Contextualized sensorimotor norms: Multi-dimensional measures of sensorimotor strength for ambiguous English words, in context.

Code ↔ Paper

6 matches between paragraphs of the paper and lines of its authors' code, computed by the harvester (lexical-v1). Click a colored paragraph or line to see its counterpart.

The 6 matches

[1] § Case study: Can large language models help scale the CS norms? › Between-dimension comparison ↔ src/analysis/contextualized_norms_analysis.Rmd, lines 2313–2381 · score 0.67 · Mantel, permutations, pairwise, Pearson, matrices, Spearman
[2] § Contextual variance in sensorimotor associations › Sense dominance and deviation from the Lancaster norms ↔ src/analysis/contextualized_norms_analysis.Rmd, lines 1170–1200 · score 0.66 · decontextualized LS norm, cosine distance, sensorimotor properties, lme4, dominant sense, closer
[3] § Comparing distributional and sensorimotor similarity › Methods › Calculating distributional distance ↔ src/analysis/contextualized_norms_analysis.Rmd, lines 1254–1326 · score 0.65 · DistilBERT, RoBERTa, ALBERT, RAW, distance, models
[4] § Contextual variance in sensorimotor associations › Deviations from the Lancaster sensorimotor norms ↔ src/analysis/contextualized_norms_analysis.Rmd, lines 766–891 · score 0.64 · standardized absolute deviation, maximally, floor, Olfaction, SD, Taste
[5] § Comparing distributional and sensorimotor similarity › Sensorimotor distance and ambiguity processing › Explanation of experimental materials ↔ src/analysis/contextualized_norms_analysis.Rmd, lines 1870–1918 · score 0.62 · primed sensibility judgment, Log RT, aggregated, accuracy, asked, word
[6] § Comparing distributional and sensorimotor similarity › Sensorimotor distance and ambiguity processing › Evaluating sensorimotor vs. distributional distance ↔ src/analysis/contextualized_norms_analysis.Rmd, lines 1870–1918 · score 0.56 · primed sensibility judgment, Log RT, accuracy, sensorimotor distance

Paper

Loaded from Europe PMC by your browser, not stored by OSCR: doi.org · Europe PMC

The paper is loaded when this pane is shown.

The paper is loaded by your browser, with JavaScript: read it at doi.org or on Europe PMC.

The authors' code

R Markdown · 2,607 lines · 81 KB · no license · 6 matches

---
title: "Analysis of contextualized sensorimotor norms"
author: "Sean Trott"
date: "January 19, 2026"
output:
html_document:
keep_md: yes
toc: yes
toc_float: yes
---
```{r setup, include=FALSE}
knitr::opts_chunk$set(dpi = 300, fig.format = "png")
```
```{r include=FALSE}
library(tidyverse)
library(lme4)
library(ggridges)
library(broom.mixed)
library(lmerTest)
library(ggcorrplot)
library(kableExtra)
```
# Introduction
In this document, we analyze the **contextualized sensorimotor norms**: judgments about the strength of different sensorimotor dimensions of ambiguous words, in context.
We use these norms in several analyses:
1. First, we compare them to the corresponding dimensions for the "decontextualized" Lancaster sensorimotor norms.
2. Second, we ask whether the **dominance** of a word sense is correlated with its sensorimotor strength, i.e., whether more concrete meanings tend to be rated as more dominant.
3. Third, we ask whether the **sensorimotor distance** between two contexts of use predicts judgments of how **related** those meanings are, above and beyond their distributional similarity and whether or not they belong to the same sense.
4. Fourth, we use **sensorimotor distance** to predict behavior on a primed sensibility judgment task.
# Characterizing dimensions
First, load the data.
```{r}
# setwd("/Users/seantrott/Dropbox/UCSD/Research/Ambiguity/SSD/cs_norms/src/analysis")
df_contextualized_meanings = read_csv("../../data/processed/contextualized_sensorimotor_norms.csv")
nrow(df_contextualized_meanings)
```
## Visualizing distributions
```{r different_dimensions}
df_contextualized_meanings_long = df_contextualized_meanings %>%
pivot_longer(cols = c(Vision.M, Hearing.M, Olfaction.M,
Taste.M, Interoception.M, Touch.M,
Mouth_throat.M, Head.M, Torso.M,
Hand_arm.M, Foot_leg.M),
names_to = "Dimension",
values_to = "Strength") %>%
mutate(Dimension = gsub('.M', '', Dimension))
df_contextualized_meanings_long %>%
ggplot(aes(x = reorder(Dimension, Strength),
y = Strength)) +
geom_violin() +
geom_jitter(alpha = .1,
width = .1) +
coord_flip() +
labs(y = "Sensorimotor strength",
x = "Dimension") +
theme_bw() +
theme(text = element_text(size=20))
```
## Correlations across dimensions
```{r corr_matrix}
columns = df_contextualized_meanings %>%
mutate(Vision = Vision.M,
Hearing = Hearing.M,
Olfaction = Olfaction.M,
Taste = Taste.M,
Interoception = Interoception.M,
Touch = Touch.M,
Mouth_throat = Mouth_throat.M,
Head = Head.M,
Torso = Torso.M,
Hand_arm = Hand_arm.M,
Foot_leg = Foot_leg.M) %>%
select(Vision, Hearing, Olfaction,
Taste, Interoception, Touch,
Mouth_throat, Head, Torso,
Hand_arm, Foot_leg)
cors = cor(columns)
# cors[lower.tri(cors, diag=TRUE)] <- 0
# Plot the correlation matrix
ggcorrplot(cors,
hc.order = FALSE,
# method = "square",
type = "upper") +
theme(
axis.text.x = element_text(size = 10, angle = 45, hjust = 1),
axis.text.y = element_text(size = 10)
)
ggcorrplot(cors, hc.order = FALSE, type = "upper",
lab = TRUE, lab_size = 2.5) +
theme(axis.text.x = element_text(size = 10, angle = 45, hjust = 1))
```
# Predicting dominance
## Load and merge data
Load the item-level means for the sensorimotor norms.
```{r}
df_contextualized_meanings = read_csv("../../data/processed/contextualized_sensorimotor_norms_with_ls.csv")
nrow(df_contextualized_meanings)
```
Load the dominance norms.
```{r}
df_dominance = read_csv("../../data/processed/dominance_norms_with_order.csv")
## Determine the specific sense/meaning of the righthand context
df_dominance = df_dominance %>%
mutate(context = substr(version_with_order, 6, 9))
## Now group by that righthand context to get relative dominance of that meaning
df_dominance_individual = df_dominance %>%
group_by(word, context) %>%
summarise(dominance = mean(dominance_right))
nrow(df_dominance_individual)
```
Merge the dominance and sensorimotor norms data.
```{r}
df_dom_plus_sm = df_contextualized_meanings %>%
inner_join(df_dominance_individual)
nrow(df_dom_plus_sm)
```
We also load and merge the Lancaster norms, as a control.
```{r}
df_lancaster <- read_csv("../../data/lexical/lancaster_norms.csv") %>%
mutate(word = tolower(Word)) %>%
rename(
Foot_leg.SD.LS = Foot_leg.SD,
Hand_arm.SD.LS = Hand_arm.SD,
Head.SD.LS = Head.SD,
Torso.SD.LS = Torso.SD
)
df_dom_plus_sm <- df_dom_plus_sm %>%
inner_join(df_lancaster)
nrow(df_dom_plus_sm)
```
## Calculating contextualized sensorimotor strength
Based on Lynott et al (2019), we **contextualized sensorimotor strength** as the *maximum* strength across all the dimensions.
