive definition that defines an entity by a more simple entity of the same class.

n! = n · (n – 1)!

[...]

^[1] Note that the term 'language' is used to refer to natural language.