Chapter 7: Semantic Memory

Semantic memory constitutes a critical aspect of long-term memory, encompassing general world knowledge that is distinct from episodic memory, which recalls specific events. This chapter explores the organization, structure, and models of semantic memory, shedding light on its fundamental role in cognition.

An interesting aspect of semantic memory is that it represents a type of understanding often missing from AI models. For instance, a robotic arm programmed to arrange blocks in a pattern might attempt to build a tower starting from the top, revealing its lack of understanding of basic concepts like gravity. Similarly, AI language models like ChatGPT can store and retrieve vast amounts of factual information but often lack true comprehension. For example, when asked to illustrate Godden and Baddeley's (1975) experiment, which demonstrated that memory improves when study and test locations match, ChatGPT produced an image of desks underwater—a strange, but possible interpretation—but also showed desks improbably floating on the water's surface, defying real-world constraints (see Figure 7.1). This highlights how AI can combine stored facts in ways that disregard fundamental physical principles, demonstrating a gap in its semantic understanding of the world.

Figure 7.1. When asked to illustrate Godden and Baddeley's (1975) experiment, ChatGPT generated this image. Here people are shown sitting at desks on top of the water along with other unlikely events. This shows that AI has gaps in its semantic understanding of the world.

"AI-generated depiction of Godden and Baddeley's experiment." by Kahan, T.A. is licensed under CC BY-NC-SA 4.0

Introduction to Semantic Memory

Semantic memory stores factual knowledge about the world, including concepts like historical facts, meanings of words, and general knowledge about objects and events. Unlike episodic memory, which recalls personal experiences, semantic memory involves information that is universally understood and applied.

Computational Models of Semantic Memory

Hierarchical Network Model

The hierarchical network model that was developed by Collins and Quillian in the late 60s, proposes that semantic memory is organized in a hierarchical manner (see Figure 7.2), with broad concepts at the top and specific instances underneath. Concepts are linked by "is-a" (superordinate) and "has" (property) links, forming a network where activation spreads based on associations. The links between layers are the “is a links” or superordinate links while the links within a layer are the “has” links or property links. This model explains how information retrieval and decision-making are influenced by the spread of activation and an intersection search. When given a statement like “a dog is an animal” in a sentence verification task activation from dog and animal spreads. If the activation meets (intersects) then a person can verify the statement as true and in instances where the activation does not intersect – “a table is an animal” – the person will indicate that the statement is false. This network also explains priming effects where reaction times to read the word “dog” are faster when preceded by a related concept (“animal”) than when preceded by an unrelated concept (“furniture”). According to this model the spread of activation from animal to dog lessens the critical amount of activation needed for dog to reach threshold and be named (i.e., causing the priming effect).

Figure 7.2. Depiction of a portion of a semantic network as described by Collins and Quillian.

"Depiction of Collins and Quillian’s semantic model." by Kahan, T.A. is licensed under CC BY-NC-SA 4.0

• Strengths: Explains sentence verification and priming effects.
• Weaknesses: Fails to account for some cognitive phenomena like typicality effects and unlearning.

• Typicality Effects: Typicality effects refer to the finding that people are faster at verifying statements involving more typical category members compared to less typical ones. For example, verifying "A robin is a bird" is quicker than "An ostrich is a bird," even though both are at the same hierarchical level in the model. The hierarchical network model assumes that all category members are equally accessible because they are stored at the same level. This uniformity makes it difficult for the model to account for the variability in response times caused by differences in how "typical" a category member feels based on real-world experience.
• Unlearning: Unlearning refers to the process of modifying or forgetting incorrect or outdated information in memory. For example, if someone initially learned that Pluto is a planet but later learned that this isn’t true, their memory needs to update. The hierarchical network model has no mechanism for modifying or removing links in the network, making it hard to adapt to new information or resolve conflicts between old and updated knowledge.