```{r sm_strength_overall}
df_dom_plus_sm = df_dom_plus_sm %>%
rowwise() %>%
mutate(max_strength = max(
c(
## Modalities
Vision.M,
Hearing.M,
Olfaction.M,
Touch.M,
Taste.M,
Interoception.M,
## Effectors
Head.M,
Mouth_throat.M,
Torso.M,
Hand_arm.M,
Foot_leg.M
)
),
minkowski3_strength = sum(c(Vision.M, Hearing.M, Olfaction.M, Touch.M, Taste.M,
Interoception.M, Head.M, Mouth_throat.M, Torso.M,
Hand_arm.M, Foot_leg.M)^3)^(1/3),
max_perceptual_strength = max(
c(
## Modalities
Vision.M,
Hearing.M,
Olfaction.M,
Touch.M,
Taste.M,
Interoception.M
)
),
max_action_strength = max(
c(
## Effectors
Head.M,
Mouth_throat.M,
Torso.M,
Hand_arm.M,
Foot_leg.M
)
)
) %>%
ungroup()
df_dom_plus_sm %>%
ggplot(aes(x = Max_strength.sensorimotor,
y = max_strength)) +
geom_point(alpha = .5) +
labs(y = "Maximum Contextualized Strength",
x = "Maximum Strength (Lancaster)") +
theme_bw()
cor.test(df_dom_plus_sm$Max_strength.sensorimotor,
df_dom_plus_sm$max_strength)
cor.test(df_dom_plus_sm$max_strength,
df_dom_plus_sm$minkowski3_strength)
```
## Does contextualized sensorimotor strength predict dominance?
The answer is **yes**: contexts with a higher *maximum* sensorimotor strength also tend to be rated as more *dominant*.
Notably, this is true above and beyond the *decontextualized* ratings of sensorimotor strength for a given word.
```{r strength_dominance}
df_dom_plus_sm %>%
ggplot(aes(x = max_strength,
y = dominance)) +
geom_point(alpha = .4) +
geom_smooth(method = "lm") +
labs(x = "Maximum Contextualized Strength",
y = "Dominance") +
theme_minimal() +
theme(text = element_text(size=16))
df_dom_plus_sm %>%
ggplot(aes(x = minkowski3_strength,
y = dominance)) +
geom_point(alpha = .4) +
geom_smooth(method = "lm") +
labs(x = "Minkowski-3 Sensorimotor Strength",
y = "Dominance") +
theme_minimal() +
theme(text = element_text(size=16))
model_full = lmer(data = df_dom_plus_sm,
dominance ~
max_strength +
Max_strength.sensorimotor + Minkowski3.sensorimotor +
(1 | word),
REML = FALSE)
model_reduced = lmer(data = df_dom_plus_sm,
dominance ~
# max_strength +
Max_strength.sensorimotor + Minkowski3.sensorimotor +
(1 | word),
REML = FALSE)
summary(model_full)
anova(model_full, model_reduced)
model_full = lmer(data = df_dom_plus_sm,
dominance ~
minkowski3_strength +
Max_strength.sensorimotor + Minkowski3.sensorimotor +
(1 | word),
REML = FALSE)
model_reduced = lmer(data = df_dom_plus_sm,
dominance ~
# max_strength +
Max_strength.sensorimotor + Minkowski3.sensorimotor +
(1 | word),
REML = FALSE)
summary(model_full)
anova(model_full, model_reduced)
```
### Examples
```{r}
### Top 10
df_dom_plus_sm %>%
arrange(desc(minkowski3_strength)) %>%
select(word, sentence, minkowski3_strength, dominance) %>%
head(10)
### Bottom 10
df_dom_plus_sm %>%
arrange(minkowski3_strength) %>%
select(word, sentence, minkowski3_strength, dominance) %>%
head(10)
### Top 10
df_dom_plus_sm %>%
arrange(desc(max_strength)) %>%
select(word, sentence, max_strength, dominance) %>%
head(10)
### Bottom 10
df_dom_plus_sm %>%
arrange(max_strength) %>%
select(word, sentence, max_strength, dominance) %>%
head(10)
```
### Which dimension best predicts dominance?
Here, we ask whether specific dimensions are particularly correlated with sense dominance.
```{r specific_dimensions_dominance}
features <- c("Vision.M", "Hearing.M", "Olfaction.M", "Touch.M", "Taste.M",
"Interoception.M", "Head.M", "Mouth_throat.M", "Torso.M",
"Hand_arm.M", "Foot_leg.M")
# Models with each feature + baseline covariate
r2_results <- map_dfr(features, function(feat) {
formula <- as.formula(paste0("dominance ~ Max_strength.sensorimotor + ", feat))
model <- lm(formula, data = df_dom_plus_sm)
tibble(
feature = feat,
R2 = summary(model)$r.squared,
beta = coef(model)[3], # coefficient for the feature (3rd term now)
p_value = summary(model)$coefficients[3, 4]
)
})
# Baseline model (just Max_strength.sensorimotor)
baseline_model <- lm(dominance ~ Max_strength.sensorimotor, data = df_dom_plus_sm)
baseline_row <- tibble(
feature = "Baseline",
R2 = summary(baseline_model)$r.squared,
beta = NA,
p_value = NA
)
# Combine
r2_results_with_baseline <- bind_rows(baseline_row, r2_results)
# Plot
r2_results_with_baseline %>%
mutate(feature = str_remove(feature, "\\.M$"),
feature = str_replace(feature, "_", "/"),
sig = ifelse(p_value < .05, "*", ""),
is_baseline = feature == "Baseline",
feature = fct_reorder(feature, R2)) %>%
ggplot(aes(x = R2, y = feature, fill = is_baseline)) +
geom_col() +
geom_text(aes(label = sig), hjust = -0.5, size = 6, na.rm = TRUE) +
scale_fill_manual(values = c("grey40", "steelblue"), guide = "none") +
labs(x = "R²", y = NULL, title = "Predicting Sense Dominance") +
theme_minimal() +
theme(text = element_text(size = 16))
cor.test(df_dom_plus_sm$Touch.M,
df_dom_plus_sm$dominance)
cor.test(df_dom_plus_sm$Vision.M,
df_dom_plus_sm$dominance)
```
It looks like vision and touch are especially strong predictors. Which items drive this, e.g., for touch?
```{r touch_dominance}
by_sense <- df_dom_plus_sm %>%
mutate(sense = str_extract(context, "M[12]")) %>%
group_by(word, sense) %>%
summarize(
mean_touch = mean(Touch.M),
mean_vision = mean(Vision.M),
mean_dom = mean(dominance),
.groups = "drop"
)
# Now 2 rows per word—compute within-word difference
sense_diffs <- by_sense %>%
pivot_wider(names_from = sense,
values_from = c(mean_touch, mean_vision, mean_dom)) %>%
mutate(
touch_diff = mean_touch_M1 - mean_touch_M2,
vision_diff = mean_vision_M1 - mean_vision_M2,
dom_diff = mean_dom_M1 - mean_dom_M2
)
sense_diffs %>%
mutate(abs_touch_diff = abs(touch_diff)) %>%
arrange(desc(abs_touch_diff)) %>%
select(word, abs_touch_diff, dom_diff) %>%
head(5)
sense_diffs %>%
mutate(abs_vision_diff = abs(vision_diff)) %>%
arrange(desc(abs_vision_diff)) %>%
select(word, abs_vision_diff, dom_diff) %>%
head(5)
# Does the sense with higher Touch or Vision tend to be more dominant?
cor.test(sense_diffs$touch_diff, sense_diffs$dom_diff)
cor.test(sense_diffs$vision_diff, sense_diffs$dom_diff)
```
# How much does context add?
Here, we inspect how much information the CS norms provide, both relative to the LS norms and just overall (i.e., how much context pushes sensorimotor vectors around).
## Comparing dimensions to Lancaster
Now, we compare each dimension to the LS Norms.
