# Graded and Ungraded Expectation Patterns: Prediction Dynamics During Active Comprehension
* **Authors:** Melinh K. Lai | Brennan R. Payne | Kara D. Federmeier
* **Journal:** Psychophysiology (2023)
* **DOI:** 10.1111/psyp.14424
---
# Overview
* **Objective:** Determine how passive versus active prediction tasks modulate neural signatures of prediction during sentence comprehension.
* **Methods:** Event‐related potentials (ERP) recorded while participants read sentences under two task conditions and classified their own predictions.
* **Key Findings:**
* N400 amplitudes show graded cloze effects, amplified under active prediction.
* Anterior positivity (AP) indexes violated predictions and **remains stable across tasks**.
* Frontal negativity (FN) indexes successful predictions, present only during active prediction.
p. 1
---
# Introduction: Context in Comprehension
* Language comprehension leverages contextual information to facilitate word processing.
* Cloze probability: a behavioral measure quantifying word predictability in context.
* **N400 component:** inversely related to cloze; lower amplitudes for higher predictability.
p. 2
---
# Introduction: Multiple Mechanisms
* Contextual facilitation arises from multiple neural processes:
1. **Predictive pre‐activation** of upcoming word features (related‐anomaly paradigms reveal N400 reductions).
2. **Non‐predictive facilitation** (semantic priming, repetition effects).
* ==Prediction effects on N400 vary with task, age, literacy, and trial‐by‐trial engagement.==
* **Raises question: To what extent do predictive processes versus general semantic facilitation contribute to N400 and other ERP effects?**
p. 2, §1.1
---
# Background: N400 Component
* Elicited 300–500 ms post word onset, maximal at central–parietal sites.
* Reflects **semantic access**: mapping stimuli to long‐term memory.
* Sensitive to:
* Repetition & associative priming → reduced amplitude.
* Cloze probability → lower N400 for more predictable words.
p. 2-3, §1.1.1
---
# Background: Anterior Positivity
* Observed 500–800 ms over frontal electrodes.
* Elicited by **unexpected but plausible** words in **strongly** constraining contexts (prediction violations).
* Hypothesized functions:
* Message‐level representation **revision**.
* Inhibition or **suppression** of the predicted word.
* Varied by individual factors (absent in older adults, low‐literacy groups).
* **Contradiction**: Graded v.s. State
p. 3–4, §1.1.2
---
# Background: Frontal Negativity
* Also occurs 500–800 ms frontal, but negative polarity.
* Larger for **expected** endings in **strongly** versus weakly constraining contexts.
* Might index confirmation of specific prediction or selection among competitors.
* ==Shows **different** sensitivity patterns from anterior positivity.==
p. 4–5, §1.1.3
---
# Gap & Rationale
* Both anterior positivity and frontal negativity share timing and distribution but **may reflect distinct processes**.
* Previous active‐prediction paradigms (e.g., Brothers et al., 2017) **couldn't dissociate** them due to stimulus limitations.
* Present study uses a broader constraint range (Federmeier et al., 2007 stimuli) to separate effects.
p. 5–6, §1.2
---
# Study Aims
1. **Aim 1:**
* Investigate whether anterior positivity scales with prediction effort (passive vs. active tasks).
3. **Aim 2:**
* Investigate whether the anterior positivity and the frontal negativity
**are actually separable from each other** or whether they
may in fact stem **from the same mechanism**.
3. **Aim 3:**
* Link ERP patterns to trial‐by‐trial participant classifications (exact match, similar, unexpected).
p. 6, §1.2
---
# Participants
* **N = 40** undergraduates (22 women), age 18–21.
* Right‐handed, monolingual English speakers, normal vision.
* Data from 2 excluded (artifacts/equipment).
* **Power analysis:** Based on d = .48 (Lai et al., 2021), α = .05, power = .80 → N ≥ 36.
>:melting_face:*A power analysis of the frontal positivity effect from Lai et al. (2021),which compared SCU items to combined WC items,found an effect size d of 0.48. With a significance criterion of α = 0.05 and power = 0.8, the minimum sample size needed for this effect was N = 36.* :melting_face:
p. 6, §2.1
---
# Materials: Sentence Stimuli
* **282 sentences** from Federmeier et al. (2007).
* **4 conditions (2×2):**
* SCE: Strong constraint, Expected ending (mean cloze = 85%)
* SCU: Strong constraint, Unexpected ending (cloze ≈ 1%)
* WCE: Weak constraint, Expected ending (mean cloze = 27%)
* WCU: Weak constraint, Unexpected ending (cloze ≈ 1%)
p. 7, §2.2 & Table 1
---
# Design Overview
* Within‐subjects 2×2 design: Constraint × Expectancy.
