Send us a comment.
Language understanding involves relating linguistic forms to meanings; language generation involves the opposite. We can certainly represent linguistic forms in the computer, but what about meanings? A computer manipulates only formal symbols. How can we say that one symbolic structure correctly represents the meaning of a sentence, whereas another does not?
There have been many different ideas about how meaning can be represented in humans and machines. One possibility involved considering the meaning of an utterance to be a procedure; that is, a set of instructions to achieve what the speaker wants. Such an idea fitted in naturally with the AI concept of the mind as a computational device following rules and instructions of some kind. The competent hearer would be able to construct this procedure and could then decide whether or not to actually run it.
According to the simplest versions of what came to be known as 'procedural semantics', the meaning of a command is a procedure to carry out the required action, the meaning of a question is a procedure to find the answer, the meaning of a statement is a procedure to add the new information conveyed to the hearer's model of the world, and so on. However, there are profound philosophical problems with any simple form of procedural semantics, not the least of which is the problem of defining exactly what is, and what is not, a procedure. If a procedure is a representation that can only be run, and whose internal structure cannot be analyzed in any way, then procedural semantics alone cannot account for the fact that a hearer may refuse to obey a command, because it seems impossible or dangerous. On the other hand, if the internal structure of a procedure is something that can be reasoned about, then it is not clear what distinguishes a procedural representation from any other representation that can sometimes be interpreted as a set of instructions.
Given the difficulty of accounting for all possible language uses in the same formalism, most logicians and AI workers have concentrated on representing the meanings of simple statements about the world. It is at least arguable that other uses of language (such as asking questions or issuing commands) involve statement meanings as a component of their meaning. The problem of representing the meanings of statements about the world merges into that of knowledge representation in general, and this is a central concern for AI research.
Some early formalisms for knowledge representation, based on the idea of networks, were motivated by intuitive psychological considerations. Intuitively, it seems, we think of the world in terms of concepts; when we hear something new, we may discover new concepts, or we may discover new relationships between existing concepts. Thus, we can think of representing knowledge in terms of a graph, where the nodes represent the concepts and the links between them represent the relationships that are known. Such 'semantic net' systems, as they are known, are designed to facilitate certain restricted kinds of inference, such as inheritance of properties from node to node (see Chapter 9 for further discussion).
One network-based notation that has provoked a lot of attention is Schank's Conceptual Dependency. This was explicitly intended as a psychological model of how people represent the meaning of sentences. In 1963, the philosophers Katz and Fodor had promulgated the idea of having a fixed set of primitives out of which all meanings were to be constructed. Schank continued in this tradition by proposing a set of (about) 11 primitive actions - for instance, PROPEL (apply a force to) and MTRANS (mental transfer of information). The idea was that these could be combined to express any event in the world. Conceptual dependency was designed to be independent of its use with any particular natural language, to allow it to act, for example, as an 'interlingua' for translating between languages. Associated with each of Schank's primitive actions was a set of underlying argument positions which could be filled differently for each instance of the action. For example, every time an instance of PROPEL is represented, there has to be an 'actor' (the performer of the action), an 'object' (the thing the force acts on) and a 'direction' (where the activity is coming from and going to).
Nowadays, many of the ideas in the early network-based systems have been subsumed in systems with more structure. Unadorned semantic networks have no way of expressing the fact that certain concepts and relationships should be grouped together into larger chunks. Thus, they cannot explain how a participant might 'focus' on a particular group of things in a discourse or how a reader might have sensible expectations about what is coming next, after recognizing a familiar situation. Some of the systems that have emerged in response to this need are partitioned semantic networks, the knowledge representation language KL-ONE and various systems based on Minsky's 'frames'. One of the latter, scripts, will be discussed in Section 1.4.
The network-based systems of the 1970s have been criticized by a number of researchers for being semantically ill defined and for failing to capture certain distinctions that critics have taken to be important - for example, the distinction between type and token, or that between the membership and subset relations. Indeed, until recently, little effort was spent on producing formal theories of what these networks actually meant and how exactly we should go about expressing any given knowledge in them. A rival class of knowledge representation languages, which has received a great deal of formal attention of this kind, is that of logics - in particular, first-order logic (see Chapter 9).