```{r ls_deviation_market}
df_diffs_raw <- df_dom_plus_sm %>%
mutate(vision_diff = (Vision.M - Visual.mean),
auditory_diff = (Hearing.M - Auditory.mean),
intero_diff = (Interoception.M - Interoceptive.mean),
olfactory_diff = (Olfaction.M - Olfactory.mean),
touch_diff = (Touch.M - Haptic.mean),
taste_diff = (Taste.M - Gustatory.mean),
torso_diff = (Torso.M - Torso.mean),
hand_arm_diff = (Hand_arm.M - Hand_arm.mean),
foot_leg_diff = (Foot_leg.M - Foot_leg.mean),
head_diff = (Head.M - Head.mean),
mouth_throat_diff = (Mouth_throat.M - Mouth.mean)) %>%
pivot_longer(cols = ends_with("_diff"),
names_to = "Dimension", values_to = "Diff") %>%
mutate(Dimension = gsub('_diff', '', Dimension),
Dimension = case_when(
Dimension == "intero" ~ "Interoception",
Dimension == "auditory" ~ "Hearing",
Dimension == "olfactory" ~ "Olfaction",
TRUE ~ str_to_title(Dimension)
))
df_diffs_raw$Dimension = factor(df_diffs_raw$Dimension,
levels = rev(c(
'Vision',
'Hearing',
'Olfaction',
'Taste',
'Interoception',
'Touch',
'Mouth_throat',
'Head',
'Torso',
'Hand_arm',
'Foot_leg'
)))
df_diffs_raw %>%
filter(word == "market") %>%
ggplot(aes(x = Dimension,
y = Diff,
fill = Dimension)) +
geom_bar(stat = "summary") +
geom_vline(xintercept = 0, linetype = "dotted") +
theme_bw() +
coord_flip() +
labs(x = "Dimension",
y = "Deviation from Lancaster Norms") +
scale_fill_manual(values = viridisLite::viridis(11, option = "mako",
begin = 0.8, end = 0.15)) +
facet_wrap(~sentence) +
theme(text = element_text(size=16)) +
guides(fill = FALSE)
df_paired_market <- df_dom_plus_sm %>%
filter(word == "market") %>%
select(sentence,
Vision.M, Hearing.M, Olfaction.M, Taste.M, Interoception.M, Touch.M,
Mouth_throat.M, Head.M, Torso.M, Hand_arm.M, Foot_leg.M,
Visual.mean, Auditory.mean, Olfactory.mean, Gustatory.mean,
Interoceptive.mean, Haptic.mean, Mouth.mean, Head.mean,
Torso.mean, Hand_arm.mean, Foot_leg.mean) %>%
pivot_longer(cols = -sentence, names_to = "raw_dim", values_to = "Rating") %>%
mutate(
Source = case_when(
str_detect(raw_dim, "\\.M$") ~ "CS Norms (contextualized)",
TRUE ~ "LS Norms (decontextualized)"
),
Dimension = case_when(
raw_dim %in% c("Vision.M", "Visual.mean") ~ "Vision",
raw_dim %in% c("Hearing.M", "Auditory.mean") ~ "Hearing",
raw_dim %in% c("Olfaction.M", "Olfactory.mean") ~ "Olfaction",
raw_dim %in% c("Taste.M", "Gustatory.mean") ~ "Taste",
raw_dim %in% c("Interoception.M", "Interoceptive.mean") ~ "Interoception",
raw_dim %in% c("Touch.M", "Haptic.mean") ~ "Touch",
raw_dim %in% c("Mouth_throat.M", "Mouth.mean") ~ "Mouth_throat",
raw_dim %in% c("Head.M", "Head.mean") ~ "Head",
raw_dim %in% c("Torso.M", "Torso.mean") ~ "Torso",
raw_dim %in% c("Hand_arm.M", "Hand_arm.mean") ~ "Hand_arm",
raw_dim %in% c("Foot_leg.M", "Foot_leg.mean") ~ "Foot_leg"
)
)
df_paired_market$Dimension <- factor(df_paired_market$Dimension,
levels = rev(c('Vision','Hearing','Olfaction','Taste','Interoception',
'Touch','Mouth_throat','Head','Torso','Hand_arm','Foot_leg')))
df_paired_market %>%
ggplot(aes(x = Dimension, y = Rating, fill = Source)) +
geom_bar(stat = "summary", position = position_dodge(width = 0.8), width = 0.7) +
theme_bw() +
coord_flip() +
labs(x = "Dimension", y = "Sensorimotor Strength", fill = NULL) +
scale_fill_manual(values = c("CS Norms (contextualized)" = "#2c7a7b",
"LS Norms (decontextualized)" = "#7c5e9c")) +
facet_wrap(~sentence) +
theme(text = element_text(size = 14), legend.position = "bottom")
```
### Distribution of deviations
```{r largest_dev}
df_diffs_raw %>%
mutate(abs_diff = abs(Diff)) %>%
summarise(mean_abs_diff = mean(abs_diff),
median_abs_diff = median(abs_diff),
sd_abs_diff = sd(abs_diff),
min_abs_diff = min(abs_diff),
max_abs_diff = max(abs_diff))
df_diffs_raw %>%
mutate(abs_diff = abs(Diff)) %>%
ggplot(aes(x = abs_diff)) +
geom_histogram() +
labs(x = "Absolute Deviation") +
theme_minimal() +
facet_wrap(~Dimension) +
theme(text = element_text(size=16))
df_by_sentence_raw = df_diffs_raw %>%
mutate(abs_diff = abs(Diff)) %>%
group_by(word, sentence) %>%
summarise(mean_abs_diff = mean(abs_diff),
min_diff = min(abs_diff),
max_diff = max(abs_diff)) %>%
ungroup()
df_by_sentence_raw %>%
summarise(mean_diff = mean(mean_abs_diff),
median_abs_diff = median(mean_abs_diff),
sd_abs_diff = sd(mean_abs_diff),
min_abs_diff = min(mean_abs_diff),
max_abs_diff = max(mean_abs_diff))
df_by_sentence_raw %>%
ggplot(aes(x = mean_abs_diff)) +
geom_histogram() +
labs(x = "Mean Absolute Deviation (By Sentence)") +
theme_minimal() +
theme(text = element_text(size=16))
df_by_sentence_raw %>%
ggplot(aes(x = max_diff)) +
geom_histogram() +
labs(x = "Max Absolute Deviation (By Sentence)") +
theme_minimal() +
theme(text = element_text(size=16))
df_by_sentence_raw %>%
arrange(desc(max_diff)) %>%
head(10)
df_by_word_raw = df_diffs_raw %>%
mutate(abs_diff = abs(Diff)) %>%
group_by(word) %>%
summarise(mean_abs_diff = mean(abs_diff)) %>%
ungroup()
df_by_word_raw %>%
summarise(mean_diff = mean(mean_abs_diff),
median_abs_diff = median(mean_abs_diff),
sd_abs_diff = sd(mean_abs_diff),
min_abs_diff = min(mean_abs_diff),
max_abs_diff = max(mean_abs_diff))
df_diffs_raw %>%
mutate(abs_diff = abs(Diff)) %>%
group_by(word) %>%
summarise(mean_abs_diff = mean(abs_diff),
sd_abs_diff = sd(abs_diff),
range_diff = max(abs_diff) - min(abs_diff)) %>%
arrange(desc(mean_abs_diff)) %>%
head(5)
```
### Individual words
```{r individual_words}
df_diffs_raw = df_diffs_raw %>%
mutate(sense = str_extract(context, "M[12]"), # "M1" or "M2"
context_within_sense = str_extract(context, "[ab]")) # "a" or "b"
df_diffs_raw %>%
filter(word == "jam") %>%
mutate(sentence = fct_reorder(sentence, as.numeric(factor(sense)))) %>%
ggplot(aes(x = Dimension,
y = Diff,
fill = Dimension)) +
geom_bar(stat = "summary") +
geom_vline(xintercept = 0, linetype = "dotted") +
theme_bw() +
coord_flip() +
labs(x = "Dimension",
y = "Deviation from Lancaster Norms") +
scale_fill_manual(values = viridisLite::viridis(11, option = "mako",
begin = 0.8, end = 0.15)) +
facet_wrap(~sentence) +
theme(text = element_text(size=16)) +
guides(fill = FALSE)
df_diffs_raw %>%
mutate(sentence = fct_reorder(sentence, as.numeric(factor(sense)))) %>%
filter(word == "punch") %>%
ggplot(aes(x = Dimension,
y = Diff,
fill = Dimension)) +
geom_bar(stat = "summary") +
geom_vline(xintercept = 0, linetype = "dotted") +
theme_bw() +
coord_flip() +
labs(x = "Dimension",
y = "Deviation from Lancaster Norms") +
scale_fill_manual(values = viridisLite::viridis(11, option = "mako",
begin = 0.8, end = 0.15)) +
facet_wrap(~sentence) +
theme(text = element_text(size=16)) +
guides(fill = FALSE)
df_diffs_raw %>%
mutate(sentence = fct_reorder(sentence, as.numeric(factor(sense)))) %>%
filter(word == "match") %>%
ggplot(aes(x = Dimension,
y = Diff,
fill = Dimension)) +
geom_bar(stat = "summary") +
geom_vline(xintercept = 0, linetype = "dotted") +
theme_bw() +
coord_flip() +
labs(x = "Dimension",
y = "Deviation from Lancaster Norms") +
scale_fill_manual(values = viridisLite::viridis(11, option = "mako",