* Sentences split into two lists; each participant sees each context once.
* Blocks (Passive, Active) counterbalanced.
* Stimuli matched on length and frequency across conditions.
:point_right: each block contained 35 or36 sentences of each type
p. 7-8, §2.2
---
# Procedure: Passive Reading
* Fixation cross (500 ms), blank (850 ms) before sentence.
* RSVP display: 200 ms word, 300 ms interstimulus
* After sentence: prompt "Press button to continue."
* Emphasis: read for comprehension; occasional memory test.
p. 7-8, §2.3
---
# Procedure: Active Prediction
* Instructions: guess final word while reading.
* 800 ms interstimulus
* After final word, rate prediction:
* Exact match、Similar meaning、Completely different
* Blinks allowed during rating.
p. 7-8, §2.3
---
# O‐Span & Recognition Tasks
* **O‐Span:** Letters + math task between blocks to shift cognitive set.
* **Recognition:** 300 words; 100 from passive, 100 active, 100 new foils.
* Participants circle recognized final words.
```mermaid
flowchart TD
A[Start] --> B[Show fixation<br/>500 ms]
B --> C[Show blank screen<br/>850 ms]
C --> D{Task type}
D -->|Passive reading| E1[RSVP<br/>200 ms display + 300 ms ineterstimuli]
D -->|Active prediction with instruction| E2[RSVP<br/>200 ms display + 800 ms ineterstimuli]
E1 --> F1[Show “Press button to continue”]
E2 --> F2[Show “Did the sentence match your expectation?”]
F1 --> G1[Participant presses button]
G1 -->|Memory test| H
F2 --> G2[Classify prediction accuracy<br/>Exact match, Similar meaning, Completely different]
G2 --> H[O-Span task]
H --> I[Another task]
I --> J[End trial]
```
p. 7-8, §2.3
---
# EEG Recording
* 26 electrodes, referenced to left mastoid (reanalyzed offline to averaged mastoids).
* EOG electrodes monitor blinks/eye movements.
* Sampling rate = 1 kHz; 0.016–250 Hz bandpass.
* Offline: 30 Hz low‐pass, trial rejection (\~14% trials excluded).
p. 8, §2.4
---
# ERP Analysis
* Epoch window: −200 to 1000 ms.
* Baseline: −200 to 0 ms.
* **N400:** 300–500 ms; 6 centroparietal channels.
* **Late anterior effects:** 500–800 ms; 11 frontal channels.
* **Conditions & Comparisons:**
* N400: Compare SCE, WCE, Unexpected words.
* Late Anterior: SCU & SCE vs. Combined WC (Weak Constraint) items.
* WC as Baseline: Prior studies showed no WCE/WCU difference.
p. 8-9, §2.5 & Fig. 1
---
# Results
## **Behavioral recognition**
* **Overall d′ = .81** indicates reliable memory for sentence‐final words.
`d' = Z(hit rate) - Z(false alarm rate)`
* Recognition higher in active vs. passive blocks.
* SCE words recognized less often than other conditions.
p. 9, §3.1 & Table 2
---
# ERP Results: Anterior Positivity
* **Compare SCU vs. combined WC (WCE+WCU)**.
* SCU elicits increased frontal positivity (0.53 μV), p = .02.
* repeated-measures ANOVA → ==No Block × Condition interaction → **stable across tasks**.==
p. 9-11, §3.2.1
---
# ERP Results: Frontal Negativity
* **Compare SCE vs. combined WC.**
* SCE elicits increased frontal negativity (0.65 μV), p = .02.
* Interaction: ==effect present only during active prediction== (1.47 μV), p < .001.
p. 9-11, §3.2.1
---
# Block × Condition Interactions
* **Anterior positivity**: consistent magnitude in passive & active.
* **Frontal negativity**: absent in passive, robust in active.
* Suggests prediction violation is automatic; prediction success requires active engagement.
p. 9-11, §3.2.1
---
# ERP Results: N400 Effects
* **repeated-measures ANOVA :** Constraint × Block × Ending.
* ` The interaction between block and constraint and
the three-way interaction between block, constraint, and
ending were not significant (p's > .05)`
* Graded pattern: SCE < WCE < Unexpected.