Feature Comparison Model

The feature comparison model posits that concepts in semantic memory are defined by their features (see Figure 7.3), which can be defining (essential characteristics) or characteristic (typically associated). Decision-making involves comparing feature lists in a two-stage process: global comparison followed by specific feature analysis. If a match between all of the features (defining and characteristic) can be made quickly then a fast yes response is made (e.g., a robin is a bird). Similarly, if there is very little overlap in features then a fast no response can be made (e.g., a chair is a bird). However, if there is some overlap then a comparison of defining features is needed (stage 2) and reaction times are slower (e.g., “a penguin is a bird” or “a bat is a bird”).  This model addresses the typicality effect because typical members of a category can be verified in stage 1 while atypical members are not verified until stage 2. However, this model struggles with defining feature identification and property statements. For example, it is difficult to develop a list of defining features for the concept “game” or “home” that everyone would agree upon. Likewise, it is difficult to develop a list of defining and characteristic features for a property like “brown”.

Figure 7.3. Depiction of the two-stage feature comparison model.

"Depiction of the feature comparison model." by Kahan, T.A. is licensed under CC BY-NC-SA 4.0

• Strengths: Accounts for typicality effects in decision-making.
• Weaknesses: Limited to yes/no responses in the sentence verification task and overlooks tasks like priming.

Revised Network Model

The dominant theory in cognitive psychology concerning semantic memory organization is the Revised Network Model (see Figure 7.4). Unlike earlier hierarchical models, this framework emphasizes a network structure where concepts are interconnected based on semantic relatedness rather than a strict hierarchical arrangement. In this model, concepts that are strongly related or frequently co-occur in experience are stored closer together within the network. This closeness is represented by stronger connections or shorter distances between nodes (Collins & Loftus, 1975).

Figure 7.4. Depiction of Collins & Loftus’ revised network model.

"Depiction of the revised network model." by Kahan, T.A. is licensed under CC BY-NC-SA 4.0

Hyperspace Analog to Language

A key concept contributing to the understanding of semantic memory formation is the Hyperspace Analog to Language (HAL) model proposed by Curt Burgess (Miller & Charles, 1991). HAL operates on the principle that semantic knowledge is acquired through statistical co-occurrences of words in natural language contexts. Burgess demonstrated this by analyzing vast amounts of textual data—over 300 million words—extracted from internet sources. The data were transformed into a high-dimensional space (140,000 dimensions), where each dimension captures a different aspect of word co-occurrence patterns.

The Hyperspace Analog to Language (HAL) model represents word meanings as points in a high-dimensional semantic space derived from word co-occurrence patterns in large language corpora. In HAL, the semantic meaning of a word is determined by its "context" within a sliding window of surrounding words, and these relationships are captured in a co-occurrence matrix.

Figure 7.5. Depiction of two of the 140k dimensions in the HAL model. Here the dimension shown on the x axis seems to measure “level of domestication” and the y axis measures “size”.

"Depiction of the Hyperspace Analog to Language (HAL) model." by Kahan, T.A. is licensed under CC BY-NC-SA 4.0

The dimensions of this matrix are not predetermined and the values in the cells reflect how often and how closely words appear together. Words with similar usage patterns are represented as vectors close to each other in this semantic space, which reflects their relatedness in meaning. For example, "dog" and "puppy" would have vectors near each other because they often occur in similar contexts. See Figure 7.5 for an example where the dimensions of size (Y axis) and domestication (X axis) are represented. Importantly, these two dimensions (of the 140,000 total dimensions) were not pre-determined by the researchers but instead were extracted by the model based on word co-occurrence. HAL's strength lies in its ability to model the acquisition of semantic relationships through exposure to language, providing insights into how contextual learning shapes semantic memory.

Prototypes

Prototypes play a crucial role in categorization and recognition processes within semantic memory. According to prototype theory, categories are represented by idealized, central examples that embody the typical features of the category (Rosch, 1978). For instance, a prototype of a bird might include features such as wings, feathers, and the ability to fly. Research by Langlois and Roggman (1990) extended prototype theory to the domain of face perception, showing that averaged faces—created by blending multiple facial images—tend to be perceived as more attractive and prototypical representations of human faces. For example, a face that is created by averaging 4 faces is rated as less attractive than a face that is created by averaging 32 faces.

