[ Konzeptu ]. This review is based on "Philosophy of Language", by Yorik Wilks, in Eugene Charniak and Yorik Wilks (comp.), pp. 205-233. Computational Semantics: An introduction to Artificial Intelligence and Natural Language Comprehension North-Holland, 1976.
Richard Montague and Ludwig Wittgenstein represent views which are diametrically opposed on the key issue of formalization: of whether, and in what ways, formalisms to express the content of natural language (NL) should be constructed. The contrast between their views comes down to two issues:
Two very broad trends in the philosophy of language, as regards the role and importance of formalization can be distinguished:
The "?" indicates a tentative list of philosophers, but one should use it with care. Their views are not comparable in every sense and may differ in many ways, particularly concerning the object of formalization, i.e. language as a whole, or just meaning, etc.
Leibniz, in the 17th century, was not only a highly original formal logician; he also believed that the formalism he proposed was the real structure of ordinary language (and, indeed, of the physical world), though without its awkward ambiguities, vaguenesses, and fallacies. In his most fantastic moments he envisaged the replacement of ordinary language by this "Universal Characteristic", to the general improvement of clarity and precision.
Whitehead and Russell's 1925 Principia Mathematica was a decisive contribution to the definition of propositional and general logic and the investigation of their properties. They were applied to the formalization of the notion of mathematical proof, but already Russel at least was setting out the ways in which this approach to logic was also, for him, a formalization of natural language. Like Leibniz, he wished to clear away what he thought of as the confusions of ordinary language. He was much concerned with the grammatical similarity of sentences like "Tigers are existent" and "Tigers are fierce" and how, in this view, this similarity had led philosophers into the error of thinking that tigers therefore had a property of existence as well as one of fierceness.
Wittgenstein was closely associated with Russell during the period between 1910 and 1920, and was developing what is now thought of as his early philosophy. This was set out in a curious early work called the Tractatus Logico-Philosophicus (1922) where he proposed what is now called the "picture theory of truth", according to which sentences in ordinary language signify or mean because their structures reflect or exhibit the same relation as that which holds between the things mentioned by the sentences. His theory of meaning at this point was more obscure than that of Russell, or Russell's predecessor, Frege.
Frege, in the late 19th century, had proposed a "dualist" theory of meaning in which each word signified in two ways (his famous example of "the Morning Star" and "the Evening Star"):
These meaning theories are referential, even when they talk of "sense". For, whatever the status or nature of the thing that is "the meaning", it is an entity that is somehow pointed at by the word. In Frege's case the word points in two different ways to two quite different sorts of entity, whereas in Wittgenstein more obscure "picture theory" there was only one kind of pointing.
[The truth of expressions can be established by truth tables and is demostrated as semantic. This is different from the method referred as proof theoretic (that comes from the sequential relation of structures in a proof)]. The word "semantic" is so troublesome that it cannot be introduced into the discussion without a word of warning, because on its ambiguities rests much of the difficulty of our whole subject.[...] It will be written as "L-semantic" when used in connection with formal logic.
Formal logic of the Twenties rests in two fundamental ideas, namely:
Montague and Wittgenstein represent respectively a reaction to, and a extension of that logic. In the hands of Tarski'35 and Carnap'37 the above concepts took on new names and properties:
Carnap in his "Logical Syntax of Language" developed the notion of ill formed expression (as "Mortality is Socrates"). Carnap distinguished: (Chomsky was many years later a student of Carnap)
Carnap was a formalizer in the Leibnizian sense since, to him, these distinctios applied equally well to proper formalization of natural language.
The distinction of language and meta-language was due to Tarski who proposed it to solve an apparently intractable problem of logic that he had inherited from Russell, who had discovered certain logical paradoxes. Tarski's statement of the problem of the paradoxes was usually in terms of the example:
| The sentence in this rectangle is false |
where any attempts to assign a truth-value to the sentence led to trouble. Tarski thought that this problem would be solved if we only used "true" in a meta-language and never in an object -language. Thus "That John is happy is true" would be a sentence with level confusion, for the proper name would be the meta-language sentence "'John is happy' is true", where the sentence "John is happy" is mentioned in the meta-language, but not used in it.