* Expectancy effects larger in **active block**.
p. 11, §3.2.2
---
# Participant Classification Analysis
* Trials sorted by participant’s own ratings:
1. Exact match 2. Similar meaning 3. Unexpected
* Allows linking ERP effects to **subjective prediction** rather than cloze only.
* Classifications for SC sentences were largely **aligned with** the experimental designations
p. 11-12, §3.2.3
---
# Frontal Effects by Classification (Exact Matches)
* **SCE vs. WCE within exact‐match bin**.
* SCE shows greater frontal negativity (1.89 μV), p = .02.
* Confirms frontal negativity reflects **true** prediction success.
> constraint/cloze probability modulates post-N400 responses even among items that are all classified by participants as being a match to their expectations.
>
p. 12-13, Fig. 4a
---
# Frontal Effects by Classification (Unexpected Ratings)
* **SCU vs. combined WC within unexpected bin**.
* SCU shows greater anterior positivity (0.93 μV), p = .02.
* Confirms anterior positivity reflects **actual** prediction violations.
p. 12-13, Fig. 4c
---
# N400 by Classification (Exact Matches)
* No difference observed in the N400 (300–500 ms) time window (?)
* The global constraint of the sentence did not affect facilitation on the N400
* After the N400 there is an apparent **posterior positivity (?)**
p. 12-13, Fig. 4e
---
# N400 by Classification
* **WCE vs. WCU in “similar meaning” bin**.
* WCE shows reduced N400 (2.48 μV), p = .01.
* Indicates semantic facilitation based on **cloze** persists even when ending is judged unexpected but related. (==a global effect of cloze probability on the N400==)
p. 12-13, Fig. 4e
---
# Key ERP Findings Overview
* **N400**
* Modulated by cloze probability (graded effect).
* Larger in **active prediction** condition.
* Strongest for **strongly constraining, expected completions (SCE)**.
* **Anterior positivity**
* Elicited by **strongly constraining, unexpected completions (SCU)**.
* **Not modulated by task**, suggesting it reflects automatic violation response.
* **Frontal negativity**
* Present primarily in **active prediction** blocks.
* Most pronounced for **SCE** items.
* Suggests sensitivity to **prediction confirmation**.
p. 13-14
---
# Interpretation – Anterior Positivity
* Elicited when a **strongly predicted item is not confirmed (but plausible)**.
* May reflect the **cost of suppressing or reanalyzing** the unfulfilled prediction.
* ` The lack of any late anterior effects in
weakly constraining sentences here does not immediately support inhibition of the predicted word`
* `...the results may support an account that links the anterior positivity to revision of the message representation... `
* **Unaffected by task**, supporting the view that it is **automatic and stimulus-driven**.
p. 14-15, §4.1.1
---
# Interpretation – Frontal Negativity
## Frontal Negativity: Beyond Anterior Positivity
* Occurs only in the **active prediction** task, indicating **task dependence**.
* **Distinct from Anterior Positivity:**
* Observed in:
* Highly literate young adults (who also show Anterior Positivity).
* Older adults and individuals with lower literacy skills (who often *don't* show Anterior Positivity)
* Thus:
* unlike the anterior positivity, the frontal negativity **does not** seem to be as yoked to prediction
* **Frame-Shifting Account (Wlotko and Federmeier):**
* Good joke comprehenders → Frontal Negativity (Coulson & Kutas, 2001)
* The frontal negativity effect was specific to those encountered in **strongly constraining** sentences, showing that this effect does not simply reflect subjective expectancy.
* **Potential Underlying Mechanisms:**
* **Ambiguity Resolution:**
* Sentences may have multiple possible interpretations.
* Frontal Negativity might reflect processes of **selecting/solidifying** the message-level representation supported by the most globally probable word
* **Deeper Processing:**
* Predictive task demands might **augment the tendency** to more deeply process sentences and consider multiple interpretations.
* Leading to an enhanced effect.
p. 15-16, §4.1.2
---
# Theoretical Implications
* Supports a **multi-component model** of prediction:
* **N400**: reflects the degree of contextual facilitation and **semantic access**.
* **Anterior positivity**: reflects **violation detection**.
* **Frontal negativity**: marks **prediction confirmation**.
* Challenges the idea that prediction is **unitary and continuous**.
* Indicates that **violations and confirmations** may rely on **separate neural processes**.
* **Task demands modulate** ERP effects → importance of **top-down goals**.