The evolutionary theory posits that average faces are considered more attractive because they signal genetic health and fitness. According to this perspective, faces that are closer to the population average may indicate a lack of genetic abnormalities, a diverse gene pool, and developmental stability, making them appealing from an evolutionary standpoint.

However, there are critiques of this theory. Evidence shows that attractiveness does not strongly correlate with health or longevity, undermining the idea that average-looking individuals are inherently healthier. Additionally, averageness preferences extend beyond human faces to inanimate objects like watches and even birds, suggesting that the preference for averageness may be a broader cognitive bias rather than an indicator of genetic fitness. This leads to alternative explanations, such as symmetry: morphed faces naturally become more symmetrical, and symmetry is widely regarded as a marker of attractiveness.

Another compelling alternative is prototype theory, which suggests that humans find attractiveness in stimuli that closely resemble mental prototypes. For faces, this means the most familiar or "average" features across many exposures become more appealing due to their alignment with a cognitive prototype. This theory also accounts for preferences in composites of profile views, where symmetry is not possible.

Together, these studies suggest that symmetry and prototypes may be better explanations of the attractiveness of morphed faces, rather than the evolutionary view that these faces signal genetic fitness.

Semantic Memory and Priming

Semantic memory encompasses our general world knowledge, including facts and concepts not tied to specific personal experiences. One of the primary methods used to study semantic memory is through priming experiments. These experiments aim to observe how exposure to one stimulus (the prime) influences the processing of a subsequent stimulus (the target).

Types of Priming Tasks

1. Pronunciation Tasks:

• Participants are instructed to read aloud a word quickly after seeing a related or unrelated prime. For example, saying "boy" faster after seeing the prime "girl."

2. Lexical Decision Tasks:

• Participants decide whether a string of letters forms a word or not. This task measures how quickly and accurately participants respond to a target word after a related or unrelated prime.

3. Word Stem Completion:

• Participants complete a word stem with the first word that comes to mind, influenced by a preceding prime. For instance, completing "B___" with "boy" more often after seeing the prime "girl."

Effects of Priming: Facilitation and Inhibition

Primarily, priming results in facilitation, where responses to the target are quicker and more accurate due to the activation of related information in semantic memory. However, inhibition can also occur, where the response to the target is slowed down by a conflicting or unrelated prime.

Factors Influencing Priming Effects

1. Semantic Network Activation:

• The spread of activation theory suggests that exposure to a prime activates related concepts in the semantic network. This activation facilitates the processing of related targets.

2. Timing and SOA (Stimulus Onset Asynchrony):

• Priming effects can vary based on the time interval between the prime and the target (SOA).

3. Types of Priming:

• Semantic priming involves related meanings (e.g., "dog" priming "cat").
• Phonological priming involves similar sounds (e.g., "moose" priming "juice").
• Orthographic priming involves similar spellings (e.g., "touch" priming "couch").
• Repetition priming involves all of these (e.g., "bamboozle" priming "bamboozle").

Automaticity of Semantic Activation

Research indicates that semantic activation in priming tasks occurs automatically, even when participants are not consciously aware of the primes or their effects. This automaticity is evident in experiments where masked primes (briefly presented and obscured) still influence subsequent processing speeds.

Cross-Linguistic Studies

Studies across different languages reveal universal patterns in semantic priming, despite linguistic differences. This suggests that the mechanisms underlying semantic memory and priming are fundamental aspects of cognition.

Practical Applications

Understanding semantic memory and priming has practical applications in fields like marketing, where priming techniques can subtly influence consumer behavior.

Automaticity in Word Processing and Semantic Activation

Neely's work demonstrates the inevitability of word processing upon visual encounter (see Neely & Kahan, 2001, for an extensive review of the evidence showing the automatic activation of words). Regardless of intentions or subsequent expectations, individuals invariably engage in processing the meaning of encountered words. This automaticity is vividly illustrated in tasks like the Stroop task, where participants are instructed to ignore the word and focus solely on color naming. Despite these instructions, the interference from the word's meaning consistently affects response times. For instance, words related to colors (e.g., "red" written in blue ink) delay color-naming responses, showcasing the automatic processing of semantic information.