Tarski's fundamental achievement was a theory of truth and logical consequence for formalized languages (a L-semantic theory).
Montague extended it to sentences in natural languages: a system in which the truth conditions of complex statements could be seen to be composed from the truth conditions of simple ones. Making even stronger claims than Tarski's, because Tarski did not think that such a theory could be constructed for a natural, non-formal, language. Otherwise the object-language and the meta-language would be the same, and Tarski thought that was bound to lead to trouble.
Wilks'76 and others in linguistics and AI do not share the assumption that the stating of truth conditions is explaining the meaning of a sentence. Those in the Tarskian tradition are highly critical of much work in linguistics and AI because meanings are not stated in that way, and it is an interesting open question whether AI work on NL could be so stated in general. Wilks does not criticize formal analysis of NL as such. What is being opposed is pointless formalizations of NL [such as Montague's], and it is suggested that Wittgenstein is on of the few philosophers who can provide insights into what a fruitful approach to NL migh be.
Logical positivists also had a thesis about the dependence of meaning on truth but should not be confused with the "truth-condition" approach followed by Tarski. Their principle was called the Principle of Verification, and it said "The meaning of a statement is the procedures we would carry out to establish its truth of falsehood". However wrong this principle might be (and it is wrong), it was at least serious, in a way that the modern truth condition approach is not.
Blackburn'9? (for one in a large school of logicians) holds general principles concerning the applicability of Tarski-like theories to NL. The thesis that they share is that of the "truth theory of meaning": namely that the meaning of a sentence is determined by its truth conditions, in the Tarskian sense of that phrase. It would be quite possible to reject Montague's detailed "Semantics for NL" and still accept the general tenents of the school about meaning and truth, namely that truth conditions determine the meaning of the senteces of NL in just the way they can be said to do so for the sentences of logic and methematics.
Dreyfus, a full-time opponent of the very possibility of AI, has made much use of Wittgenstein's arguments against formal logic in his own arguments against AI. However, the general structure of this chapter is that AI has real philosophical importance because it has completely changed the old debate between formalists and anti-formalists. For to handle language "formally" on a computer it is in no way necessary to accept the tenets of Tarski, Montague or any other approach based on formal logic. On the contrary, the most fruitful approaches to understanding natural language are precisely those not subservient to a powerful logical or L-semantic theory.
AI has provided a sense to the notion of the precise manipulation of language that is not necessarily open to the attacks of the anti-formalists like Wittgenstein. And a large part of the credit for breaking the old formalist/anti-formalist opposition in a new way must go to Chomsky.
Chomsky's theory of grammar is a precise theory of language: it has the form of a logic, but not the content. Chomsky took the structure of proof-theoretic (what Carnap called transformation rules), namely the repeated derivation of structures from other structures by means of rules of inference, but he let its content be no more than what Carnap had meant by Formation Rules, namely the separation of the well-formed from the ill-formed. Thus, in transformational grammar the inference (transformational) rules were to apply to axioms and theorems, but to produce not new theorems but well-formed English sentences.
Thus, Chomsky has a precise system of handling language, but with no semantic definition of truth, and not even a syntactic, proof-theoretic one either. Thus, Chomsky's was the first concrete proposal to breach the wall between formal and anti-formal approaches.
AI has gone further, and indeed Chomsky's paradigm still shares many of the drawbacks to formal approaches: in particular the rigid "derivational" structure common to logical proofs and to Chomskyan transformational derivations.
[ Montague'70, Wittgenstein'58 ]
Those appalled by the generality of the title of this chapter will be relieved to hear right away that it attempts no more than a brisk introduction followed by some detailed, though not very detailed, discussion of only two philosophers: Richard Montague and Ludwig Wittgenstein. These philosophers have been chosen not so much for their influence on our subject matter, which has been small, but because their views are diametrically opposed on the key issue of formalization: of whether, and in what ways , we should construct formalisms to express the content of natural language. I shall argue, too, that the influence of Wittgenstein has been largely beneficial while that of Montague has been largely malign. Much of what follows will be a justification of that rather sweeping judgment. It should be said that those who find even the simplified formalism of the Montague section too strenuous may turn directly to the Wittgenstein since they are almost entirely independent of each other.
