Reading Models and Eye Movement in Reading (Comprehensive Notes)

Overview

  • Today’s session covers two additional reading models and a look at eye movement/attention research in reading.
  • Focus shifts between single-word reading and sentence-level reading to highlight how context changes processing.
  • Emphasis on reading aloud and the processing steps needed to pronounce unfamiliar words correctly.
  • Introduction to pseudowords and how readers map graphemes to phonemes under time pressure.

Key Concepts in Reading and Pronunciation

  • Reading aloud involves rapid, often error-free pronunciation despite encountering unfamiliar spellings.
  • Grapheme-phoneme mapping is a core mechanism readers rely on, especially for pseudowords.
  • Pseudowords are pronounceable strings that are not real words (e.g., fas or fas vs. fast). They are used in experiments to probe phoneme-grapheme mapping.
  • Word frequency affects pronunciation and recognition; familiar/high-frequency words are processed faster.
  • Regular words map directly from spelling to pronunciation (e.g., cat, fat, bat, lot, hippopotamus).
  • Irregular words do not map straightforwardly from spelling to pronunciation (e.g., yacht, champagne).
  • Non-words can be pronounceable (e.g., a pseudoword) or unpronounceable; pronounceable non-words can be read phonologically but are not real words.
  • Semantic meaning (semantics) can influence how a word is read aloud, especially for irregular or unfamiliar words.

Models of Reading Aloud

  • Two main computational approaches to explain reading aloud:
    • Dual Route Cascaded (DRC) Model
    • Distributed/Connectionist Triangle Model (often called the Triangle Model)

Dual Route Cascaded Model (DRC) — Max Coltheart and colleagues

  • Core idea: reading aloud uses two (often described as three) routes from orthography to phonology.
  • Pathways start with the orthographic input (the printed word).
  • Routes:
    • Route 1: Non-lexical (non-word) pathway
    • Function: map letters to sounds for pronounceable non-words or when no word is recognized.
    • Output: pronunciation via direct grapheme-phoneme conversion.
    • Route 2/3: Lexical pathway(s) for real words
    • Route 2: Direct lexical-to-phonology route (without semantic access)
    • Route 3: Lexical route through semantics to phonology
    • Function: recognize a word as a whole, then access its pronunciation; or, if needed, access its meaning (semantics) before producing pronunciation.
  • Orthographic analysis: initial stage where letters and spelling are analyzed to decide if the string is a real word or a non-word.
  • If a string is a non-word, you follow Route 1 (non-lexical).
  • If a string is a real word, you may follow Route 2 (direct to phonology) or Route 3 (via semantics).
  • “Response buffer” and articulation: once a pronunciation is determined, it is buffered and spoken aloud.
  • Cascaded processing: activation spreads to downstream stages before finishing processing at the current level, allowing rapid, parallel processing.
  • Distinctions within word types:
    • Regular words: pronounced exactly as spelled (e.g., cat, fat, sat, bat, mat; hippopotamus as a long but regular example).
    • Irregular words: pronunciation cannot be predicted purely from spelling (e.g., yacht, champagne).
    • Non-words: pronounceable vs. unpronounceable; readers rely on grapheme-to-phoneme rules for pronounceable non-words.
  • Lexical vs. non-lexical routes: the green route (lexical) involves recognizing a real word and may involve semantic processing; the red route (non-lexical) relies on letter-to-sound mappings for pronounceable strings that are not words.
  • How the model explains dyslexia types:
    • Surface dyslexia: reliance on Route 1 (non-lexical) with poor access to irregular words; relatively good pronunciation of regular words and non-words in some cases.
    • Phonological dyslexia: difficulty mapping letters to sounds for non-words; relatively better reading of high-frequency, regular words that have consistent spellings.
    • Deep dyslexia: not well explained by the DRC; often associated with brain damage; suggests involvement of right-hemisphere language processing in some cases.
  • Practical notes on dyslexia assessment:
    • Comprehensive assessments can determine subtypes (e.g., surface vs. phonological dyslexia) and guide strategies, while screenings may only indicate dyslexia without subtype detail.
  • Limitations of DRC:
    • Does not account well for individual differences in reading strategies.
    • Assumes near-perfect knowledge of letter positions within words; may not reflect real-world variability.
    • Phonological processing speed and the speed of mapping may be faster in practice than the model predicts.
    • English-specific model limitations; not all languages align with the same reading processes.
  • Notable example: Baby X demonstration by Mark Sager (Soul Machines) illustrating a machine learning approach to reading as a human-like process; highlights challenges of teaching machines to read and the role of context in word recognition.