Challenges to Automatic Semantic Activation

Despite widespread acceptance of automatic semantic activation theories, empirical findings present challenges. One such challenge is the phenomenon where priming effects diminish under conditions of focused attention on individual letters. This finding suggests that semantic activation may not occur under all conditions and highlights the role of attentional focus in modulating cognitive processes. For example, if a person is shown the word "girl" with the letter "g" repeated over it and is first required to decide whether the letter "g" appears in the word "girl" before naming the target word "boy" aloud, no priming effect is observed.

Automaticity, Repetition Priming, and ERP Evidence

Neely, VerWys, and Kahan (1998) found surprising evidence that the repetition of a prime (e.g., presenting "girl" twice before "boy") eliminates the priming effect. This finding contradicts the notion of automatic semantic activation, suggesting that semantic processing might be more nuanced than previously thought. However, event-related potential (ERP) studies reveal that even when priming effects are absent in behavioral measures (reaction times), neural activity patterns suggest ongoing semantic processing, which supports the automaticity of semantic activation. Specifically, ERP components like the N400 indicate that words are processed for meaning even when this processing does not manifest in overt behavioral effects.

Strategic Influences on Semantic Priming

Jim Neely, seen pictured with his former Ph.D. student Todd Kahan in Figure 7.6, has also found evidence for strategic effects in semantic priming studies (Neely, 1977). In Neely’s now classic experiment, he investigated whether semantic activation occurs automatically or is influenced by expectations. Participants were tasked with deciding as quickly as possible whether a target item was a word or not (i.e., lexical decision) after being shown a prime word. For example, the prime word "body" was sometimes paired with the unexpected but semantically related target "heart" or with the expected but unrelated target "door."

Figure 7.6. Photo of Todd Kahan with his Ph.D. advisor Jim Neely. Photo taken by Kathy Mathis at the Psychonomic Conference in Boston in November of 2022.

"Photo of Todd Kahan with his Ph.D. advisor Jim Neely." by Kahan, T.A. is licensed under CC BY-NC-SA 4.0

The key variable was the stimulus onset asynchrony (SOA)—the time interval between the presentation of the prime and the target. Neely manipulated the SOA to examine how different amounts of time affected the influence of automatic associations and expectations on performance.

At a short SOA, the pattern of results showed priming for unexpected but semantically related items (e.g., "body-heart"). This indicates that automatic semantic activation was at play, as there was not enough time for participants to form expectations. However, at a long SOA, priming occurred for expected but unrelated items (e.g., "body-door"), reflecting the influence of participants' expectations, which had time to develop (see Figure 7.7).

Figure 7.7. Critical results from Neely’s 1977 study. Facilitation is found at a short prime-target SOA for related but unexpected pairs but facilitation for unrelated but expected prime-target pairs is only found at a long SOA.

"Critical results from Neely’s 1977 experiment." by Kahan, T.A. is licensed under CC BY-NC-SA 4.0

These findings demonstrate that both automatic semantic activation and conscious expectations contribute to semantic priming. At short SOAs, automatic processes dominate, but given enough time, expectations can override automatic associations to influence performance. This study highlights the dynamic interplay between automatic and controlled processes.

Neely's 1977 work was groundbreaking (cited by over 1,300 research studies) because it demonstrated how strategic, conscious processes can interact with—or counteract—automatic, unconscious components of cognition. This insight has profoundly influenced many areas of psychology. For instance, in social psychology, it has informed our understanding of automatic biases that may conflict with a person’s conscious beliefs. Despite good intentions, these automatic biases can still exert powerful and often insidious effects. This is evident in tasks like the shooter bias task, where participants are more likely to mistakenly "shoot" unarmed individuals from stereotyped groups, or the Implicit Association Test (IAT), which reveals unconscious associations, such as linking certain groups with negative attributes. These findings, inspired by Neely's work, underscore the dual influence of automatic and strategic processes, shaping interventions to mitigate biases and deepen our understanding of human behavior.