First-order logic has been used as a representation language in AI right from the beginnings of the subject. Thus, Woods's LUNAR program for answering questions about lunar rock samples, which was produced at roughly the same time as SHRDLU, used a logical representation for the meanings of questions. The current interest in logic programming systems is just one symptom of the growing realization of the importance of logic. Indeed, some researchers argue that many existing network notations, insofar as they are well defined, are equivalent to, or weaker than, first-order logic, and that we should therefore dispense with them. On the other hand, others argue that logic is too neutral a representation language, not tuned to the kinds of meanings that natural language conveys.
The idea of representing natural language meanings in logic is one aspect of a vein of 'logicism' that is present in much of the recent work on NLP. The aim of this enterprise is to produce a formal account of how natural language conveys meaning by drawing on work in formal semantics and the philosophy of language. Almost all computer models of language understanding make use of some concepts from this work - for example, the idea that meaning can be obtained compositionally, an idea that goes back to the work of Frege in the 1890s.
The compositionality principle, also known as Frege's principle, maintains that the meaning of any phrase can be obtained by some operation on the meanings of its parts. So, given a structural description of a sentence, it is possible to work out what the sentence means by first finding the meanings of the individual words, then combining these together to construct the meanings of small phrases, then combining these to construct the meanings of larger phrases, and so on until the meaning of the whole sentence is formed.
There is no necessity for things to work this way: we can readily imagine (artificial) languages in which, say, the meaning of an expression is determined, in part, by its linear position in the string - for example, a language in which 'one hundred and seven' denotes 93 if it occurs sentence-initially, 107 if it occurs sentence-medially, and 700 if it occurs sentence-finally. The semantics of such languages would not be stateable compositionally - at least, not in any straightforward way. The work of the philosophical logician Montague is taken by many as an important demonstration of the utility of the compositionality principle in the analysis of natural language. Montague showed that many previously problematic, semantic phenomena in natural language were amenable to formal treatment within a strictly compositional regime. Montague's account was not computational in form and left important computational issues, such as the role of inference and the representation of word meanings, untouched. Nevertheless, his methodology has proved an inspiration to many workers in NLP.
As we have seen, a processing account of language comprehension must provide some explanation of how ambiguity is resolved. This involves specifying extra mechanisms for filtering out syntactic or semantic analyzes that are inappropriate. Many computer programs for manipulating natural language have made use of selectional restrictions, originally proposed by Katz and Fodor in the early 1960s. Selectional restrictions were introduced as a way of accounting for the fact that, for instance, a sentence as a whole can be unambiguous even though individual words may have several alternative possible senses. The theory was based on the idea that with each sense of a word it is possible to associate semantic markers, specifying features of the meaning as well as conditions on the features of word senses that could combine with it. So, for instance, one possible sense of the word 'spirit' (alcoholic fluid) would have the marker 'physical object', whereas another sense (vital principle) would not. The adjective 'yellow', when describing a colour, requires that the noun it qualifies has the marker 'physical object'. Another sense of 'yellow' (cowardly) requires the noun to have the marker 'animate', but neither of the senses for 'spirit' considered here has this marker. Thus, the phrase 'yellow spirit' can be shown quite formally to have only one possible sense here (yellow-coloured alcoholic fluid), rather than four - to keep things simple, we ignore the other possible senses of 'yellow' and 'spirit').
Within a processing framework, it is possible to extend the basic Katz and Fodor ideas in a number of ways. For instance, it is possible to ensure that semantic marker analysis causes a syntactic analysis to be rejected if no corresponding semantic readings can be found. The semantic tests can be applied to the referents of phrases rather than simply the word senses. And a notion of 'preference' can be introduced and dispreferred readings allowed to succeed when no highly ranked reading is available (cf., 'Cowardice is a yellow spirit that haunts the field of battle').
The use of semantic markers and selectional restrictions is a crude technique that can, nevertheless, be computationally effective, especially within restricted domains. It is, however, only the tip of a very large iceberg of possible ways in which knowledge can resolve ambiguity. This is well illustrated by considering an example due to Winograd. If somebody says 'The city councillors refused the demonstrators a permit because they feared violence', most people will understand without apparent difficulty that 'they' refers to the councillors. If instead somebody says 'The city councillors refused the demonstrators a permit because they advocated revolution', most people will decide that 'they' means the demonstrators. Knowledge of some kind is enabling people to resolve this ambiguity, and yet this is considerably more subtle than anything that could be readily expressed by semantic markers and selectional restrictions.
Send us a comment.