Concentrating on two philosophers in this way means that a number of the names of the great and the good, in what is normally thought of as the philosophy of language, will not appear. To attempt to say everything is of course to say nothing. Even so, we can distinguish two very broad trends in the philosophy of language, as regards the role and importance of formalization. Broadly speaking, one group of philosophers has been for it, and for as much of it as possible, while the other group has been uncompromisigly against it. With a little stretching of the imagination one can reach back and assign even the Greek philosophers to one or other of these groups. It is clear, for example, that Aristotle was considerably more preoccupied by logic, in its conventional sense, than was Plato.
But it is with Leibniz, in the 17th century, that the positive attitude to formalization suddently appears in full bloom. Leibniz was not only a highly original formal logician; he also believed that the formalism he proposed was the real structure of ordinary language (and, indeed, of the physical world), though without its awkward ambiguities, vaguenesses, and fallacies. In his most fantastic moments he envisaged the replacement of ordinary language by this "Universal Characteristic", to the general improvement of clarity and precision. He went further: "For once missionaries are able to introduce this universal language, then will also the true religion, which stands in intimate harmony with reason, be established, and there will be as little reason to fear any apostasy in the future as to fear the renunciation of arithmetic and geometry once they have been learnt". You will see already that the formalist attitude is not necessarily a dull and small-minded one!
By and large, the two centuries that followed saw this position sink almost without trace. The important change, for our purposes, came with the rise of formal logic at the turn of the century. It began with the definition of propositional and general logic and the investigation of their properties. The earliest account of these calculi in English is Whitehead and Russels' Principia Mathematica in which they were applied to the formalization of the notion of mathematical proof, but already Russel at least was setting out the ways in which this approach to logic was also, for him, a formalization of natural language. Like Leibniz, he wished to clear away what he thought of as the confusions of ordinary language. He was much concerned with the grammatical similarity of sentences like "Tigers are existent" and "Tigers are fierce" and how, in this view, this similarity had led philosophers into the error of thinking that tigers therefore had a property of existence as well as one of fierceness.
In the First Order Predicate Calculus, as it is now called, the first sentence might go into some form such as:
EXISTS(X) (TIGER(X))
while the second might go into some form such as:
FOR-ALL(X) (TIGER(X) IMPLIES FIERCE(X))
The important thing here (and there are many alternative forms for these sentence codings) is that in none of them is there any predicate symbol for "exists", in the way there is for tigerness or fierceness. Or, to put it another way, the assertions of existence are always in the quantifiers of the expression. And so, in the Predicate Calculus, the similarity of form between the two English sentences completely disappears.
Russel was not philosophically neutral about all this, for he believed that serious intellectual errors had followed from what he believed to be the confusion of two logical forms caused by their having the same grammatical form. One classical argument about God's existence, for example, centered on the question as to whether a perfect being (i.e. God) had to have the property of existence if he was to be perfect. In Russell's view one could not reasonably talk about the "property of existence" at all once one had seen that the two forms of sentence above did not both translate into "property forms" in his logic.
Wittgenstein was closely associated with Russell during the period between 1910 and 1920, and was developing what is now thought of as his early philosophy. This was set out in a curious early work called the Tractatus Logico-Philosophicus, which wll not be discussed in any detail here, for we shall be concentrating on his late philosophy. In that early work he proposed what is now called the "picture theory of truth", according to which sentences in ordinary language signify or mean because their structures reflect or exhibit the same relation as that which holds between the things mentioned by the sentences. Thus, a logical form of fact like "catONmat" would be true if the relation between the entity symbols "cat" and "mat", and the relational symbol "ON", reflected the relation between the appropriate entities in the world. The problems for Wittgenstein's commentators (including himself) have always been (a) about what "reflected" could mean there, which might be clear if the relational symbol had been LEFTOF, and "cat" was to the left of "mat" on the page so that the sentence would be true if the real cat was indeed to the left of the real mat and "catLEFTOFmat" would indeed be a sort of picture of the fact described. However, none of the above was so clear if the relations were more complex and realistic such as MOTHEROF. The second difficulty (b) has been about what the "entities in the world" were that the symbols referred to. It seemed clear that Wittgenstein did not mean the real objects out there in the world, like the cat and the mat themselves. He used the German word "Gegenstände" (which sometimes means "objects"), yet no one has ever been quite clear what these entities were to be. His theory of meaning at this point was more obscure than that of Russell, or Russell's predecessor, Frege.
