Knowledge Concepts and Categories

Koen Lamberts,David Shanks

• Editorial: Psychology Press
• Año: 2013
• ISBN:
  9781135064402
• Edición: 1
• Categoría: Psychology
• Categoría: History & Theory in Psychology

Knowledge, Concepts and Categories brings together an overview of recent research on concepts and knowledge that abstracts across a variety of specific fields of cognitive psychology. Readers will find data from many different areas: developmental psychology, formal modelling, neuropsychology, connectionism, philosophy, and so on. The book can be divided into three parts. Chapters 1 to 5 each contain a thorough and systematic review of a significant aspect of research on concepts and categories. Chapters 6 to 9 are concerned primarily with issues related to the taxonomy of human knowledge. Finally, Chapters 10 to 12 discuss formal models of categorization and function learning. The purpose of these three chapters is to provide a few examples of current formal modelling of conceptual behaviour. Knowledge, Concepts and Categories will be welcomed by students and researchers in cognitive psychology and related areas as an unusually wide-ranging and authoritative review of an important subfield of psychology.

CHAPTER ONE

Knowledge and Concept Learning¹

Evan Heit

It has been remarked of sophisticated computer data bases that “everything is deeply intertwingled” (Nelson 1987). This observation also applies especially well to concept learning by humans. Conceptual knowledge has a highly interrelated nature. What a person learns about a new category is greatly influenced by and dependent on what this person knows about other, related categories.

For example, imagine two people who are learning to drive a manual transmission automobile. In effect, these people are learning about a new concept, manual transmission cars. Say that one person has had many years of experience driving cars with automatic transmissions, and the other person has never driven a car before. The first person’s learning will be facilitated greatly by previous knowledge of the category automatic transmission cars, so that this person will be able to quickly find and operate the steering wheel, brakes, radio, etc. in the new car. Yet this prior knowledge would not be of much help as this person is learning about how to shift gears in manual transmission cars. In fact, all of this experience with automatic transmissions might make it especially difficult to learn to operate a manual transmission. Now imagine the situation of the second person, who has never driven before. Overall, this person will probably learn very slowly compared to the first person, because of this person’s lack of relevant prior knowledge. This second person’s learning will likely be a drawn-out process with much trial-and-error practice involved. On the positive side, though, the second person might have some advantage over the first person in learning how to shift gears, because the second person would not have to overcome negative transfer from experience with automatic transmissions.

As another example, imagine that you are an explorer visiting a remote island, with the purpose of writing a book about the people that you see there. You bring to this island many forms of prior knowledge that will guide you in learning about these new people. For example, based on your experiences in other places, you would expect to see males and females, younger and older people, shy people and arrogant people. You would also have certain hypotheses at a more abstract level, for example, that the clothes that someone wears may be related to the person’s age and gender. (Goodman 1955, referred to such abstract hypotheses as overhypotheses.) In a way, these biases resulting from previous knowledge might seem to be undesirable. After all, wouldn’t it be better to be a detached, unbiased observer? However, such biases can make learning much more efficient. Without any prior expectations about what the important categories are on this new island, you would likely spend too much time on unimportant information. For example, you might spend the first month of your visit categorizing people in terms of whether they have small ears or large ears, and the second month trying to notice the relation between ear size and how fast people walk. Without the guidance of your prior knowledge, you could spend an interminable amount of time trying to learn about all the possible categories and the relations among categories. Clearly, some use of prior knowledge of old categories would be critical in learning about the new categories on this island. (See Keil 1989, and Peirce 1931–1935, for related arguments.)

The past decade has been an exciting time for categorization research. Our understanding of the “intertwingledness” or interrelatedness of concept learning has been building steadily. There are numerous situations, such as learning about new objects (like manual transmission cars) or visiting new locations (whether they are new islands or just new restaurants) in which category learning is influenced by what is already known. This chapter will review the experimental evidence for the claim that concept learning depends heavily on prior knowledge, and describe the different ways that prior knowledge has an influence. Furthermore, this chapter will discuss current models of categorization and concept learning with the aim of improving these models to address the important influences of prior knowledge. Finally, inductive reasoning and memory, cognitive abilities that are closely related to categorization, will be discussed in terms of effects of background knowledge.

THEORETICAL ARGUMENTS

The seminal paper concerning knowledge effects on concept learning was written by Murphy & Medin (1985). They contrasted two approaches to describing concept learning, which they referred to as similarity-based and theory-based. According to similarity-based approaches, there is a simple way to tell whether something belongs to a particular category: You assess the similarity between the item and what is known about the category (see also Rips 1989). The more similar item X is to what is known about category C, the more likely you will place X in category C. This similarity-based approach does appear to be a reasonable idea, and it is consistent with several existing accounts of how people learn about categories. For example, take a standard prototype account (Hampton 1993, Rosch & Mervis 1975) of how you might learn about a category such as a novel kind of bird. You would observe members of this species of bird, and remember typical features or characteristics of these birds. These features would be summarized as a prototype, representing the average member of the species (e.g. light brown, fourteen-inch wingspan, lives in treetops). To judge whether another bird belongs to this species, you would evaluate the similarity between this bird and the prototypical list of features.

Murphy & Medin argued that although a similarity-based approach to categorization may be a reasonable start, it will ultimately prove to be incomplete. As illustrated by the earlier example of the explorer visiting an island, there may be so much information available that it will be difficult to simply observe and remember everything. A category learner needs some constraints or biases on what to observe. A related point is that the learner needs to figure out how to describe observations in terms of features. Except perhaps in nature books, birds do not come already labelled with tags such as “light brown” and “lives in treetops”. Such descriptions are inferred and applied by the learner. In addition, people have knowledge about the causal relations between these features that would not be captured by a feature list. For example, it is reasonable to expect that smaller birds will tend to live closer to the ground and larger birds would be more likely to live in treetops, because larger birds can better sustain exposure to wind and severe weather.

