Neural Sensitivity to Word Frequency Modulated by Morphological Structure: Univariate and Multivariate fMRI Evidence from Korean

This fMRI study on Korean word recognition demonstrates that neural sensitivity to surface frequency, rather than obligatory pre-lexical decomposition into stems and affixes, drives the processing of morphologically complex words, suggesting that morphological access is jointly constrained by structural and usage-based statistical factors.

The Gist

The Big Question: Do We Read Words Like Lego or Like Photos?

Imagine you are reading a word on a sign. When your brain sees a complex word like "HUNTER," does it instantly recognize it as a single, familiar picture (a photo of a hunter)? Or does it automatically break it apart into its building blocks, "HUNT" + "ER," before understanding what it means?

For decades, scientists have debated this. Some say our brains are like Lego masters that always take complex words apart first. Others say our brains are like photo albums that recognize frequent words as whole images, only breaking them down if they are rare or strange.

This study set out to solve this mystery by looking inside the brains of Korean speakers while they read words.

The Perfect Test Case: Korean as a "Transparent" Language

Why Korean? Think of Korean as a very clear, transparent language.

• In English, the word "corner" has "corn" in it, but "corn" doesn't really mean the same thing. It's a messy puzzle.
• In Korean, words are built like clear glass blocks. You have a main block (the stem) and you snap a smaller block (a suffix) onto the end to change the meaning (like adding "the" or "to" in English).

Because the blocks are so clear, the researchers could create a perfect experiment. They could compare:

1. Simple Words: Like "Corn" (one block).
2. Inflected Words: Like "Corn-on-the-table" (two blocks snapped together).

They could then ask: Does the brain care more about how often the whole phrase is used, or how often the main block is used?

The Experiment: A Speedy Game in the MRI Machine

The researchers put 21 Korean speakers in an MRI machine (a giant camera that takes pictures of brain activity). They played a quick game:

• A word would flash on the screen.
• The person had to press a button to say, "Is this a real word or a fake one?"
• The words varied in how common they were. Some were super popular (like "Hello"), and some were rare.

They used two types of analysis:

1. The "Volume" Check (Univariate): How loud is the signal in a specific brain area? (Is the brain working harder?)
2. The "Pattern" Check (RSA): This is the fancy part. Instead of just looking at volume, they looked at the shape of the activity. If you see two similar words, do the brain patterns look similar? If you see two very different words, do the patterns look different? It's like checking if the brain has a unique "fingerprint" for every word.

The Results: The Brain Loves the "Whole Package"

Here is what they found, and why it matters:

1. The "Whole Word" Wins
The brain didn't seem to care much about the frequency of the main block (the stem). Instead, it was obsessed with the frequency of the entire word.

• Analogy: Imagine you have a favorite coffee shop. If you go there every day (high frequency), you recognize the whole building instantly. You don't stop to count the bricks or analyze the door handle.
• The study found that for complex words, the brain treats them like that favorite coffee shop. If the whole phrase is common, the brain accesses it as a single unit.

2. The "Lego" Theory is Wrong (for frequent words)
If the "Lego" theory were true, the brain should have been very active analyzing the main block ("HUNT") regardless of whether "HUNTER" was common or rare.

• The Finding: The brain ignored the main block's frequency. It only cared about the whole word's frequency.
• Where did this happen? In the Frontal Lobe (the brain's executive office) and the Parietal Lobe (the brain's integration hub). These are areas responsible for retrieving meaning and making decisions, not just breaking things apart.

3. Complexity Makes the "Whole Word" Even Stronger
Interestingly, the effect was stronger for the complex words (the two-block words) than the simple ones.

• Analogy: It's like a VIP pass. The more complex the word is, the more the brain relies on its "memory" of the whole thing if it's a familiar word. The brain says, "I know this whole package! I don't need to take it apart."

The Conclusion: We Are Flexible, Not Rigid

The study concludes that our brains are smart and flexible, not rigid robots.

• Old View: Our brains are like a factory assembly line that always takes words apart first.
• New View: Our brains are like a smart librarian.
  • If a book is very popular (high frequency), the librarian grabs the whole book off the shelf instantly.
  • If a book is rare or weird, the librarian might have to open it and look at the chapters (decompose it) to understand it.

The Takeaway:
We don't have to break words down to understand them. If a complex word is common enough, our brains store it as a single, solid unit. We only break it apart when we need to. This suggests that our language skills are built on experience and statistics (how often we see things) rather than just rigid rules.

In short: Familiarity breeds wholeness. The more you see a complex word, the more your brain treats it as a single, unbreakable unit.

Technical Summary

1. Problem and Research Question

The central debate in psycholinguistics regarding visual word recognition concerns whether morphologically complex words are obligatorily decomposed into stems and affixes before meaning access (pre-lexical decomposition) or whether they can be accessed as whole-word units (post-lexical access), particularly when they are frequent and familiar.

• The Conflict: Pre-lexical models predict that neural processing should be driven by base frequency (the cumulative frequency of the stem), as the stem is accessed first. Conversely, post-lexical models predict that surface frequency (the frequency of the whole word form) should drive neural responses, as frequent forms are stored and retrieved as holistic units.
• The Gap: While behavioral studies suggest surface frequency often dominates, neuroimaging evidence has been mixed. Few studies have directly contrasted surface and base frequency effects in a single language system where these variables can be cleanly dissociated.
• The Specific Context: The study focuses on Korean nominal inflection. Korean's agglutinative structure features invariant stems and transparent suffixes (case/topic markers), allowing for a precise separation of stem frequency (base) and whole-word frequency (surface).