In the late 19th century Frege had proposed a "dualist" theory of meaning in which each word signified, if it did signify, in two ways: one way (Bedeutung), referred to some entity, and one (Sinn) referred to the "sense" of the word. The details of this distinction have preoccupied philosohers ever since, but the broad outline is clear. His famous example of "the Morning Star" and "the Evening Star" is still the best illustration: those pharses mean something different in that they have different SENSES; however it is also true that they refer to the same heavenly entity or REFERENT (i.e. Venus). These meaning theories need not detain us, but it should be noted that they are all in some way referential, even when they talk of "sense". For, whatever the status or nature of the thing that is "the meaning", it is an entity that is somehow pointed at by the word. In Frege's case the word points in two different ways to two quite different sorts of entity, whereas in Wittgenstein more obscure "picture theory" there was only one kind of pointing.
Yet two ideas surfaced in Wittgenstein's early work that were to be very important in the logic of the Twenties: first the notion of significant and insignificant combination. For him the symbols for the Gegenstände (like "cat") could only be substituted into the pictorial forms of fact (like "catONmat") in certain ways and not others. That is to say "Socrates is Mortal" reflected some relation of entities in the real world, but "Mortality is Socrates" did not, and not because those entities were not in that relation, but because the symbols "Mortality" and "Socrates" could not be substituted in that particular representation of fact at all since the combination made no sense.
With this theory of "picture forms of fact" Wittgenstein was also putting forward the idea, even though hazily, that a theory of meaning required a theory of truth, because the way in which the combination of symbols reflected, or failed to reflect, a fact (i.e. was true or false) was the same thing, in some sense, as the way the statement made sense. It made sense only insofar as it reflected (was true) or failed to reflect (was false) a fact. In particular, he developed a very elementary theory of truth for the Propositional Calculus, a method called "truth tables", discovered independently at the same time by C. S. Pierce in the United States.
This notion is important for what follows, so let us just set it out quickly here: the Propositional Calculus contains variables like p, q, etc. that stand for the proposition expressed by any simple sentence such as "John is happy". These simple propositions can be made into more complex ones by means of connectives NOT, AND, OR and IMPLIES. [...]
[Thus, the truth of expressions can be established by truth tables and is demostrated as semantic. This is different from the method referred as proof theoretic (that comes from the sequential relation of structures in a proof)]. The word "semantic" is so troublesome that it cannot be introduced into the discussion without a word of warning, because on its ambiguities rests much of the difficulty of our whole subject.[...] It will be written as "L-semantic" when used in connexion with formal logic.
We have introduced two of the fundamental ideas on which the formal logic of the Twenties rests, namely, significant combination and L-semantic demostration. That logic is the heart of the background to this chapter, because the two authors represent respectively a reaction to, and a extension of that logic. In the hands of Tarski and Carnap the above concepts took on new names and properties: "logical syntax" (replacing "L-semantic demonstration"), to which was added a third concept "meta-language".
Carnap in his "Logical Syntax of Language" developed the notion of "ill formed expression" that we met earlier in discussing "Mortality is Socrates". Carnap distinguished, for any logic its rules of formation and rules of transformation (those who think this is beginning to sound familiar should remember that Chomsky was many years later a student of Carnap). The rules of formation determined what were, and were not, well formed expressions in the logical language. Thus, in the Propositional Calculus, (P IMPLIES (OR Q)) was not, while (P OR Q) was a well formed expression. The rules of transformation then operated on those well formed expressions that were also thrue so as to produce theorems in logic. Carnap was a formalizer in the Leibnizian sense since, to him, these distinctios applied equally well to proper formalization of natural language. Carnap constructed such a formalization in which he distinguished two types of what he called "pseudo-statement":
(1) Caesar is and
(2) Caesar is prime number
The first was counter-syntactic; the second violated the rules of logica syntax. In his system he also made use of a distinction between an object language and a meta-language. [...]