begin = 0.8, end = 0.15)) +
facet_wrap(~sentence) +
theme(text = element_text(size=16)) +
guides(fill = FALSE)
df_diffs_raw %>%
mutate(sentence = fct_reorder(sentence, as.numeric(factor(sense)))) %>%
filter(word == "pen") %>%
ggplot(aes(x = Dimension,
y = Diff,
fill = Dimension)) +
geom_bar(stat = "summary") +
geom_vline(xintercept = 0, linetype = "dotted") +
theme_bw() +
coord_flip() +
labs(x = "Dimension",
y = "Deviation from Lancaster Norms") +
scale_fill_manual(values = viridisLite::viridis(11, option = "mako",
begin = 0.8, end = 0.15)) +
facet_wrap(~sentence) +
theme(text = element_text(size=16)) +
guides(fill = FALSE)
df_diffs_raw %>%
mutate(sentence = fct_reorder(sentence, as.numeric(factor(sense)))) %>%
filter(word == "band") %>%
ggplot(aes(x = Dimension,
y = Diff,
fill = Dimension)) +
geom_bar(stat = "summary") +
geom_vline(xintercept = 0, linetype = "dotted") +
theme_bw() +
coord_flip() +
labs(x = "Dimension",
y = "Deviation from Lancaster Norms") +
scale_fill_manual(values = viridisLite::viridis(11, option = "mako",
begin = 0.8, end = 0.15)) +
facet_wrap(~sentence) +
theme(text = element_text(size=16)) +
guides(fill = FALSE)
```
### Which dimensions have the largest deviations?
```{r}
df_diffs_raw %>%
mutate(abs_diff = abs(Diff)) %>%
group_by(Dimension) %>%
summarise(mean_abs_diff = mean(abs_diff),
sd_abs_diff = sd(abs_diff),
range_diff = max(abs_diff) - min(abs_diff)) %>%
arrange(desc(range_diff))
```
## Z-scored deviations
```{r zscored_diffs}
SD_FLOOR <- 0.1
df_diffs_z_floored <- df_dom_plus_sm %>%
mutate(
vision_diff = (Vision.M - Visual.mean) / pmax(Visual.SD, SD_FLOOR),
auditory_diff = (Hearing.M - Auditory.mean) / pmax(Auditory.SD, SD_FLOOR),
intero_diff = (Interoception.M - Interoceptive.mean) / pmax(Interoceptive.SD, SD_FLOOR),
olfactory_diff = (Olfaction.M - Olfactory.mean) / pmax(Olfactory.SD, SD_FLOOR),
touch_diff = (Touch.M - Haptic.mean) / pmax(Haptic.SD, SD_FLOOR),
taste_diff = (Taste.M - Gustatory.mean) / pmax(Gustatory.SD, SD_FLOOR),
torso_diff = (Torso.M - Torso.mean) / pmax(Torso.SD.LS, SD_FLOOR),
hand_arm_diff = (Hand_arm.M - Hand_arm.mean) / pmax(Hand_arm.SD.LS, SD_FLOOR),
foot_leg_diff = (Foot_leg.M - Foot_leg.mean) / pmax(Foot_leg.SD.LS, SD_FLOOR),
head_diff = (Head.M - Head.mean) / pmax(Head.SD.LS, SD_FLOOR),
mouth_throat_diff = (Mouth_throat.M - Mouth.mean) / pmax(Mouth.SD, SD_FLOOR)
) %>%
pivot_longer(cols = ends_with("_diff"),
names_to = "Dimension", values_to = "Diff") %>%
mutate(Dimension = gsub('_diff', '', Dimension),
Dimension = case_when(
Dimension == "intero" ~ "Interoception",
Dimension == "auditory" ~ "Hearing",
Dimension == "olfactory" ~ "Olfaction",
TRUE ~ str_to_title(Dimension)
)) %>%
select(word, sentence, Dimension, Diff, context, dominance)
# Distribution
sentence_dev_floored <- df_diffs_z_floored %>%
mutate(abs_diff = abs(Diff)) %>%
group_by(word, sentence) %>%
summarise(mean_abs_diff = mean(abs_diff, na.rm = TRUE),
max_abs_diff = max(abs_diff),
.groups = "drop")
sentence_dev_floored %>%
ggplot(aes(x = mean_abs_diff)) +
geom_histogram(bins = 30) +
geom_vline(xintercept = 1, linetype = "dashed", color = "red") +
labs(x = "Mean Absolute Standardized Deviation (By Sentence)",
y = "Count of sentences") +
theme_minimal() +
theme(text = element_text(size=16))
sentence_dev_floored %>%
ggplot(aes(x = max_abs_diff)) +
geom_histogram(bins = 30) +
geom_vline(xintercept = 1, linetype = "dashed", color = "red") +
labs(x = "Max Absolute Standardized Deviation (By Sentence)",
y = "Count of sentences") +
theme_minimal() +
theme(text = element_text(size=16))
### Overall % sentences more than one on average
sentence_dev_floored %>%
mutate(greater_than_one = mean_abs_diff > 1) %>%
summarise(prop_more_than_one = mean(greater_than_one),
max_dev = max(mean_abs_diff))
### Overall % sentences more than one maximally
quantile(sentence_dev_floored$max_abs_diff,
probs = c(0.25, 0.5, 0.75, 0.9, 0.95, 0.99),
na.rm = TRUE)
### Overall deviations by dimension
df_diffs_z_floored %>%
mutate(abs_diff = abs(Diff)) %>%
ggplot(aes(x = abs_diff)) +
geom_histogram() +
labs(x = "Standardized Absolute Deviation") +
geom_vline(xintercept = 1, linetype = "dashed", color = "red") +
theme_minimal() +
facet_wrap(~Dimension) +
theme(text = element_text(size=16))
### Overall % of contexts exceeding 1 SD
df_diffs_z_floored %>%
mutate(abs_diff = abs(Diff)) %>%
mutate(greater_than_one = abs_diff > 1) %>%
summarise(prop_more_than_one = mean(greater_than_one),
max_dev = max(abs_diff))
#### Additional summary statistics
df_diffs_z_floored %>%
mutate(abs_diff = abs(Diff)) %>%
summarise(mean_abs_diff = mean(abs_diff),
median_abs_diff = median(abs_diff),
sd_abs_diff = sd(abs_diff),
min_abs_diff = min(abs_diff),
max_abs_diff = max(abs_diff))
df_by_sentence_z_floored = df_diffs_z_floored %>%
mutate(abs_diff = abs(Diff)) %>%
group_by(word, sentence) %>%
summarise(mean_abs_diff = mean(abs_diff)) %>%
ungroup()
df_by_sentence_z_floored %>%
summarise(mean_diff = mean(mean_abs_diff),
median_abs_diff = median(mean_abs_diff),
sd_abs_diff = sd(mean_abs_diff),
min_abs_diff = min(mean_abs_diff),
max_abs_diff = max(mean_abs_diff))
df_by_word_z_floored = df_diffs_z_floored %>%
mutate(abs_diff = abs(Diff)) %>%
group_by(word) %>%
summarise(mean_abs_diff = mean(abs_diff)) %>%
ungroup()
df_by_word_z_floored %>%
summarise(mean_diff = mean(mean_abs_diff),
median_abs_diff = median(mean_abs_diff),
sd_abs_diff = sd(mean_abs_diff),
min_abs_diff = min(mean_abs_diff),
max_abs_diff = max(mean_abs_diff))
```
### Compare to raw differences
```{r}
# Raw (unstandardized) mean absolute deviation per sentence
sentence_dev_raw <- df_diffs_raw %>%
mutate(abs_diff = abs(Diff)) %>%
group_by(word, sentence, context) %>%
summarise(mean_abs_diff = mean(abs_diff, na.rm = TRUE),
.groups = "drop") %>%
mutate(version = "Raw")
# Standardized (floored) mean absolute deviation per sentence
sentence_dev_z <- df_diffs_z_floored %>%
mutate(abs_diff = abs(Diff)) %>%
group_by(word, sentence, context) %>%
summarise(mean_abs_diff = mean(abs_diff, na.rm = TRUE),
.groups = "drop") %>%
mutate(version = "Standardized")
# By sentence
sentence_compare <- sentence_dev_raw %>%
select(word, sentence, context, raw = mean_abs_diff) %>%
inner_join(sentence_dev_z %>%
select(word, sentence, context, standardized = mean_abs_diff),
by = c("word", "sentence", "context"))
cor.test(sentence_compare$raw, sentence_compare$standardized, method = "spearman")
cor.test(sentence_compare$raw, sentence_compare$standardized, method = "pearson")
sentence_compare %>%
ggplot(aes(x = raw, y = standardized)) +
geom_point(alpha = 0.3) +
geom_smooth(method = "lm", se = FALSE) +
labs(x = "Raw mean absolute deviation",
y = "Standardized mean absolute deviation",
title = "By-sentence deviations") +
theme_minimal() +
theme(text = element_text(size=16))
```
## Within-word range
Here, we calculate the *range* of sensorimotor associatoins for each dimension across the contexts of use for each word.
