Lexical Processing and Semantics: Concepts, Models, and Implications
Tutorial Announcements
Room changes: Check the e-student portal for updated tutorial locations due to higher enrollment and to avoid last-minute room capacity issues.
Apology: Acknowledgement of inconvenience for students who went to the wrong room.
Tutorial details: Check updated tutorial rules and tutors on the portal; if you couldn't register for tutorials, email the course coordinator Alice Wu urgently to get placed in a tutorial.
Practical nature of tutorials: Tutorials are supposed to be practical, interactive, and physical. If you need access, contact the tutors to discuss details.
Materials: Tutorial slides and activities for each week are posted under the respective week; you should have access to these.
Assessment: First assessment is a critique of an article. An example from last year is provided to illustrate expectations (specific and concise writing). The current article/theme for this year’s critique is posted under the “critique of an article” section. The deadline requires posting your written critique to the designated link.
Questions: Clarified that questions about tutorials or the first assessment should be addressed in the tutorial or via the course coordinator.
This week start: The lecturer shifts to content on lexical processing and word recognition, starting with revision of last week’s concepts and moving into deeper material.
Lexical Processing: Key Concepts and Terms
Levels of representation of a word (revision):
Semantics (high-level meaning)
Orthography (spelling/letter structure)
Phonology (sound structure)
Lexical decision task: Decide whether a string of letters is a word in the language (e.g., English). Demonstrates word vs. non-word processing.
Naming: Task uses naming a picture; measures reaction time (RT) typically in milliseconds and accuracy.
Reaction time (RT): Measured in milliseconds. Concept: extremely fast processing; millisecond-level processing underpins language comprehension.
Milliseconds: 1 ms = 1/1000 second.
Naming measures: RT and accuracy; RTs reflect retrieval and production processes.
Access and post-access processing:
Post-lexical/post-access processes occur after initial lexical access, such as verifying relationships in priming paradigms.
Example: in a priming task, the relationship between a target word and a prime is checked after lexical access.
Eye Movement Research in Reading
Two primary measurable variables:
Fixations: how long the eye remains on a word (processing time per word).
Saccades: rapid eye movements between fixations; regressive saccades move back to earlier text to verify information.
Other measures mentioned:
Pupil dilation: correlated with cognitive load and information processing.
Paradigms and terms:
Visual word paradigm: another eye-tracking approach (relevant in multilingual contexts).
Perceptual span: the region of text that can be effectively perceived during a fixation; comprises focal (phoria) and peripheral (paraphoria) areas.
Phoria vs Paraphoria: focal vs less important areas within the perceptual span.
Perceptual span loosely defines how much information is captured during reading.
Why eye movements reveal reading processes:
The eyes are a window into cognitive processing; pupil size can index processing load beyond fixation duration.
Eye movements provide millisecond-level timing information about when and how readers access and integrate information.
Priming, Frequency Effects, and Semantic Influence
Frequency effect: Greater exposure to a stimulus facilitates subsequent processing (faster recognition).
Priming paradigms:
Masked priming: brief, often unconscious presentation of a prime that facilitates processing of the target.
Semantic priming: related word pairs show facilitated processing (e.g., doctor–nurse; dogs–cats).
Empirical results from priming studies inform theoretical models of lexical access and semantic networks.
Models of Lexical Processing
Two broad model types discussed:
Search model (early, serial processing): information is processed one step at a time (e.g., D → O → G for a word like DOG).
Interactive Activation (IA) model, a connectionist approach: parallel processing across multiple levels (features → letter units → word); processing is distributed and interactive with bidirectional activation.
Core distinction:
Serial search processes information step-by-step.
IA/connectionist models enable parallel processing with spreading activation among levels.
Structure of the IA model (three levels):
Features (low-level properties)
Letter units (orthographic representations)
Word units (lexical representations)
Mechanisms:
Activation spreads across the network; both bottom-up and top-down feedback influence processing.
The most supported view is that the interactive activation model captures reading/lexical processing more accurately than strict serial search models.
