Comprehensive Notes on Word Recognition: WSE, PSE, IAM, Semantic Priming, and Context Effects

Class context and introductory remarks

• Student-faculty banter about the lecture room and overflow: one room is empty due to overflow; frustration about moving to a different room, which affects attendance and perceived efficiency.
• Scheduling friction on Mondays and Fridays noted (five or six students in several sessions per day; fatigue toward end of week).
• Instructor acknowledges start time and aims to finish earlier to allow weekend downtime.
• Brief discussion about reading today: topic is word superiority effect and pseudoword superiority effect, plus the Interactive Activation Model (IAM) of reading, and how these relate to single words vs. sentences.
• Overview: next, a model of reading, semantic priming, and eventually linkage to higher-level language processes.

Key goals of the lecture

• Define and explain the word superiority effect (WSE) and pseudoword superiority effect (PSE).
• Describe the Interactive Activation Model (IAM) of reading: three levels (feature, letter, word) and the roles of bottom-up vs. top-down processing.
• Discuss experimental evidence for WSE and PSE, including the role of orthographic neighbors and word frequency.
• Introduce semantic priming, lexical decision tasks, and effects of context (sentence-level priming).
• Compare short-letter-string experiments with sentence-context experiments and discuss predictability (close scores).
• Highlight strengths and limitations of IAM; preview other models to be covered later.

Reading at the core: what reading involves

• Reading requires perceiving that a sequence of letters forms a word, then reading multiple words and using prior knowledge to interpret meaning.
• Studies show university students read about
• r \approx 300 \ \[ \text{words/min} \] which corresponds to about
• $t \approx \frac{1}{r} \times 60 \text{ s} \approx 200 \text{ ms/word}$ (roughly two hundred milliseconds per word). This is fast but not as fast as spoken language.
• There is debate about what counts as word recognition: some define it as purely visual processing (recognizing letters regardless of meaning), while others include access to word meaning.
• When meaning is engaged, processing can slow down if the focus shifts to semantic processing rather than mere recognition.
• Visual word processing is often equated with reading by the general public, but true reading involves sentences/paragraphs and text-level processing; visual word processing can be studied with constrained, short word strings (3–5 letters) to control experimental variables.

The Interactive Activation Model (IAM): framework and mechanics

• IAM is a computational model from the early 1980s that uses parallel processing and combines top-down and bottom-up processing.
• Three levels of recognition:
  • Feature level: detect basic letter features (e.g., a vertical stroke on the edges of a letter). Example: letters with vertical edges activate excitatory connections for those features and inhibit letters without those features.
  • Letter level: identify the specific letter from activated features (requires some top-down knowledge about letter forms).
  • Word level: activate words that are composed of the identified letters, based on prior knowledge (lexicon).
• Activation and inhibition flow simultaneously across levels (parallel processing): features → letters → words, with top-down input from word knowledge to support letter identification and from word knowledge to influence feature/letter activation.
• Bottom-up processing: feature and letter information build up to activate word representations.
• Top-down processing: prior knowledge about which words are likely given context or letter sequences speeds up recognition.

Word superiority effect (WSE) and pseudoword superiority effect (PSE)

• WSE: real words are recognized more quickly and accurately than non-words when presented as a string of letters briefly (e.g., with a mask following the string).
• In IAM terms, recognizing a word at the word level increases activation for its constituent letters and inhibits competing letters, speeding up recognition.
• Example task: briefly present a four-letter string (e.g., e _ t) followed by a mask; participants choose which letter fits into the blank (from two choices). If the string is a real word, letter identification is faster than if the string is a non-word.
• Orthographic neighbors: words that differ by one letter (or one sound) form an orthographic (or phonological) neighborhood; larger neighborhoods can slow recognition if frequent neighbors compete.
  • High-frequency neighbors tend to produce stronger inhibition, slowing the target word’s recognition.
  • Frequency of the target word itself matters: high-frequency words (e.g., seat) are processed faster than low-frequency words (e.g., sect or sent in many contexts).
  • Individual differences matter: word frequency can vary by context and by person (e.g., a post office worker may read SENT more often and so may rate it as higher frequency than someone else).
• Cambridge reading sample (context effect): a classic example where the first and last letters are in place but middle letters are jumbled; many readers still recognize the word, illustrating strong top-down influence and familiarity effects. This challenges purely bottom-up letter-by-letter processing and supports a role for prior knowledge and context in reading.
• Peripheral vision limitation: the WSE appears to diminish or disappear when letter strings are presented in peripheral vision, suggesting top-down processing is more effective in central vision.
• Pseudoword superiority effect (PSE): pronounceable non-words (pseudowords) facilitate processing more than unpronounceable letter strings. This supports the role of phonological processing and inner speech in reading; readers exploit pronounceability to speed up processing even when the string is not a real word.
• Phonological processing and language differences:
  • The facilitation from pseudowords relies on phonological rules of the reader’s language; bilingual readers may show different pseudoword effects depending on language proficiency and phonotactics.
  • Lexical neighbors can influence processing depending on whether the neighbor is more or less frequent than the target word.