These critical influences of knowledge are not explained by similarity-based approaches, Murphy & Medin argued. In contrast, theory-based approaches would consider people’s knowledge about the world, including their intuitive theories about what features are important to observe and how they are related to each other. The Murphy & Medin article did not propose a particular theory-based model of categorization so much as to lay out the challenges that researchers would face in developing a more complete account of categorization that addresses the influences of knowledge. Much of the categorization research published after Murphy & Medin (1985) has presented experimental evidence for, and more detailed empirical accounts of, knowledge effects on concept learning. Also, some work has begun to develop more complete models of categorization that address some of the issues raised by Murphy & Medin. The next two sections of this chapter will review the empirical work on knowledge and concept learning, and the following section will discuss categorization models that address these experimental results.

EXPERIMENTAL EVIDENCE FOR SPECIFIC INFLUENCES OF KNOWLEDGE

At this point, there is quite a bit of amassed evidence on ways that knowledge influences category learning. Before describing this evidence in detail, it is possible to draw some generalizations about what is known. Perhaps the most fundamental generalization is that in learning about new categories, people act as if these categories will be consistent with previous knowledge. People seem to act with economy, so that previous knowledge structures are reused when possible. This generalization is apparent in a few different ways. In general it is easier to learn a new category when it is similar to a previously-known category, as in the earlier example of learning about manual transmission cars. Also, people’s beliefs about new categories include their knowledge from other categories; in effect, there is leakage from one category to another. Likewise, people’s strategies in learning new categories are consistent with their beliefs about other categories. For example, an explorer’s strategies in studying people on a new island would reflect what the explorer knows about the social structure of other places. In the following sections, four different kinds of experimental results will be described, indicating different effects of prior knowledge.

Integration effects

One of the basic influences of prior knowledge on the learning of new categories is integration of prior knowledge with new observations (Heit 1994). That is, the initial representation of a new category is based on prior knowledge, and this representation is updated gradually as new observations are made. For example, imagine that you are walking through some forest for the first time. A nearby forest has large and aggressive birds, so you initially expect the same in the new forest. However, most of the birds you first see are small and unaggressive. As you observe more birds, you gradually revise your beliefs to reflect the local conditions. After just a few observations, your beliefs about the new category of birds might represent an average of your prior knowledge and what you observe. With an even larger number of observations, your beliefs mostly reflect the data from the new forest (small and unaggressive birds) rather than your previous beliefs based on the other forest (large and aggressive). This process is similar to an anchor-and-adjust method of estimation, which Tversky & Kahneman (1974) have argued is a widespread form of reasoning. Also, this process is similar to Bayesian statistical procedures for estimation, in which an initial estimate is revised as new data are encountered (Edwards et al. 1963, Raiffa & Schlaifer 1961).

Recent experiments by Heit (1994, 1995) obtained results that are consistent with an integration account. Instead of being brought to a forest, the subjects in these experiments were shown descriptions of people in a fictional city. Heit assessed subject’s initial beliefs about the city as well as their beliefs after they observed members of categories from this city. For example, the subjects learned about a category of joggers. Initially, subjects expected that about 75 per cent of these joggers would own expensive running shoes. Some subjects then saw descriptions of joggers such that 75 per cent did own expensive running shoes, whereas other subjects saw other proportions (0%, 25%, 50%, 100%) of joggers with expensive running shoes. In their final judgements, subjects acted as if they were taking a weighted average of the expected proportion of joggers with expensive running shoes and the observed proportion. For example, subjects who observed 75 per cent expensive running shoes continued to make judgements of about 75 per cent. Subjects who observed only 25 per cent running shoes ultimately made judgements of about 50 per cent. Furthermore, Heit found that subjects who were given a larger number of descriptions of people in the city tended to discount their prior knowledge more, again consistent with the integration account. (For further experimental evidence of integration effects, see Hayes & Taplin 1992, 1995.)

Clinical psychologists sometimes show similar anchoring effects in their categorizations, or diagnoses, of patients (see Mumma 1993, for a review). Clinicians often show suggestion effects, so that their diagnoses represent an integration of their previous knowledge and their own observations. A typical source of suggestion effects would be a diagnosis made by a colleague. For example, a clinician might categorize a patient as having borderline personality disorder if another clinician has previously reported this diagnosis, even if the patient’s symptoms would fit with a number of other disorders as well. Here, the previous clinician’s analysis of the patient serves as an anchor or initial representation when the new clinician learns about the patient.

A critical aspect of integration effects is the initial category representation that people assemble based on prior knowledge. Ward (1994) has developed a technique for studying these initial representations. This work sheds light on how people borrow information from related categories as they begin learning about a new category. Ward’s task placed people in a creative situation in which they imagined the members of new categories. For example, subjects were asked to draw pictures of animals that might appear on another planet. These imagined animals were very likely to have familiar appendages such as arms, legs, or wings, and to have sense organs such as eyes and ears. Consistent with the idea of integration, Ward concluded that these initial category representations contained a great deal of specific, borrowed information from established categories of animals on Earth.

Selective weighting effects

Several researchers (Keil 1989, Murphy & Medin 1985, Murphy & Wisniewski 1989) have argued that selective weighting effects of prior knowledge are critical in cat...