### Across all contexts
```{r within_word_range}
df_rawc_with_norms = read_csv("../../data/processed/sentence_pairs_with_sensorimotor_distance.csv") %>%
drop_na(sensorimotor_distance) %>%
select(sensorimotor_distance, action_distance, perceptual_distance,
word, same, ambiguity_type, sentence1, sentence2, mean_relatedness, Class, version)
nrow(df_rawc_with_norms)
word_dim_range <- df_dom_plus_sm %>%
pivot_longer(cols = c(Vision.M, Hearing.M, Olfaction.M, Taste.M,
Interoception.M, Touch.M, Mouth_throat.M,
Head.M, Hand_arm.M, Foot_leg.M, Torso.M),
names_to = "Dimension", values_to = "CS_value") %>%
mutate(Dimension = gsub("\\.M$", "", Dimension)) %>%
group_by(word, Dimension) %>%
summarise(within_range = max(CS_value, na.rm = TRUE) - min(CS_value, na.rm = TRUE),
.groups = "drop")
# Aggregate to word level (mean range across dimensions)
word_range <- word_dim_range %>%
group_by(word) %>%
summarise(mean_range = mean(within_range, na.rm = TRUE),
max_range = max(within_range, na.rm = TRUE),
.groups = "drop") %>%
arrange(desc(mean_range))
print(head(word_range, 15))
print(tail(word_range, 15))
# By dimension
word_dim_range %>%
ggplot(aes(x = within_range)) +
geom_histogram(bins = 30) +
labs(x = "Within-Word Range of CS Ratings (across contexts)",
y = "Count of words") +
theme_minimal() +
facet_wrap(~Dimension)
# Distribution
word_range %>%
ggplot(aes(x = mean_range)) +
geom_histogram(bins = 30) +
labs(x = "Mean Within-Word Range of CS Ratings (across contexts)",
y = "Count of words") +
theme_minimal()
# Summary stats
word_range %>%
summarise(mean = mean(mean_range),
sd = sd(mean_range),
median = median(mean_range),
min = min(mean_range),
max = max(mean_range))
# Get per-word ambiguity type
word_ambiguity <- df_rawc_with_norms %>%
distinct(word, ambiguity_type)
# Merge with within-word range
word_range_with_amb <- word_range %>%
inner_join(word_ambiguity, by = "word")
# Summary stats by ambiguity type
word_range_with_amb %>%
group_by(ambiguity_type) %>%
summarise(mean = mean(mean_range),
sd = sd(mean_range),
median = median(mean_range),
n = n(),
.groups = "drop")
# Test the difference
summary(lm(data = word_range_with_amb,
mean_range ~ ambiguity_type))
t.test(mean_range ~ ambiguity_type, data = word_range_with_amb)
# Visualize
word_range_with_amb %>%
ggplot(aes(x = ambiguity_type, y = mean_range, fill = ambiguity_type)) +
geom_boxplot(alpha = 0.6) +
geom_jitter(width = 0.15, alpha = 0.3) +
labs(x = "Ambiguity Type",
y = "Mean Within-Word Range of CS Ratings") +
theme_minimal() +
theme(legend.position = "none")
```
### Between vs. Within sense
```{r within_word_range_senses}
# Need sense info from context column (M1/M2)
word_range_decomposed <- df_dom_plus_sm %>%
mutate(sense = str_extract(context, "M[12]")) %>%
pivot_longer(cols = c(Vision.M, Hearing.M, Olfaction.M, Taste.M,
Interoception.M, Touch.M, Mouth_throat.M,
Head.M, Hand_arm.M, Foot_leg.M, Torso.M),
names_to = "Dimension", values_to = "CS_value") %>%
mutate(Dimension = gsub("\\.M$", "", Dimension)) %>%
group_by(word, Dimension, sense) %>%
summarise(sense_mean = mean(CS_value),
sense_within_range = max(CS_value) - min(CS_value),
.groups = "drop") %>%
group_by(word, Dimension) %>%
summarise(between_sense_range = max(sense_mean) - min(sense_mean),
within_sense_range = mean(sense_within_range),
.groups = "drop") %>%
group_by(word) %>%
summarise(mean_between_sense = mean(between_sense_range),
mean_within_sense = mean(within_sense_range),
max_between_sense = max(between_sense_range),
max_within_sense = max(within_sense_range),
.groups = "drop")
word_range_decomposed %>%
inner_join(word_ambiguity, by = "word") %>%
group_by(ambiguity_type) %>%
summarise(between = mean(mean_between_sense),
within = mean(mean_within_sense),,
between_max = mean(max_between_sense),
within_max = mean(max_within_sense),
n = n())
# Test each separately
t.test(mean_between_sense ~ ambiguity_type,
data = word_range_decomposed %>% inner_join(word_ambiguity, by = "word"))
t.test(mean_within_sense ~ ambiguity_type,
data = word_range_decomposed %>% inner_join(word_ambiguity, by = "word"))
# Test max
t.test(max_between_sense ~ ambiguity_type,
data = word_range_decomposed %>% inner_join(word_ambiguity, by = "word"))
t.test(max_within_sense ~ ambiguity_type,
data = word_range_decomposed %>% inner_join(word_ambiguity, by = "word"))
word_range_decomposed_long = word_range_decomposed %>%
inner_join(word_ambiguity, by = "word") %>%
pivot_longer(cols = c(mean_between_sense, mean_within_sense),
names_to = "comparison",
values_to = "word_range")
summary(lmer(data = word_range_decomposed_long,
word_range ~ comparison + (1|word)))
summary(lmer(data = word_range_decomposed_long,
word_range ~ ambiguity_type*comparison + (1|word)))
word_range_decomposed_long %>%
mutate(comparison = case_when(
comparison == "mean_between_sense" ~ "Different Sense",
comparison == "mean_within_sense" ~ "Same Sense"
)) %>%
ggplot(aes(x = word_range,
y = ambiguity_type,
fill = comparison)) +
geom_density_ridges2(aes(height = ..density..),
color = NA,
alpha = 0.5,
scale=0.85,
stat="density") +
labs(x = "Mean Within-Word Range of CS Ratings",
y = NULL,
fill = NULL) +
theme_minimal() +
scale_fill_manual(
values = c("Same Sense" = viridisLite::viridis(2, option = "mako", begin = 0.8, end = 0.15)[1],
"Different Sense" = viridisLite::viridis(2, option = "mako", begin = 0.8, end = 0.15)[2])
) +
theme(text = element_text(size = 20))
word_range_decomposed_long = word_range_decomposed %>%
inner_join(word_ambiguity, by = "word") %>%
pivot_longer(cols = c(max_between_sense, max_within_sense),
names_to = "comparison",
values_to = "word_range")
summary(lmer(data = word_range_decomposed_long,
word_range ~ comparison + (1|word)))
summary(lmer(data = word_range_decomposed_long,
word_range ~ ambiguity_type*comparison + (1|word)))
word_range_decomposed_long %>%
mutate(comparison = case_when(
comparison == "max_between_sense" ~ "Different Sense",
comparison == "max_within_sense" ~ "Same Sense"
)) %>%
ggplot(aes(x = word_range,
y = ambiguity_type,
fill = comparison)) +
geom_density_ridges2(aes(height = ..density..),
color = NA,
alpha = 0.5,
scale=0.85,
stat="density") +
labs(x = "Max Within-Word Range of CS Ratings",
y = NULL,
fill = NULL) +
theme_minimal() +
scale_fill_manual(
values = c("Same Sense" = viridisLite::viridis(2, option = "mako", begin = 0.8, end = 0.15)[1],
"Different Sense" = viridisLite::viridis(2, option = "mako", begin = 0.8, end = 0.15)[2])
) +
theme(text = element_text(size = 20))