Distributed Connectionist (Triangle) Model

  • Core idea: reading involves continuous interaction among three components: orthography (spelling), phonology (sound), and semantics (meaning).
  • Also a computational model; emphasizes interaction rather than strict routes.
  • Structure: three components connected with bidirectional interactions (double-headed arrows): orthography, phonology, semantics.
  • Reading process:
    • Word is seen; pronunciation can be produced directly via a phonological route, or via semantic access before pronunciation.
    • All three components can influence each other during learning and word recognition.
  • Learning and performance:
    • The network was trained on roughly 3,000 words and could pronounce many words aloud; the model also produced correct pronunciations for a substantial portion of non-words.
    • When encountering inconsistencies in pronunciation (e.g., "though", "bow", "through", "rough"), the model slowed down, mirroring human performance.
    • Infrequent words were slower to pronounce, and the slowdown was greater for inconsistent words; performance for non-words remained relatively high (about 90%) despite slower processing.
  • Relationship to real-world reading:
    • Supports the idea that readers use all three components (orthography, phonology, semantics) rather than relying solely on a phoneme-by-phoneme rule or a purely per-word lookup.
    • Provides a framework for understanding different dyslexia types and their manifestations beyond what the DRC can offer.
  • Brain mapping & neurobiology:
    • Visual processing is primarily posterior in the occipital lobe; phonological processing is more anterior; semantic processing is deep cortical and sits between visual and phonological processing.
  • Strengths of the triangle model:
    • Accounts for interactions among spelling, sound, and meaning; includes explicit learning processes.
    • Provides a mechanism to explain semantic involvement in reading aloud and the effects of word meaning on pronunciation.
  • Weaknesses/limitations:
    • Still emphasizes monosyllabic words in many implementations; may oversimplify how complex words are processed.
    • Explains semantics to some extent but does not fully articulate how meaning drives pronunciation in all contexts.
    • Less explicit about cross-language generalization; primarily developed and tested in English.
  • Strengths vs. DRC:
    • DRC emphasizes regularity and a more rigid route-based structure; Triangle model emphasizes consistency and semantic involvement with richer interactivity.
    • Triangle model highlights learning dynamics and semantic contributions more overtly, whereas DRC centers on route selection and word-type processing.

Eye Movement and the Easy Reader Model

  • Eye movement research asks: how do readers move their eyes (saccades) and where do fixations occur during reading?
  • Easy Reader Model (serial, step-by-step):
    • Assumes readers move their eyes from one group of words to the next in a serial fashion, with fixations in between.
    • Fixation duration: typically around 200−250extms200-250 ext{ ms} per fixation.
    • Saccades cover roughly eight letters or spaces before the eye moves to the next location.
    • Peripheral vision provides information from words to the left and right of the fixation point, enabling some preview processing.
  • Key findings:
    • Reading can proceed smoothly when text is predictable and words are common and consistently pronounced.
    • Predictability speeds up reading; readers anticipate the next word and process more efficiently.
    • When predictability fails (unexpected word), fixation duration increases and readers may regress (move backward) to reread.
  • Demonstrations and real-world examples:
    • An example sentence about Crystal Palace shows how readers’ fixations are influenced by semantic context and predicted content.
    • Readers often fixate on a central word while using peripheral vision to glean surrounding words.
  • Predictability and sign-reading examples:
    • People often read signs with errors or unusual phrasing because of top-down expectations about meaning; top-down expectations can override actual orthography in perceived meaning.
  • Limitations of the Easy Reader Model:
    • Works best for short, simple, easily understood words; not as accurate for long, multi-syllabic words.
    • Better suited to competent readers; may not accurately reflect learning to read or struggling readers.
    • English-centric; does not generalize seamlessly to other languages with different orthographies or reading directions.
  • Relationship to the other models:
    • The Easy Reader Model aligns with serial processing assumptions; in contrast, the Triangle Model and DRC emphasize interactive or parallel processing and multiple pathways.

Dyslexia, Consistency, and Reading Speed

  • Consistency matters: faster reading and higher accuracy when pronunciation is consistent and predictable.
  • Regularity vs. consistency:
    • DRC emphasizes regularity (how words map to pronunciation directly from spelling).
    • Triangle Model emphasizes consistency (how reliably pronunciation maps across contexts).
  • Reading speed findings:
    • When pronunciation is consistent across exposure, reading speed tends to be faster; inconsistent pronunciations slow reading more, especially for infrequent words.
  • Non-words and reading aloud:
    • DRC is relatively inflexible in mapping non-words to sounds.
    • Triangle Model is more flexible, drawing on individual experiences and past knowledge to map non-words.
  • Dyslexia subtypes in practice:
    • Surface dyslexia: performs better on regular words and non-words, struggles with irregular words; often relies on letter-to-sound rules.
    • Phonological dyslexia: strong with familiar words, difficulty reading non-words or pseudo-words; relies less on phoneme-grapheme rules for novel strings.
    • Deep dyslexia: usually associated with brain damage; may involve right-hemisphere language processing; not well captured by the DRC.
  • Assessments:
    • Subtype diagnosis and detailed strategies depend on the comprehensiveness of the assessment.
    • Not all assessments differentiate subtypes; some are screenings that confirm a dyslexia diagnosis but don’t specify subtype.
  • Real-world implications:
    • Understanding dyslexia helps tailor strategies for decoding, reading aloud, and comprehension in classrooms and daily life.
    • Helpful for designing educational tools and interventions that address both phonological decoding and semantic processing.