Carnap was far from dispassionate in all this, in that his purpose was what Russell's had been, namely to do away with certain kinds of sentence, and particularly those that arose in certain kinds of philosophical writing, by showing that they violated the rules of logical syntax.
The distinction of language and meta-language did not arise from any considerations about the structure of natural language, however, but was due to Tarski who proposed it to solve an apparently intractable problem of logic that he had inherited from Russell, who had discovered certain logical paradoxes. Tarski's statement of the problem of the paradoxes was usually in terms of the example:
| The sentence in this rectangle is false |
where any attempts to assign a truth-value to the sentence led to trouble. Tarski thought that this problem would be solved if we only used "true" in a meta-language and never in an object -language. Thus "That John is happy is true" would be a sentence with level confusion, for the proper name would be the meta-language sentence "'John is happy' is true", where the sentence "John is happy" is mentioned in the meta-language, but not used in it.
Tarski's fundamental achievement was a theory of truth and logical consequence for formalized languages, or what, using our convention, we are calling a L-semantic theory.[...]
One might say, at the risk of enormous simplification, that Tarki's theory of consequence and truth is a systematic generalization of this method of replacement of Leibniz's slogan "Logical truth is truth in all possible worlds" and of truth-table notion that the logical truth of compound expressions is to be settled in terms, and only in terms, of the truth-conditions of the simpler propositions of which they are constructed. "Possible world" is not meant technically in that slogan, and its more modern meaning will be explained later.
[...] The truth conditions of a sentence are the conditions under which it is true. The truth definition of a sentence is the specification of its truth conditions.[...]
It is difficult to give a simplified and brief account of how Traski constructed a recursive definition of truth for formal languges; one from which the truth definitions of individual sentences could be calculated, rather tan being simply listed in advance.
The key feature of such a definition, though, is that it has the form:
DEF: "A sentence is true if and only if it has property X"
where X is some complicated property (in fact involving a technical notion of "satisfaction by all sequences of objects") that can be calculated. [...]
Montague's work is itself intended to yield such recursive truth definitions for sentences in natural languages: a system in which the truth conditions of complex statements could be seen to be composed from the truth conditions of simple ones [...].
Montague's work is thus making even stronger claims than Tarski's, because Tarski did not think that such a theory could be constructed for a natural, non-formal, language.
One reason why Tarski did not think such a theory applicable to any single natural language was because that would mean [...] the object-language and the meta-language would be the same, and Tarski thought that was bound to lead to trouble.
One assumption behind all of this work is that the stating of truth conditions is explaining the meaning of a sentence. This is an assumption not shared by much of the semantic work [in Charniak and Wilks 1976]; very little of it explained meaning as what would be true if such and such was true. Those in the Tarskian tradition are highly critical of much work in linguistics and AI because meanings are not stated in that way, and it is an interesting open question whether AI work on NL could be so stated in general.
This "truth-condition" approach following Tarski should not be confused with the claims of the logical positivists. They also had a thesis about the dependence of meaning on truth [...]. Their principle was called the Principle of Verification, and it said "The meaning of a statement is the procedures we would carry out to establish its truth of falsehood". [...] However wrong this principle might be (and it is wrong), it was at least serious, in a way that the modern truth condition approach is not.
[...] There is a school of logicians sharing general principles concerning the applicability of Tarski-like theories to NL [...]. The thesis that they share is that of the "truth theory of meaning": namely that the meaning of a sentence is determined by its truth conditions, in the Tarskian sense of that phrase. It would be quite possible to reject Montague's detailed "Semantics for NL" and still accept the general tenents of the school about meaning and truth, namely that truth conditions determine the meaning of the senteces of NL in just the way they can be said to do so for the sentences of logic and methematics.
[...Montague shifted] to the notion of "truth relative to a model" interpretation or possible world as the standard sense of "true", and not that defined in DEF above. This involves giving up Tarski's convention that there was a basic sense of truth independent of particular models or interpretations, and sticking only to the notion of "truth within a model" or interpretation, or what is usually called a relativistic notion of truth. [...]