```
# How does dominance relate to deviation from the LS Norms?
This question can in turn be decomposed into two questions:
First, are more dominant senses **closer** to the LS Norms overall? We might expect this to be the case if the LS Norms reflect the dominant sense; that is, when people rate the sensorimotor properties of a decontextualized word, they might be more likely to index properties associated with the most dominant contexts or meanings of that word.
And the answer is **yes**: more dominant senses are indeed more *similar* (less distant) from the Lancaster norm in terms of their sensorimotor profile.
```{r dominance_ls_deviation}
model_with_dominance = lmer(data = df_dom_plus_sm,
distance_to_lancaster ~ dominance + (1 | word),
REML = FALSE)
model_no_dominance = lmer(data = df_dom_plus_sm,
distance_to_lancaster ~ (1 | word),
REML = FALSE)
summary(model_with_dominance)
anova(model_with_dominance, model_no_dominance)
df_dom_plus_sm %>%
ggplot(aes(x = dominance,
y = distance_to_lancaster)) +
geom_point(alpha = .5) +
geom_smooth(method = "lm") +
labs(x = "Dominance",
y = "Cosine Distance to Decontextualized LS Norm") +
theme_bw()
cor.test(df_dom_plus_sm$dominance, df_dom_plus_sm$distance_to_lancaster)
```
And second: does dominance predict the **direction** of difference?
The earlier analysis of dominance suggests that more dominant senses are more concrete than less dominant senses. Thus, we might expect that more dominant senses are also more concrete on average than the decontextualized norms.
We find that this is true: that is, more dominant senses are *more concrete* on average than the LS norm.
```{r}
df_diffs_avg = df_diffs_raw %>%
group_by(word, sentence, context) %>%
summarise(mean_diff = mean(Diff))
df_diffs_avg = df_diffs_avg %>%
left_join(df_dom_plus_sm)
model_with_dominance = lmer(data = df_diffs_avg,
mean_diff ~ dominance + (1 | word),
REML = FALSE)
model_no_dominance = lmer(data = df_diffs_avg,
mean_diff ~ (1 | word),
REML = FALSE)
summary(model_with_dominance)
anova(model_with_dominance, model_no_dominance)
cor.test(df_diffs_avg$dominance, df_diffs_avg$mean_diff)
```
# Predicting relatedness
Next, we ask about the **sensorimotor distance** between two sentence pairs, and whether it correlates both with `same/different sense` and the `mean_relatedness` judgments for those sentence pairs.
Here, we load a version of the dataset that also contains a *baseline* measure: the sensorimotor distance as calculated using a bag-of-words approach (i.e., using the original Lancaster Norms).
## Load data
```{r}
df_rawc_with_norms = read_csv("../../data/processed/sentence_pairs_with_sensorimotor_distance.csv") %>%
drop_na(sensorimotor_distance) %>%
select(sensorimotor_distance, action_distance, perceptual_distance,
word, same, ambiguity_type, sentence1, sentence2, mean_relatedness, Class, version)
nrow(df_rawc_with_norms)
```
## Load English RAW-C data
```{r}
df_bert = read_csv("../../data/processed/models_english/rawc-distances_model-bert-base-uncased.csv") %>%
mutate(Model = "BERT-base-uncased",
Multilingual = "Monolingual")
df_bert_cased = read_csv("../../data/processed/models_english/rawc-distances_model-bert-base-cased.csv") %>%
mutate(Model = "BERT-base-cased",
Multilingual = "Monolingual")
df_xlm = read_csv("../../data/processed/models_english/rawc-distances_model-xlm-roberta-base.csv") %>%
mutate(Model = "XLM-RoBERTa",
Multilingual = "Multilingual")
df_ab1 = read_csv("../../data/processed/models_english/rawc-distances_model-albert-base-v1.csv") %>%
mutate(Model = "ALBERT-base-v1",
Multilingual = "Monolingual")
df_ab2 = read_csv("../../data/processed/models_english/rawc-distances_model-albert-base-v2.csv") %>%
mutate(Model = "ALBERT-base-v2",
Multilingual = "Monolingual")
df_al = read_csv("../../data/processed/models_english/rawc-distances_model-albert-large-v2.csv") %>%
mutate(Model = "ALBERT-large-v2",
Multilingual = "Monolingual")
df_axl = read_csv("../../data/processed/models_english/rawc-distances_model-albert-xlarge-v2.csv") %>%
mutate(Model = "ALBERT-xlarge-v2",
Multilingual = "Monolingual")
df_axxl = read_csv("../../data/processed/models_english/rawc-distances_model-albert-xxlarge-v2.csv") %>%
mutate(Model = "ALBERT-xxlarge-v2",
Multilingual = "Monolingual")
df_rb = read_csv("../../data/processed/models_english/rawc-distances_model-roberta-base.csv") %>%
mutate(Model = "RoBERTa-base",
Multilingual = "Monolingual")
df_rl = read_csv("../../data/processed/models_english/rawc-distances_model-roberta-large.csv") %>%
mutate(Model = "RoBERTa-large",
Multilingual = "Monolingual")
df_db = read_csv("../../data/processed/models_english/rawc-distances_model-distilbert-base-uncased.csv") %>%
mutate(Model = "DistilBERT",
Multilingual = "Monolingual")
df_mb = read_csv("../../data/processed/models_english/rawc-distances_model-bert-base-multilingual-cased.csv") %>%
mutate(Model = "Multilingual BERT",
Multilingual = "Multilingual")
df_all = df_bert %>%
bind_rows(df_bert_cased) %>%
bind_rows(df_xlm) %>%
bind_rows(df_ab1) %>%
bind_rows(df_ab2) %>%
bind_rows(df_al) %>%
bind_rows(df_axl) %>%
bind_rows(df_axxl) %>%
bind_rows(df_rb) %>%
bind_rows(df_rl) %>%
bind_rows(df_db) %>%
bind_rows(df_mb)
df_merged = df_rawc_with_norms %>%
inner_join(df_all)
```
## Predicting same/different sense
### Sensorimotor distance and sense boundaries
```{r sm_sense}
df_rawc_with_norms = df_rawc_with_norms %>%
mutate(Same = case_when(
same == TRUE ~ "Same Sense",
same == FALSE ~ "Different Sense"
))
df_rawc_with_norms %>%
ggplot(aes(x = sensorimotor_distance,
y = ambiguity_type,
fill = Same)) +
geom_density_ridges2(aes(height = ..density..),
color = NA,
alpha = 0.5,
scale=0.85,
stat="density") +
labs(x = "Sensorimotor Distance",
y = NULL,
fill = NULL) +
scale_fill_manual(
values = c("Same Sense" = viridisLite::viridis(2, option = "mako", begin = 0.8, end = 0.15)[1],
"Different Sense" = viridisLite::viridis(2, option = "mako", begin = 0.8, end = 0.15)[2])
) +
theme_minimal() +
theme(text = element_text(size = 20))
model_full = lmer(data = df_rawc_with_norms,
sensorimotor_distance ~ same +
(1 + same | word),
control=lmerControl(optimizer="bobyqa"),
REML = FALSE)
model_reduced = lmer(data = df_rawc_with_norms,