Lexicon, neighbors, and frequency considerations

• Lexicon: our mental dictionary of words; bilinguals may have multiple lexicons or a single lexicon accessed differently depending on language context.
• Lexical neighborhood: set of words that differ from the target by one letter or one phoneme.
• Effects of neighbors:
  • If a neighboring word is highly frequent, it can slow down recognition of the target word due to competition.
  • If a neighboring word is less frequent than the target, neighbor activation can speed up recognition, depending on context and exposure.
• Orthographic vs phonological neighborhoods: both can influence recognition; activation spread can be facilitated or inhibited by neighbors depending on frequency and similarity.
• Word frequency and individual experience:
  • Frequency lists exist but are not universal; individuals may have different exposure histories that affect how high-frequency a word feels to them.
  • A word like SENT may be more frequent for someone who works in shipping than for someone else; thus, frequency effects are context-dependent.
• Lexicon and bilingualism nuances:
  • Some researchers argue for multiple lexicons in bilinguals; others argue for a single lexicon with language-dependent access routes.
  • Access pathways and cross-language activation can influence reading speed and accuracy in bilinguals.

Strengths and limitations of the Interactive Activation Model

• Strengths:
  • Demonstrates a clear example of connectionist processing and how top-down and bottom-up processes interact during reading.
  • Predicts the word superiority effect well and aligns with some empirical data on letter and word recognition.
• Limitations:
  • Does not account for the meaning of words or semantic processing directly.
  • Insufficient emphasis on phonological processing beyond a general role for pronunciation in pseudowords.
  • Focuses on letter position within the word, which may be less critical than assumed (e.g., words can be read with letters scrambled, as in the Cambridge sample).
  • Experimental designs often used four-letter uppercase words; real-world reading uses lowercase and longer words, which IAM does not fully address.
• Practical implications: IAM provides a theoretical basis for reading research and for designing reading instruction and assessment, but it should be supplemented with models that incorporate semantics and longer text processing.
• Looking ahead: two other reading models (not detailed here) will be covered next week, offering additional perspectives on word recognition and reading comprehension.

Semantic priming: meaning-based processing in reading

• Semantic priming definition: processing a target word is facilitated when it is semantically related to a preceding prime word.
• Methods: many studies use a lexical decision task (decide if a string of letters is a word) with a prime word presented beforehand.
• Early findings (1970s): rapid semantic priming occurs when the prime and target are semantically related; this supports automatic activation of related concepts.
• The 1970s prime-time manipulation study (Bird/Body examples):
  • Prime word category or category name is followed by a letter string; participants decide if the letter string is a word.
  • Four conditions to disentangle automatic vs. attention-driven priming:
    1) Expected and semantically related
    2) Unexpected but semantically related
    3) Expected but semantically unrelated
    4) Unrelated and not semantically related (control)
  • Time points studied: 250 ms, 400 ms, 700 ms after the prime.
  • Findings: automatic semantic relations predicted faster and more accurate responses at 250 ms; attention-controlled expectations slowed responses at later times (700 ms). This supports a stronger role for automatic semantic priming than for deliberate attention-driven priming.
• Meta-analysis of semantic priming studies: about 17 studies analyzed; found a small but consistent contextual priming effect across studies; semantic relation between prime and target predicts speed/accuracy better than explicit expectation alone.
• Sentence-level context vs. two-word priming:
  • Sentence context provides richer information, increasing the likelihood and strength of semantic priming effects.
  • Larger context helps reveal top-down processing, world knowledge, and semantic expectations in reading longer texts.
• Predictability and close scores (cloze probability):
  • Cloze probability measures how often readers predict a given next word in a sentence context when given only the initial portion of the sentence.
  • Large samples (e.g., 10,000–20,000 participants) yield stable estimates of predictability as a percentage.
  • Predictable words in a sentence are fixated for shorter times, speeding processing relative to unpredictable words.
• Example sentence task (English, “The day was breezy, so the boy went outside to fly a … in the park”):
  • The choice between an/ a affects expectation and processing; the experiment manipulated whether the next word would plausibly start with a vowel after “an” vs a consonant after “a.”
  • Evoked potential differences observed in brain activity when predictable vs. less predictable options are encountered, illustrating predictive processing in reading.
• Theoretical interpretations of predictability: two main accounts
  • Lexical predictive account: readers activate a single specific word prior to its presentation based on sentence context (highly specific prediction).
  • Graded prediction account: readers generate a broader set of properties or categories (e.g., part of speech, semantic category) without necessarily predicting the exact word.
  • Evidence leaning toward graded predictions: low cost of incorrect lexical predictions and relatively few exact lexical-prediction errors; readers often rely on partial information and graded expectations.
• Practical takeaway: context and semantics drive reading efficiency; top-down processing interacts with bottom-up input to shape recognition speed and accuracy.