2. Methodology

Participants

• Sample: 25 native Korean speakers recruited initially; final neuroimaging sample was N = 21 (11 female, mean age ~26.4 years) after exclusions for poor behavioral performance or data corruption.
• Criteria: Right-handed, normal/corrected vision, no neurological history.

Stimuli Design

• Total Items: 120 real Korean nouns and 120 pseudowords.
• Conditions:
  • Simple Nouns (N=60): Monomorphemic (e.g., 옥수수 "corn"). Here, surface frequency equals base frequency.
  • Inflected Nouns (N=60): Stem + Case/Topic particle (e.g., 까닭은 "reason-TOPIC"). Here, surface frequency and base frequency are distinct.
• Frequency Manipulation:
  • Surface Frequency (Zipf-word): Log-transformed frequency of the whole form.
  • Base Frequency (Zipf-stem): Log-transformed frequency of the stem.
  • Control: Word length (syllables) and first-syllable frequency were matched across conditions.
• Task: Rapid event-related Lexical Decision Task (LDT). Participants judged if a string was a real word or a pseudoword.

Imaging and Analysis

• Scanner: 3T Siemens Magnetom Trio.
• Regions of Interest (ROIs): Six left-hemisphere language regions were defined using the Harvard-Oxford Atlas:
  • Inferior Frontal Gyrus (pars opercularis & triangularis)
  • Supramarginal Gyrus (SMG) & Angular Gyrus (AG)
  • Middle Temporal Gyrus (MTG) & Fusiform Gyrus (FuG)
• Analytical Approaches:
  1. Univariate Analysis: Linear mixed-effects models (LMMs) on trial-wise beta values to test for main effects of frequency and interactions with morphological condition.
  2. Representational Similarity Analysis (RSA): Multivariate approach correlating neural dissimilarity matrices (RDMs) with theoretical RDMs derived from surface and base frequency distances. This tests whether the pattern of activation reflects frequency statistics.

3. Key Results

Behavioral Performance

• Both surface and base frequency facilitated reaction times (faster responses for higher frequency).
• No significant interaction was found between morphological condition (simple vs. inflected) and frequency in behavioral data, suggesting both frequency types aid recognition efficiency.

Univariate fMRI Results

• Surface Frequency: Showed a significant main effect in the Inferior Parietal Lobule (IPL), specifically the Angular Gyrus (AG) and Supramarginal Gyrus (SMG).
• Interaction: A significant Condition × Surface Frequency interaction was observed in the SMG. Activation increased with surface frequency for inflected nouns but not for simple nouns.
• Base Frequency: Showed no reliable positive effects in any ROI.
• Whole-Brain: Both conditions activated a left-lateralized fronto-temporo-occipital network, but the direct contrast (Inflected > Simple) yielded no suprathreshold clusters, suggesting similar overall system engagement.

Multivariate (RSA) Results

• Surface Frequency Encoding: Robust positive correlations were found between neural patterns and surface frequency models in the IFG pars opercularis (IFGoper) and SMG. Crucially, this encoding was significantly stronger for inflected nouns than for simple nouns.
• Base Frequency Encoding: No ROI showed significant positive encoding of base frequency. In fact, simple nouns showed negative correlations with base frequency in IFGoper and SMG.
• Conclusion: The neural representational geometry in frontoparietal regions tracks whole-word statistics, not stem-level statistics.

4. Key Contributions

1. Dissociation of Frequency Types: The study provides the first fMRI evidence in Korean explicitly dissociating surface and base frequency effects, demonstrating that the brain encodes surface frequency for inflected words while ignoring base frequency in sustained neural responses.
2. Multivariate Evidence: By using RSA, the authors show that the structure of neural representations in the IFG and IPL is shaped by whole-word usage statistics, challenging the idea that morphological complexity necessitates obligatory stem-based decomposition.
3. Regional Specificity: The findings localize frequency sensitivity to IFGoper and IPL (SMG/AG), regions associated with lexical retrieval and semantic integration, rather than early visual/orthographic processing regions (like the fusiform gyrus), suggesting these effects occur at a post-lexical stage.
4. Behavioral-Neural Dissociation: The study highlights that while both frequency types aid behavioral speed, only surface frequency shapes the underlying neural representational geometry, indicating that behavioral metrics may mask distinct representational mechanisms.

5. Significance and Theoretical Implications

• Refutation of Obligatory Decomposition: The results argue against strict pre-lexical decomposition models (e.g., Taft & Forster, 1975), which predict that stem frequency should drive early activation. Instead, the data supports post-lexical or hybrid models (e.g., Augmented Addressed Morphology, AAM).
• Usage-Based Processing: The findings support the view that morphological processing is distributional and usage-driven. High-frequency inflected forms become entrenched as independent lexical entries (whole-word representations) rather than being decomposed on every encounter.
• Cross-Linguistic Generalizability: The study demonstrates that the dominance of surface frequency in morphological processing, previously observed in Indo-European languages, also holds for Korean, a language with a transparent agglutinative structure. This suggests that the mechanism of frequency-driven access is a universal property of the visual word recognition system, regardless of morphological type.
• Methodological Advancement: The combination of univariate and RSA approaches in a single design offers a more nuanced view of how lexical statistics constrain neural activity, moving beyond simple "activation vs. deactivation" to understanding the geometry of neural representations.

Limitations Noted: The fMRI temporal resolution cannot capture rapid pre-lexical segmentation (which may occur transiently <200ms, as seen in ERP studies). The study relies on a lexical decision task, which may differ from natural reading. Future work combining EEG/MEG with fMRI is recommended to map the full temporal dynamics of decomposition vs. whole-word access.