sensorimotor_distance ~ # same +
(1 + same | word),
control=lmerControl(optimizer="bobyqa"),
REML = FALSE)
summary(model_full)
anova(model_full, model_reduced)
model_full_at = lmer(data = df_rawc_with_norms,
sensorimotor_distance ~ same * ambiguity_type +
(1 + same | word),
control=lmerControl(optimizer="bobyqa"),
REML = FALSE)
model_reduced_at = lmer(data = df_rawc_with_norms,
sensorimotor_distance ~ same + ambiguity_type +
(1 + same | word),
control=lmerControl(optimizer="bobyqa"),
REML = FALSE)
summary(model_full_at)
anova(model_full_at, model_reduced_at)
df_rawc_with_norms %>%
group_by(same, ambiguity_type) %>%
summarise(mean = mean(sensorimotor_distance),
sd = sd(sensorimotor_distance))
```
### Predicting sense boundaries
Here, we run the actual analysis:
```{r aic_sense_boundaries}
model_sm_only <- glm(same ~ sensorimotor_distance,
data = df_rawc_with_norms,
family = binomial)
aic_sm_baseline <- AIC(model_sm_only)
print(paste("AIC:", round(aic_sm_baseline, 1)))
### Get aic
aic_results <- df_merged %>%
group_by(n_params, Model, Layer) %>%
summarize(
# BERT only
aic_bert = {
model <- glm(Same_sense ~ Distance, family = binomial)
AIC(model)
},
# Combined
aic_combined = {
model <- glm(Same_sense ~ Distance + sensorimotor_distance, family = binomial)
AIC(model)
},
aic_sm = {
model <- glm(Same_sense ~ sensorimotor_distance, family = binomial)
AIC(model)
},
.groups = "drop"
) %>%
mutate(
# AIC Delta
aic_bert_vs_sm = aic_bert - aic_sm_baseline,
aic_combined_vs_sm = aic_sm_baseline - aic_combined,
aic_combined_vs_dist = aic_bert - aic_combined
)
# ============================================================
# 3. Best layer per model
# ============================================================
best_layers <- aic_results %>%
group_by(Model, n_params) %>%
slice_min(aic_bert, n = 1) %>%
select(Model, Layer,
aic_bert_vs_sm, aic_combined_vs_sm, aic_combined_vs_dist,
aic_sm, aic_bert, aic_combined)
best_layers = best_layers %>%
mutate(aic_bert_vs_sm2 = aic_bert - aic_sm_baseline)
print(best_layers)
## AIC difference from SM baseline
best_layers %>%
select(Model, aic_bert_vs_sm, aic_combined_vs_sm) %>%
pivot_longer(cols = c(aic_bert_vs_sm, aic_combined_vs_sm),
names_to = "type", values_to = "aic_diff") %>%
mutate(type = ifelse(type == "aic_bert_vs_sm",
"Distributional only",
"Distributional + Sensorimotor"),
type = factor(type, levels = c("Distributional only",
"Distributional + Sensorimotor")),
Model = fct_reorder(Model, aic_diff)) %>%
ggplot(aes(x = aic_diff, y = Model, fill = type)) +
geom_col(position = "dodge") +
geom_vline(xintercept = 0, linetype = "dashed", color = "red") +
labs(x = "ΔAIC vs. Sensorimotor baseline",
y = NULL,
fill = NULL) +
scale_fill_manual(values = c("Distributional only" = "gray60",
"Distributional + Sensorimotor" = "steelblue")) +
theme_minimal() +
theme(text = element_text(size = 16),
legend.position = "bottom")
best_layers %>%
ungroup() %>%
select(Model, aic_bert, aic_combined) %>%
mutate(across(starts_with("aic"), round)) %>%
kbl(col.names = c("Model", "Distributional", "Hybrid"),
format = "latex",
booktabs = TRUE,
caption = "AIC values for predicting sense boundaries using Distributional Distance alone or both Distributional Distance and Sensorimotor Distance (AIC for Sensorimotor Distance was 707)",
label = "aic") %>%
kable_styling(latex_options = "hold_position")
aic_results %>%
mutate(dist_better_than_sm = aic_bert_vs_sm > 4,
hybrid_better_than_sm = aic_combined_vs_sm > 4,
hybrid_better_than_dist = aic_combined_vs_dist > 4) %>%
summarise(mean(dist_better_than_sm),
mean(hybrid_better_than_sm),
mean(hybrid_better_than_dist))
best_layers %>%
ungroup() %>%
mutate(dist_better_than_sm = aic_bert_vs_sm < -4,
hybrid_better_than_sm = aic_combined_vs_sm > 4,
hybrid_better_than_dist = aic_combined_vs_dist > 4) %>%
summarise(mean(dist_better_than_sm),
mean(hybrid_better_than_sm),
mean(hybrid_better_than_dist))
### Visualize raw AIC
aic_summary_full <- best_layers %>%
select(Model, Layer, aic_sm, aic_bert, aic_combined) %>%
pivot_longer(cols = c(aic_sm, aic_bert, aic_combined),
names_to = "type", values_to = "aic") %>%
mutate(type = case_when(
type == "aic_sm" ~ "Sensorimotor",
type == "aic_bert" ~ "Distributional",
type == "aic_combined" ~ "Hybrid"
),
type = factor(type, levels = c("Sensorimotor", "Distributional", "Hybrid")))
# Get sensorimotor baseline (should be the same for all models)
aic_sm_baseline <- aic_summary_full %>%
filter(type == "Sensorimotor") %>%
pull(aic) %>%
unique()
# Filter to just distributional and hybrid
aic_summary_no_sm <- aic_summary_full %>%
filter(type != "Sensorimotor")
aic_summary_se_no_sm <- aic_summary_no_sm %>%
group_by(type) %>%
summarize(
mean_aic = mean(aic),
se_aic = sd(aic) / sqrt(n())
)
ggplot() +
# Sensorimotor baseline as dashed line
geom_hline(yintercept = aic_sm_baseline,
linetype = "dashed", color = "coral", linewidth = 1) +
# Raw points for each model/layer
geom_jitter(data = aic_summary_no_sm,
aes(x = type, y = aic, color = type),
alpha = 0.3, width = 0.1, size = 2) +
# Mean + SE
geom_point(data = aic_summary_se_no_sm,
aes(x = type, y = mean_aic, color = type),
size = 5) +
geom_errorbar(data = aic_summary_se_no_sm,
aes(x = type, ymin = mean_aic - se_aic,
ymax = mean_aic + se_aic, color = type),
width = 0.15, linewidth = 1.2) +
annotate("text", x = 1.5, y = aic_sm_baseline + 30,
label = "Sensorimotor", color = "coral", size = 5)+
labs(x = NULL,
y = "AIC (lower is better)",
title = "Predicting Sense Boundary") +
scale_color_manual(values = c("Distributional" = "gray60",
"Hybrid" = "steelblue")) +
theme_minimal() +
theme(text = element_text(size = 16),
legend.position = "none")
```
## Predicting relatedness
### Comparing to each distributional relatedness measure
```{r relatedness_comparison}
df_rawc_with_norms %>%
ggplot(aes(x = sensorimotor_distance,
y = mean_relatedness,
color = Same)) +
geom_point(alpha = .5) +
scale_color_manual(
values = c("Same Sense" = viridisLite::viridis(2, option = "mako",
begin = 0.8, end = 0.15)[1],
"Different Sense" = viridisLite::viridis(2, option = "mako",
begin = 0.8, end = 0.15)[2])
) +
theme_minimal() +
labs(x = "Sensorimotor Distance",
y = "Mean Relatedness",