Current emphasis: semantics as a difficult component to pin down; meaning is stored and accessed through more than just orthographic/phonological codes.
Semantics: Denotation, Connotation, and Representations
Denotation vs Connotation:
Denotation: core, essential properties that define a category (e.g., what makes a dog a dog — functions, general attributes).
Connotation: culturally variable, context-dependent attributes (e.g., dog can be viewed as friendly, dangerous, dirty, etc.).
Semantic memory theories:
Semantic networks (spreading activation): concepts as nodes; related concepts connected by links; priming results from spreading activation.
Difficulties: the theory is highly powerful and flexible, but this makes falsifiability challenging (a good theory should yield falsifiable predictions).
Feature-based theories as alternative:
Concepts decomposed into perceptual and functional features.
Two subtypes within feature theories:
Classical (defining) view: categories have defining features; membership is necessary and sufficient; examples include defining features of birds.
Prototype (family resemblance) view: no single set of necessary features; categories are represented by typical exemplars or prototypes; membership is graded and context-dependent.
Perceptual vs functional features:
Perceptual: features derived from perception (color, shape, texture).
Functional: features describing typical uses and utility (e.g., an object is edible or can be used for sitting).
Context dependence in meaning:
Some meanings and features are context-independent (e.g., semantic core of ‘bank’ can contain money in specific contexts).
Other meanings are context-dependent and revealed only in situ (e.g., roof can be walked upon in one scenario but not inferred from a prior sentence).
Context can shift category membership or the weight of particular features in real-time processing.
Defining vs contextual features:
Some features are stable across contexts; others vary with context, situation, or culture.
Classical vs Prototype Theories: Examples and Implications
Classical (defining features) issues:
Some categories have fuzzy boundaries (e.g., bread vs cake, game, sports, abstract terms like love or freedom).
Some categories are hard to define with a finite list of features (e.g., 'game' or 'sports'; abstract terms).
Some items are ambiguous depending on culture or perspective (e.g., is a tomato a fruit or vegetable? depends on context and culture).
Prototype theory (family resemblance) advantages:
Explains graded category membership and typicality effects (e.g., which fruit is most representative: blueberry vs coconut?; kiwifruit vs orange; strawberry vs cucumber).
Accounts for fuzzy boundaries and items belonging to multiple categories.
Explains graded structure within a category (e.g., 13 vs 57 as odd numbers; 13 more prototypical than 57; frequency or typicality affects judgments).
Challenges for prototype theory:
Difficult to identify a single best exemplar for abstract concepts (e.g., truth, love, freedom).
Some intuitive tasks (e.g., odd-number judgments) may not align with a simple frequency-based prototype account for all individuals or cultures (cultural knowledge and numeracy can influence typicality judgments).
Cultural and linguistic differences can alter prototypical exemplars (e.g., the Chinese cultural association with the number 4 due to death homophone; this affects typicality judgments and processing).
Integration and parallels:
Prototype theory has influenced broad areas including speech perception and phoneme categorization (e.g., most representative phoneme exemplars shaping category perception).
Real-world categories often exhibit both definitional and prototype-like properties; many researchers incorporate both perspectives.
Context-Dependent Meaning and Sentence-Level Evidence
Sentence verification tasks as a method to test context dependence:
Bank example: two sentences explored whether the bank can contain money vs. can be robbed; results suggest some meanings are immediately accessible (definitional/essential attributes) while others require context to retrieve relevant properties.
Roof example: contrast where a roof can be walked upon is contextually primed, while the roof’s primary function (to keep out rain) is less related to that property; this demonstrates context-dependent retrieval of certain semantic properties.
Implications for lexical semantics modelling:
Some meanings are robust and context-independent, retrieved quickly when a word is encountered.
Other meanings are activated only when supported by local context or specific semantic relations, underscoring the role of context in meaning construction.