Formal analysis of NL is not being criticized as such. What is being opposed is pointless formalizations of NL [such as Montague's], and it is suggested that Wittgenstein is on of the few philosophers who can provide insights into what a fruitful approach to NL migh be.
We must tread carefully here because Wittgenstein was a foe of all attempts to apply formal logic to the analysis and understanding of NL.
Moreover, the philosopher Dreyfus, a full-time opponent of the very possibility of AI, has made much use of Wittgenstein's arguments against formal logic in his own arguments against AI. However, the general structure of this chapter is that AI has real philosophical importance because it has completely changed the old debate between formalists and anti-formalists. For to handle language "formally" on a computer it is in no way necessary to accept the tenets of Tarski, Montague or any other approach based on formal logic. On the contrary, the most fruitful approaches to understanding natural language are precisely thos not subservient to a powerful logical or L-semantic theory.
Thus it is that AI has provided a sense to the notion of the precise manipulation of language that is not necessarily open to the attacks of the anti-formalists like Wittgenstein. And a large part of the credit for breaking the old formalist/anti-formalist opposition in a new way must go to Chomsky.
Chomsky's theory of grammar is a precise theory of language: it has the form of a logic, but not the content. Chomsky took the structure of proof-theoretic (what Carnap called transformation rules), namely the repeated derivation of structures from other structures by means of rules of inference, but he let its content be no more than what Carnap had meant by Formation Rules, namely the separation of the well-formed from the ill-formed. Thus, in transformational grammar the inference (transformational) rules were to apply to axioms and theorems, but to produce not new theorems but well-formed English sentences.
Thus, Chomsky has a precise system of handling language, but with no semantic definition of truth, and not even a syntactic, proof-theoretic one either. Thus, Chomsky's was the first concrete proposal to breach the wall between formal and anti-formal approaches [...]. AI has gone further, and indeed Chomsky's paradigm still shares many of the drawbacks to formal approaches: in particular the rigid "derivational" structure common to logical proofs and to Chomskyan transformational derivations.
If Chomsky's grammar could be seen as a move to preserve the advantages of formalization without its "logicism", Montague saw his task as reversing that move; he intended to tacke formalization of NL in a more serious way.
Montague's approach can be illustrated by the way he analyzes a sentence like "Every man loves some woman", with two readins to any logician, which cannot be adequately treated by transformational grammar:
(i) for every man there is some woman that he loves.
(ii) there is some woman such that every man loves her.
[...] The basic idea is to construct a L-semantical object that states what it is, in set-theoretic terms, for the corresponding reading of some sentence to be true. And the production of the whole statement "Every man loves some woman" is true if and only if... (the L-semantic object's content) is precisely the Tarskian goal that Montague set out to reach for a NL: that is, a whole theory, functioning as a Tarskian definition of truth, that produces as non-trivial consequences the truth-definitions of individual sentences of the language.
[...] What can one say in general about an L-semantics of this sort?
1) Syntax is arbitrary and unmotivated.
2) The L-semantics is entirely a reflexion of the sintax and the two could not in principle diverge. This seems extraordinary and very implausible.
3) The assumption at every stage is that there is a molecular confrontation between language and the world. [...] Montague's assumption is far from any AI view of meaning on which we cannot talk about meaning independently of large structures of knowledge existing outside teh sentence examined. There is no place for that in Montague's system because meaning is to be built up only from very simple L-semantical objetcs, attached to the items of the sentence direcly.
4) The notion of "truth condition" in a clear sense is not computable.
A cynic might say that, whatever the value of L-semantics as a subsequent axiomatization, or reconstruction, of linguistic computations, it could never be a research tool; one in which important contenful linguistic rules were established initially. In the same way, science is never done by thinking about the axioms of formal scientific systems. The notions of L-semantics are all mathematical notions and belong there. In the world of NL, they are, in Ryle's phrase, "on holiday" and cannot be expected to earn their keep.
