Relevance To Lectures

There are two related reasons for combining this reading with this week's lecture materials. First, we consider in detail in class only one theory of how words become meaningful. With this reading, I wanted to ensure that you understand that there are many, many other approaches to this issue, and that in fact it is an enormously complex (and interesting) research problem. Second, in the lecture I make an argument that when we understand meaning, we are dealing with a computational level issue. Many people might believe that this "sterilizes" this research topic, and I want to make sure that you don't walk away with such a feeling. Also, many people might argue that the study of meaning has computational elements, but must also be dealt with at other levels too - a view that is supported by some of the research reviewed in the chapter. I'm sympathetic to this position too, and want the chapter to serve as a counterbalance to some of the directions that we take in the lecture.

Margin Notes


"In this chapter we address the following questions: what do we mean by meaning? How do individuals achieve their goals through conversation? Why do they use the language they do?" Communication among individuals requires a common code, a code that pairs meanings with sounds. But an utterance must also be understood in context. There can be a difference between the literal meanings of words and the intended meaning of an utterance that uses them. While the semantics of a sentence is invariant across contexts, the meaning actually conveyed by a sentence is highly determined by context. "What do we mean by the meaning of words? In order to think about the meaning of words, we have to address a deep question: how do language, the mind, and the world connect up?"

The Meaning Of Words

Meaning cannot just be a reference to the world. Also, it can't just be a mental image. "We need to distinguish the sense of a word from its reference and from individuals' associations. The sense of a word is the concept associated with the word. The reference of a word is a set of things in the world to which it can be applied. The concepts of a 'young child' and 'sister' are different though both can refer to the same individual."

The goal of some utterances is to get addressees to pick out some entity. "Analytic approach": "There is a long tradition according to which lexical items can be defined in terms of certain basic concepts or semantic components which fully define them." But there are problems with this approach - many words are difficult to define in terms of necessary and sufficient conditions.

So what should we do? We could accept that there are different kinds of concepts, some of which are analytic, but some of which are not. "In such cases, individuals may know how to recognize these instances and their knowledge of these concepts can be of three sorts: they may know which properties are necessary (if any); they may know which properties normally occur (default properties); and they may know those properties that vary quite freely." There can be vagueness in the specification of a concept. But in the end a set of basic subconcepts is still going to be needed, and these must be related to the world in some way. "Subconcepts are the means by which people think, that is construct propositions about the world. ... Language, mind, and the world interrelate then because language allows the construction of a mental model." This amounts to the view that understanding is the same thing as building a mental model of the world. (NB: This is quite different than covariance theory as described in the lecture, and this mental model view is not without controversy.)

Problem with this approach - what mental model do you build when an infinite number of different models might be consistent with a sentence? The response to this is to assume that a "representative model" will be built, and perhaps later modified. Linguistic evidence suggests that "model building" is part of understanding. For example, elliptical sentences like 'The dogs were too' must be interpreted in the context of a model. "The explicit content of an utterance is usually just a point of departure to capture a state of affairs. The addressee has to fill out the details. The appropriate domain of such an account is not natural language sentences, but propositional representations which permit the construction of models. These models can then make direct contact with models of states of affairs in the world."

Informative And Communicative Intentions

How do individuals determine the intended meaning of a sentence? "Ostensive-inferential communication" involves utterances that have two parts - something that is said (the informative intention) and the intention that what is said is supposed to be recognized (the communicative intention). "According to relevance theory, individuals pay attention to the most relevant phenomena, that is, phenomena which are likely to have cognitive/contextual effects without costing too much processing effort." (NB: later in the chapter, relevance theory becomes more and more important, and some of the meat of this sentence is put on the table.)


What is needed to make a conversation successful? According to the autonomous view, each participant says the right thing at the right time, with one person taking the lead. Grounding theory provides a second view, in which participants collaborate to establish the mutual belief that the speaker is being understood.

In grounding a conversation, checks are periodically made to ensure that understanding is happening. How is such checking done? According to the contribution model, contributions to conversations are divided into two phases - presentation and acceptance. In the presentation phase, something is said. In the acceptance phase, the speaker looks for negative or positive evidence to ensure that they have been understood. These phases can be embedded hierarchically - for instance, within an acceptance phase one might find sub- presentation and sub-acceptance phases that reflect a subcomponent of the conversation.