Broader implications and connections

• Reading is an interaction between bottom-up letter/feature processing and top-down knowledge (world knowledge, word meaning, context).
• Processing efficiency depends on multiple intersecting factors: word frequency, orthographic neighborhood size, phonological processing, and sentence-level predictability.
• Bilingualism and cross-language effects complicate lexical access and priming effects, suggesting variability in lexicon structure and access routes across individuals.
• Real-world reading (continuous text) relies heavily on semantic priming, predictability (cloze probability), and contextual cues to guide comprehension and speed.
• Educational and cognitive implications:
  • Emphasize strategies that leverage context and semantic scaffolding to improve reading fluency.
  • Consider phonological processing training to support pseudoword decoding and decoding of unfamiliar words.
  • Be aware of individual differences in word frequency exposure and lexical knowledge when assessing reading skills.

Summary of key concepts and terms

• Word superiority effect (WSE): Real words facilitate faster and more accurate letter identification than non-words in brief letter-string tasks.
• Pseudoword superiority effect (PSE): Pronounceable non-words (pseudowords) can also facilitate processing due to phonological processing.
• Interactive Activation Model (IAM): A three-level, parallel-processing model of reading with feature, letter, and word levels; combines bottom-up and top-down processing.
• Orthographic neighborhood: Set of words differing by one letter; neighborhood frequency can speed up or slow down word recognition depending on neighbor frequencies.
• Lexicon: Mental dictionary of words; access can be language-specific in bilinguals.
• Semantic priming: Faster processing of a target word when preceded by a semantically related prime.
• Lexical decision task: Experimental task of deciding whether a string is a real word.
• Cloze probability (close scores): Predictability measure of how likely readers are to guess the next word in a sentence.
• Evoked potential: Brain response measured in experiments; used to study the neural correlates of prediction and processing speed.
• Cambridge reading sample: Classic demonstration that readers can recognize words with jumbled middle letters, highlighting top-down influences in reading.
• Limitations of IAM: Lacks a robust treatment of word meaning, phonology, and longer-word processing; mostly tested with short, uppercase words (4 letters).

Connections to prior and future topics

• Earlier lectures on visual processing and the role of top-down vs bottom-up processing align with the IAM framework.
• The subsequent two reading models (to be covered next week) will address semantics and longer-word/paragraph processing, providing alternatives to IAM.
• Practical implications for design of reading materials and remediation for reading difficulties through an understanding of how context and phonology influence recognition.

Practical practice prompts (quick study aids)

• Explain how the word superiority effect would manifest in a pattern-mask task with a real word vs a non-word string.
• Describe how orthographic neighbors can both help and hinder word recognition, depending on neighbor frequency.
• Compare the automatic vs attention-controlled aspects of semantic priming in the 1970s experiments and summarize what the time-course data suggest.
• Discuss why the Cambridge reading sample challenges a strictly bottom-up view of word recognition.
• Outline the main strengths and limitations of the Interactive Activation Model and what aspects of reading it cannot capture well.