[...] Much of the motivation of the Philosophical Investigations was to set out why [the view of the "picture theory of truth"] and its associated doctrines were wrong. Also in the background are Tarski and Carnap, who still advocated formalist views long after Wittgenstein had given them up. Most of the views attacked in Philosophical Investigations were held in one form or another by Tarski. The change in Wittgenstein was that he ceased to believed that words in NL had meaning chiefly because they pointed at objects in the world, and that sentences were true because they matched up to the world in some direct one-to-one way. He became more and more convinced that what was important about language was its "deep grammatical forms", and it was from here that the metaphor of "depth" in modern linguistics took off. [...]
(i) Reference
Thesis: words do not in general have meaning in virtue of pointing at objects in the real world (or "conceptual objects" either).
Pointing or referring is in principle a vague activity. It can only be made clear by explaining from within the language what we are pointing at, i.e. pointing already assumes the whole language. Hence it is not that pointing explains how we mean, as the formalists thought when they defined the denotations of their symbols as objects, or sets of objects, because the pointing presumes upon the language rather than explains it.
Wittgenstein says we could have a language based on the referential notion, but it would be a language more primitive than what we call NL.
(ii) Mini languages and language games
Thesis: we can construct mini-languages obeying any rules we like, and we can think of them as games. The important question is whether these games are sufficiently like the "whole game" of NL.
W. presents the danger of taking a mini-language and assuming that their properties hold for the whole of NL. [...] The languages of semantic primitives in Schank's or Wilks' systems are also mini-languages in a wide sense. There would be similar problems if they started to postulate "conceptual objects" to which the primitives refere (and Schank has in fact sometimes suggested that his primitives "really" refer to entities in the mind or brain). If a language of primitives is given that property then it loses one essential feature of full NL, and begins to look more like [Winograd's] "blocks world" mini-language.
However, it should not be though that Wittgenstein is a defender of linguistic primitives, or primitives of any sort. Indeed, one of the attractions to him of his "truth-table" method of presenting Propositional Calculus was that it avoided the more conventional form in terms of primitive formulas from which all other true formulas could be derived. Yet, one could argue that a large part of what W. found objectionable about the notion of "primitive" in logic was the idea that there is a right set of them, if only we could discover it, to provide an infallible starting point. But in the case of semantic primitives, it is still possible to use them, as Wilks does, without claiming that there is a single right set of them, as Schank does.
(iii) Family resemblances and boundaries
Thesis: the conventional notion of concept is wrong: namely the view that a concept relates in some way to the qualities or characteristics that all things falling under the concept have. W. takes the concept of "game" and argues that one could not define a game by necessary and sufficient qualities. For any proposed necessary characteristic of being a game, W. claims he can think of a game that does not have the characteristic. Along these lines it has been argued that the entities under a concept form a something more like a family, for some members of a family share the characteristic "long nose", while some share the characteristic "tiny feet", but there need be no characteristic they all share for them to have a "family resemblance". The moral is that there are not firm boundaries to concepts, nor are there to linguistic usage, nor to the application of linguistic rules.
NL has an essential vagueness in its everyday employment, one that does not harm, and which in no way interferes with our understanding each other. It is, however, a feature that no one at the moment has any idea how to put into a computational system.
226: A feature a satisfactory language system would have to embody is the ability to redraw its own boundaries, and to be as vague or specific as necessary, in the way that we can.
These considerations might suggest that a theory of language can never be, in a straightforward way, a scientific theory, that is to say a body of rules that decides, rightly or wrongly, of any utterance, what its correct structure is. But that is another and very difficult discussion in its own right. A fuller discussion of $133 might lead us to question the whole possibility and usefulness of "ingenious knockdown counter-examples" to theories of language.
(iv) The linguistic whole and confronting the world
Thesis: a language is a whole and does not confront the world sentence by sentence for the testing of the truth or falsity of each individual part.
(v) Logicians have a false picture of how language is
Thesis: logicians think that language is like their favorite calculus, but they are quite wrong. Moreover, it is language itself and its use that is the standard for testing disputes that arise, not what logicians dictate.
This thesis clearly clashes head on not only with Montague but also with those logicians who believe in the applicability of the Predicate Calculus to language.
(vi) Understanding is not a feeling
Thesis: we have the idea that "understanding" something involves, or is associated with, a special feeling of being right. But the tests of our being right are quite different from the feeling.