Brain and Learning Implications

  • Brain-behavior mapping:
    • Visual processing: posterior occipital regions; reading begins with recognizing letter shapes.
    • Phonological processing: more anterior regions; mapping letters to sounds.
    • Semantic processing: deep cortical areas, bridging meaning with form and sound.
  • Implications for education and assistive tech:
    • Models support different instructional approaches, from grapheme-phoneme drills to semantic enrichment and context-based reading strategies.
    • Technologies that support reading (e.g., prediction, text completion) align with the idea that readers use expectations and meaning to facilitate processing.

Practical Examples and Takeaways

  • Regular vs. irregular words: expect straightforward pronunciation for regular words; irregulars require lexical/semantic knowledge or memorization.
  • Pseudowords: test of pure phoneme-grapheme mapping without semantic cues; readers rely on rules and exposure to pronounce them.
  • Word frequency effects: high-frequency words are read more quickly; rare or infrequent words slow readers, especially when pronunciations are inconsistent.
  • Consistency effects: readers are faster when pronunciation rules are consistent across contexts and similar words.
  • Sign reading and errors: reading can be influenced by expectations; people often misread signs due to top-down processing.
  • Eye movement: readers process chunks with saccades and fixations; predictability and word familiarity influence fixation duration and regression.

Connections to Prior Material and Real-World Relevance

  • Reading models connect to broader cognitive psychology ideas about how perception, language, and memory interact during complex tasks.
  • The dual-route and triangle models illustrate competing theories of reading that explain both accuracy and speed, and help interpret dyslexia profiles.
  • Eye-tracking research links to theories of attention, prediction, and peripheral processing in real-time tasks beyond reading.
  • Real-world relevance includes education, reading assessment, and the design of clearer signs and instructional materials to support diverse readers.

Summary Takeaways

  • Reading aloud involves rapid, multi-source processing: orthography, phonology, and semantics interactively (Triangle Model) or via multiple routes (DRC).
  • The DRC proposes a flexible system with a non-lexical route and lexical routes (with or without semantic access); it accounts for surface and phonological dyslexia but has limitations, including cross-language generalizability.
  • The Triangle Model emphasizes continuous interaction among orthography, phonology, and semantics, better capturing learning, inconsistency effects, and semantic involvement in reading aloud.
  • Eye movement research (Easy Reader Model) suggests serial processing with predictable benefits from word frequency and context, but has limits for complex texts and non-native readers.
  • Across models, consistency, frequency, predictability, and semantic context all shape how we read aloud and how quickly we do so, with clear implications for education and practice.
  • Finally, readings signs and real-world text reveal how top-down expectations can influence even seemingly straightforward decoding tasks.

Notable Names and Examples Mentioned

  • Max Coltheart (Macquarie University) – key figure behind the Dual Route Cascaded model.
  • Mark Sager – creator of Baby X and Soul Machines demonstration of machine reading.
  • Examples used:
    • Regular vs. irregular words: cat vs. yacht, champagne; regular long word hippopotamus; non-words like fas, keck, miz, blut.
    • Pseudoword pronunciation ambiguity: fas vs. fast.
    • Inconsistent pronunciation examples: though, bow, through, rough (same “ough” but different pronunciations).
    • Non-word reading performance: Triangle model could pronounce ~3000 words and achieve near 90% accuracy on non-words under training conditions.

Key Formulas and Numerical References

  • Pseudowords pronunciation relies on grapheme-phoneme mapping rules: not a fixed real-word mapping but rule-based pronunciation.
  • Model performance metrics:
    • DRC-like models trained on about 8,0008{,}000 words; achieved extaccuracy ≈ 99%ext{accuracy} \, \approx \, 99\% for single-syllable words and non-words under certain pronounceability constraints.
    • Triangle model: accuracy on non-words ~90%90\% with some slowing on inconsistent words or infrequent words.
  • Eye-tracking measures:
    • Fixation duration: 200 ms≤extfixationduration≤250 ms200\text{ ms} \le ext{fixation duration} \le 250\text{ ms}
    • Saccade span: about 88 letters or spaces per saccade.

Closing Note

  • The three models and eye-movement research together provide a nuanced picture of reading. They offer complementary explanations for how we decode text, how meaning enters the process, and how reading speed is shaped by regularity, consistency, frequency, and prediction. Ongoing research continues to refine these models and their applicability across languages and populations, including dyslexia subtypes and reading development.