---
# Q&A
### 陳高冰貞
**Question:**
> After learning about the two late components (frontal negativity and anterior positivity), I started wondering: is there ==a brain signal that comes even earlier than the N400 — something like the initial "hunch" we feel?== Or could that sense of anticipation come from somewhere else, like the heart or a "gut feeling"?
**Reply:**
:::spoiler Reply
Yes, there are earlier ERP components like the **P200** or **ELAN** that may reflect initial expectations, but the “gut feeling” may also involve body systems like heart rate or the gut–brain axis, not just the brain.
:::
---
### 郭旻芃
**Question 1:**
> Older adults in Wlotko (2012) exhibited the N400 expectancy effect but did not show AP while encountering unexpected words. This raises the question: ==aside from a reduced predictive processing ability, are there any other possible interpretations for the absent of AP in older adults?== For example, could it be that older adults lack sufficient cognitive resources to engage in the subsequent reanalysis process, which would explain the absence of AP following the N400?🍯
**Reply:**
:::spoiler Reply
:::
**Question 2:**
> In Brothers et al. (2017), participants were instructed to do "active prediction" during the experiment. I wonder ==whether such a requirement might be redundant or even interfere with the natural reading process==, since prediction typically occurs automatically during reading.
**Reply:**
:::spoiler Reply
Great point. While prediction is often automatic, asking participants to predict can highlight processes that might otherwise remain hidden—but yes, it may also interfere with natural reading.
:::
**Question 3:**
> I wonder whether the Anterior Negativity (AN) would still be observed when participants encounter an expected sentence ==final word that is spelled incorrectly or written with the wrong characters==. Alternatively, would such stimuli elicit different ERP components, such as the P3b?
**ChatGPT Reply:**
:::spoiler Reply
Misspellings or character errors might not elicit AN or AP but could instead trigger other ERP components like the P3b or early visual/orthographic effects like the N250.
:::
---
### 徐依綾
**Question 1:**
> The article frequently uses '**anterior**' to describe the anterior positivity but refers to the frontal negativity as '**frontal**.' I wonder ==whether there are any reasons for the use of different descriptions in these ERP components==.
**Reply:**
:::spoiler Reply
P.4 *"We use the term “anterior” when referring to the positivity and
“frontal” when referring to the negativity ==because, in the extant
literature, these effects have most often been described as an “anterior
positivity” and a “frontal negativity”==. Although relatively synonymous
with respect to the distributional characteristics implied, we thus
endeavored to remain consistent with the prior literature in our naming
conventions here. Additionally, we felt that maintaining the naming
conventions from the prior literature allowed for better readability in
distinguishing the two patterns from each other."*
:::
**Question 2:**
> I wonder whether the findings of this paper could be applied to other domains, such as in the development of natural language processing systems for modeling human prediction.
**Reply:**
:::spoiler Reply
The distinction between passive comprehension and active prediction in the study is valuable for NLP design.
==AP-> Reanalysis v.s. Deepseek -> safety alignment==
Current models often blend both: they predict (autocomplete) and comprehend (analyze input). Learning from the brain's ability to switch modes depending on task demands might inform adaptive language systems that know **when to predict aggressively and when to slow down and integrate.**
> Ganapini, M.B. et al. (2023). Thinking Fast and Slow in AI: The Role of Metacognition. In: Nicosia, G., et al. Machine Learning, Optimization, and Data Science. LOD 2022. Lecture Notes in Computer Science, vol 13811. Springer, Cham. https://doi.org/10.1007/978-3-031-25891-6_38
:::
---
### 危幼苓
**Question:**
> The frontal negativity to expected words was not apparent during passive comprehension but was greatly enhanced in the active prediction part. ==I wonder whether there are two mechanisms in charge of prediction and comprehension.==
**Reply:**
:::spoiler Reply
That’s very likely. Prediction and comprehension probably rely on partially different mechanisms—prediction may involve control or monitoring systems, while comprehension relies more on integrative processes like the N400.
:::
---
<style>
/* Styling for better readability */
h1, h2, h3 {
color: #2E86C1;
font-family: Arial, sans-serif;
border-bottom: 2px solid #ddd;
padding-bottom: 5px;
}
h1 {
font-size: 28px;
}
h2 {
font-size: 24px;
}
h3 {
font-size: 20px;
}
p {
font-size: 16px;
line-height: 1.6;
}
code {
background-color: #f4f4f4;
padding: 3px 5px;
border-radius: 4px;
}
blockquote {
font-style: italic;
color: #7B7D7D;
border-left: 4px solid #3498DB;
padding-left: 10px;
}
</style>
Sign in with Wallet
Connect another wallet