W. is making the point that it is dangerous to assess understanding other than in terms of actual and possible performances, and, if we take that to mean "performances with language" we will see that it argues against one sort of criticism that AI workers have some times made of each other's systems: that they only "appear to understand" but "didn't really do so". Those who employ that sort of criticism are, in W. terms, acting as if understanding is a special feeling.
Dreyfus has argued that AI is impossible because to be intelligent a computer would have to be exactly like us: bodies, feelings and all. He has often quoted W. in support of his own position, but it can equally well be argued that W. clear distinction between understanding-as-performance (which AI workers believe a machine can have) and understanding-as-feeling (which no doubt only we have) supports exactly the opposite position.
(vii) Application justifies our structure
Thesis: the significance of a representational structure cannot be divorced from the process of its application to actual language.
Our representational structures cannot be laid direct against the world they represent because they are not "in the same space", that is to say the representation is in symbols and the world is in chairs, tables, ocean and government buildings out there. Those who advocate Montague-type systems do seem to believe that in some way their representations are "laid against the world" directly via the L-semantic notions of "set" and "function".
I follows that to be in the "same space" as what they measure, AI representations for language must preserve the essential features of NL.
(viii) Real world knowledge and forms of life
Thesis: language understanding is not independent of very general inductive truths about our human experience
[ Konzeptu ]
Introduction |
Introducción |
|
This paper addresses the question of whether it is possible to sense-tag systematically, and on a large scale, and how we should assess progress so far. That is to say, how to attach each occurrence of a word in a context to one and only one sense in a dictionary - a particular dictionary of course, and that is part of the problem. The paper does not propose a solution to the question, though we have reported empirical findings elsewhere (Cowie et al. 1992 and Wilks et al. 1996), and intend to continue and refine that work. The point of this paper is to examine two well-known contributions critically, one (Kilgarriff 1993) which is widely taken as showing that the task, as defined, cannot be carried out systematically by humans, and second (Yarowsky 1995) which claims strikingly good results at doing exactly that. |
Esta ponencia toca la cuestión de si es posible etiquetar de manera sistemática, y a gran escala, los sentidos de las palabras, al tiempo que trata de evaluar los avances realizados hasta la fecha. Se trata de averiguar si es posible asignar a cada aparición de una palabra en un contexto una y solo una acepción en un diccionario - por supuesto, un diccionario muy particular, lo cual constituye parte del problema. Esta comunicación no popone soluciones a la cuestión , aunque con anterioridad ya hemos ofrecido resultados empíricos (Cowie et al. 1992 and Wilks et al. 1996), y pretendemos seguir y mejorar ese trabajo. El motivo de esta ponencia es examinar con criterio crítico dos contribuciones bien conocidas, la primera (Kilgarriff 1993) suele tomarse como prueba de que la tarea que hemos definido no puede ser realizada de manera sistemática por personas; la segunda (Yarowsky 1995) sorprendentemente proclama buenos resultados haciendo exactamente eso. |
|
|
Empirical, corpus based, computational linguistics has reached by now into almost every crevice of the subject, and perhaps pragmatics will soon succumb. Semantics, if we may assume the sense-tagging task is semantic, taken broadly, has shown striking progress in the last five years and, in Yarowsky's most recent work (1995) has produced very high levels of success in the 90%s, well above the key bench-mark figure of 62% correct sense assignment, achieved at an informal experiment in New Mexico about 1990, in which each word was assigned its first sense listed in Longman's Dictionary of Contemporary English (LDOCE). |
La lingüística computacional empírica, basada en corpus, ha llegado en la actualidad a casi todo **crevice de la materia, y tal vez la pragmática sucumba pronto. La semántica, si aceptamos que el etiquetado de sentidos es una tarea semántica, ha demostrado un progreso sorprendente en el último lustro y, en el trabajo más reciente de Yarowsky (1995), ha producido unos niveles de éxito del 90%, muy por encima de la cifra experimental clave del 62% conseguida de manera informal en Nuveo México, en torno a 1990, en la que cada palabra era asociada con la primera acepción de los artículos del Longman's Dictionary of Contemporary English (LDOCE). |
|