Developmental and Theory-Theory Perspectives
Theory-theory approach (developmental): knowledge-based understanding of meaning evolves with life experiences and learning; objects acquire meaning through knowledge about their properties and relations, with less emphasis on fixed feature sets.
Child vs adult categorization (developmental data):
Children tend to interpret identity more literally based on perceptual features; e.g., dyeing a tiger’s stripes leads younger children to call it a leopard or a lion, while older children and adults maintain the tiger identity if essential properties persist.
As children acquire more knowledge about object identity, their categorization aligns more with adult-like notion of essential identity despite superficial changes (e.g., dyeing the outside vs. inside identity).
Implications:
Early categorization may rely on perceptual features and surface attributes; more advanced categorization incorporates essential properties and functional knowledge.
This developmental shift supports a knowledge-based view of semantic memory, where understanding grows with experience and theory-like reasoning about objects and categories.
Rogers et al. Connectionist Model of Semantic Memory
Architecture and key components:
Verbal layer: units for words (labels) that map to semantic features and can be driven by reading/writing (orthography/phonology).
Visual layer: units for visual features (shape, color, appearance) that map to semantic content.
Semantic (higher-level) layer: a set of abstract semantic units that integrate information from verbal and visual layers.
Encyclopedic features: real-world knowledge such as typical locations or usage (e.g., a cup is usually in the kitchen).
How the network operates:
Activation flows bidirectionally between verbal, visual, and semantic units; weights reflect prior experience and learning.
Activation from a word (name) travels to related perceptual features and then to semantic units, which can in turn activate related visuals; similarly, a visual cue can activate semantic and then the corresponding word forms.
Similar or closely related concepts produce similar activation patterns (e.g., budgie vs canary share semantic and visual properties).
Advantages of this model:
Avoids the need for pre-specified feature lists; meaning is represented as a pattern of activation across a network of units.
Explains how different inputs (words vs pictures) can access the same semantic content via shared semantic representations.
Prediction and simulation can specify how knowledge is stored and retrieved by adjusting connections and training history.
Key takeaway:
The Rogers et al. model demonstrates a plausible mechanism for integrating visual, verbal, and encyclopedic knowledge in semantic processing, with activation dynamics guiding lexical retrieval and comprehension.
Summary and Takeaways
Three broad approaches to lexical semantics:
Feature-based theories (classical/defining features; prototype/fuzzy boundaries; decompositional approaches).
Knowledge-based/developmental approaches (theory-theory; context-specific and learned representations).
Connectionist/associationist approaches (semantic networks with spreading activation; parallel distributed processing and feature integration).
Core contrasts:
Classical feature theories emphasize clearly defined features and necessary membership criteria; struggle with fuzzy categories.
Prototype theories emphasize typicality and graded membership; handle fuzzy boundaries but struggle with best exemplar definitions for abstract concepts and some cross-cultural effects.
Connectionist models emphasize distributed representations and learning from experience; activation patterns determine meaning access without relying on explicit feature lists.
Context matters:
Meaning is often context-dependent, with certain features becoming salient or primed only in specific linguistic or situational contexts.
Developmental perspective:
Children’s conceptual representations evolve from perceptual and surface-level cues to more abstract, knowledge-based understanding, supporting the notion that meaning is learned and theory-driven.
Practical relevance:
A robust theory of lexical semantics needs to accommodate rapid word recognition, flexible meaning retrieval, cross-modal representations (visual, verbal), and cultural/individual differences in typicality and interpretation.
Looking ahead (next week):
The course will continue with sentence-level processing and broader semantic-syntax interaction topics, building on today’s foundations.
Connections to Prior Content and Real-World Relevance
The material connects to foundational cognitive psychology principles: memory representation, priming, categorization, and processing speed.
Real-world relevance includes understanding language comprehension in multicultural contexts (toy examples like fruit/vegetable category, tomato ambiguity) and how context shapes interpretation in everyday communication.
The content has implications for educational approaches (how children acquire language and meaning), linguistics (categorization and semantic theories), and AI language models (how to model meaning and context in computational systems).