What evidence is there for collaboration in conversations? One way to determine this is to look at how utterances that make up a conversation are formulated. It has been argued that "the appropriate way to capture the workings of the process of presentation and acceptance is the principle of least collaborative effort: participants try to minimize the cognitive work both do from initiating to accepting an utterance."

How do you study this? One paradigm involves referential communication. There are a bunch of abstract cards (tangrams) seen by both subjects. One subject sees them in order, describes them to a second subject, and the second subject tries to put a second set of the same cards in the same order. The dependent measure here are qualitative analyses of what subjects say to each other. These data support the notion that conversations are collaborative. There are more words said earlier in the experiment than later, and as the conversation evolves it depends on the recall of previous (longer) descriptions of stimuli. In this kind of study, self-repairs - corrections of utterances that attempt to increase acceptance - decrease over time. Agreement of reference (i.e., with respect to which card is being referred to) is based mostly on permanent properties of described stimuli, which makes sense because this provides a solid foundation for acceptance. Finally, those who just hear the conversation but who do not take part in it - and therefore cannot collaborate to not accept references that they do not understand - do poorer than those who participate in terms of putting the cards in the right order at the end.

"Individuals work to refine their language, then, to suit their communication goals and change the nature of the descriptive language used over the course of the task in order to access their models of the situation as effectively as possible." This has been nicely studied in a task where subjects have to collaborate their way through a computer generated maze. This kind of communication requires an adequate description of the maze - an appropriate model. In this task, it was seen that over time there were different descriptive schemes used by subject pairs. As a pair of subjects grew more acquainted with the task, they evolved into using a coordinate system to economically describe their mutual situations.

Literal And Non-Literal Uses Of Language

According to Grice, conversation works because people agree to cooperate following certain maxims: "the maxim of quantity (say as much as and no more than is needed); quality (only say what you believe to be true); relation (be relevant); and manner (be clear)." When one maxim appears to be violated, it is assumed that this couldn't be true - so an intended (vs. literal) meaning is sought to "fix" the situation. This assumes that literal meanings of utterances are given priority. But this may not in fact be true of conversation - lots of evidence exists for the primary processing of metaphors.

How are metaphors understood? One view is that metaphor is a kind of comparison of features. A second view is that metaphor is a literal class inclusion statement. Metaphor comprehension relies a great deal on context.

"Understanding any utterance, literal or non-literal, requires understanding the communicative intentions of the speaker." A good example of this is in understaning irony - why would someone say something opposite to what they intended? "According to relevance theory, [irony] is designed to communicate the speaker's attitude towards a thought, view, or utterance, which is attributed to someone other than the speaker."

Understanding metaphor and irony seem similar - but with respect to relevance theory, there is one crucial difference. Understanding irony is more complex than understanding metaphor, because in the former "the addressee has to infer that the speaker is attributing a thought to someone other than herself and is expressing her own non-supportive attitude towards it." Indeed, some clinical populations (e.g., autistic children) have great difficulty in understanding figurative language - and this is apparently due to problems with meta- representation (i.e., the ability to represent beliefs about beliefs). Meta- representational problems make it very difficult to deal with irony.

"An important notion here is the idea that individuals seek to maximize the relevance of all information, including, of course, communicated information." But the cost of cognitive processing of information has to be factored in too. "The greater the cognitive effects the greater the relevance of the informaiton. Deriving cognitive effects is not without cost or effort and, all else being equal, a given cogniive effect achieved at greater processing cost will be less relevant."

"An utterance on a given interpretation is optimally relevant if and only if it achieves enough effects to be worth the addressee's attention and puts the addressee to no unjustified effort in achieving these effects. ... every utterance creates a presumption or expectation of its own optimal relevance."

There is a principle of rationality that can be applied to comprehension: possible interpretations are examined in order of their accessibility. "The process stops when the expected level of relevance is achieved. Only one interpretation at most will meet the criterion of consistency with the expectation of optimal relevance."

Pearl Street | "An Invitation To Cognitive Science" Home Page | Dawson Home Page |