color = "") +
theme(text = element_text(size = 15),
legend.position="bottom")
cor(df_rawc_with_norms$sensorimotor_distance,
df_rawc_with_norms$mean_relatedness)
```
### Comparing AIC
```{r aic_relatedness}
# ============================================================
# 2. For each model/layer: compute R² and AIC, difference from SM baseline
# ============================================================
relatedness_results <- df_merged %>%
group_by(Model, Layer, n_params, Multilingual) %>%
summarize(
# Sensorimotor only
r2_sm = {
model <- lm(mean_relatedness ~ sensorimotor_distance)
summary(model)$r.squared
},
aic_sm = {
model <- lm(mean_relatedness ~ sensorimotor_distance)
AIC(model)
},
# BERT only
r2_bert = {
model <- lm(mean_relatedness ~ Distance)
summary(model)$r.squared
},
aic_bert = {
model <- lm(mean_relatedness ~ Distance)
AIC(model)
},
# Combined
r2_combined = {
model <- lm(mean_relatedness ~ Distance + sensorimotor_distance)
summary(model)$r.squared
},
aic_combined = {
model <- lm(mean_relatedness ~ Distance + sensorimotor_distance)
AIC(model)
},
.groups = "drop"
) %>%
mutate(
aic_bert_vs_sm = aic_bert - aic_sm,
aic_combined_vs_sm = aic_sm - aic_combined,
aic_combined_vs_dist = aic_bert - aic_combined
)
# ============================================================
# 3. Best layer per model
# ============================================================
best_layers <- relatedness_results %>%
group_by(Model, n_params, Multilingual) %>%
slice_max(r2_bert, n = 1) %>%
select(Model, Layer, n_params, Multilingual,
# r2_sm, r2_bert, r2_combined,
aic_bert_vs_sm, aic_sm, aic_bert, aic_combined,
aic_combined_vs_sm, aic_combined_vs_dist)
print(best_layers)
best_layers %>%
ungroup() %>%
select(Model, aic_bert, aic_combined) %>%
mutate(across(starts_with("aic"), round)) %>%
kbl(col.names = c("Model", "Distributional", "Hybrid"),
format = "latex",
booktabs = TRUE,
caption = "AIC values for predicting relatedness judgments using Distributional Distance alone or both Distributional Distance and Sensorimotor Distance (AIC for Sensorimotor Distance was 2134.78.)",
label = "aic") %>%
kable_styling(latex_options = "hold_position")
best_layers %>%
select(Model, aic_bert_vs_sm, aic_combined_vs_sm) %>%
pivot_longer(cols = c(aic_bert_vs_sm, aic_combined_vs_sm),
names_to = "type", values_to = "aic_diff") %>%
mutate(type = ifelse(type == "aic_bert_vs_sm", "Distributional only", "Distributional + Sensorimotor"),
type = factor(type, levels = c("Distributional only", "Distributional + Sensorimotor")),
Model = fct_reorder(Model, aic_diff)) %>%
ggplot(aes(x = aic_diff, y = Model, fill = type)) +
geom_col(position = "dodge") +
geom_vline(xintercept = 0, linetype = "dashed", color = "red") +
geom_vline(xintercept = 4, linetype = "dotted", color = "gray50") +
geom_vline(xintercept = -4, linetype = "dotted", color = "gray50") +
labs(x = "ΔAIC vs. Sensorimotor baseline",
y = NULL,
fill = NULL) +
scale_fill_manual(values = c("Distributional only" = "gray60",
"Distributional + Sensorimotor" = "steelblue")) +
theme_minimal() +
theme(text = element_text(size = 16),
legend.position = "bottom")
# ============================================================
# 7. Summary
# ============================================================
best_layers %>%
ungroup() %>%
mutate(dist_better_than_sm = aic_bert_vs_sm < -4,
hybrid_better_than_sm = aic_combined_vs_sm > 4,
hybrid_better_than_dist = aic_combined_vs_dist > 4) %>%
summarise(mean(dist_better_than_sm),
mean(hybrid_better_than_sm),
mean(hybrid_better_than_dist))
### Visualize raw AIC
aic_summary_full <- best_layers %>%
select(Model, Layer, aic_sm, aic_bert, aic_combined) %>%
pivot_longer(cols = c(aic_sm, aic_bert, aic_combined),
names_to = "type", values_to = "aic") %>%
mutate(type = case_when(
type == "aic_sm" ~ "Sensorimotor",
type == "aic_bert" ~ "Distributional",
type == "aic_combined" ~ "Hybrid"
),
type = factor(type, levels = c("Sensorimotor", "Distributional", "Hybrid")))
# Get sensorimotor baseline (should be the same for all models)
aic_sm_baseline <- aic_summary_full %>%
filter(type == "Sensorimotor") %>%
pull(aic) %>%
unique()
# Filter to just distributional and hybrid
aic_summary_no_sm <- aic_summary_full %>%
filter(type != "Sensorimotor")
aic_summary_se_no_sm <- aic_summary_no_sm %>%
group_by(type) %>%
summarize(
mean_aic = mean(aic),
se_aic = sd(aic) / sqrt(n())
)
ggplot() +
# Sensorimotor baseline as dashed line
geom_hline(yintercept = aic_sm_baseline,
linetype = "dashed", color = "coral", linewidth = 1) +
# Raw points for each model/layer
geom_jitter(data = aic_summary_no_sm,
aes(x = type, y = aic, color = type),
alpha = 0.3, width = 0.1, size = 2) +
# Mean + SE
geom_point(data = aic_summary_se_no_sm,
aes(x = type, y = mean_aic, color = type),
size = 5) +
geom_errorbar(data = aic_summary_se_no_sm,
aes(x = type, ymin = mean_aic - se_aic,
ymax = mean_aic + se_aic, color = type),
width = 0.15, linewidth = 1.2) +
annotate("text", x = 1.5, y = aic_sm_baseline + 30,
label = "Sensorimotor", color = "coral", size = 5)+
labs(x = NULL,
y = "AIC (lower is better)",
title = "Predicting Relatedness") +
scale_color_manual(values = c("Distributional" = "gray60",
"Hybrid" = "steelblue")) +
theme_minimal() +
theme(text = element_text(size = 16),
legend.position = "none")
```
## Correlation between sensorimotor and distributional distance
```{r corr_dist_sm}
# Get best layer
df_by_layer = df_merged %>%
group_by(Model, Multilingual, Layer, n_params) %>%
summarise(r = cor(mean_relatedness, Distance, method = "pearson"),
r2 = r ** 2,
rho = cor(mean_relatedness, Distance, method = "spearman"),
count = n())
df_best_layer <- df_by_layer %>%
group_by(Model) %>%
slice_max(r2, n = 1) %>%
select(Model, Layer, r2)
# Filter for only best layers
df_best <- df_merged %>%
semi_join(df_best_layer, by = c("Model", "Layer"))
# Pivot wider
df_wide <- df_best %>%
mutate(`Sensorimotor Distance` = sensorimotor_distance) %>%
select(word, sentence1, sentence2, `Sensorimotor Distance`, Model, Distance) %>%
pivot_wider(names_from = Model, values_from = Distance)
# Compute correlation matrix
cor_matrix <- df_wide %>%
select(`Sensorimotor Distance`, where(is.numeric)) %>%
cor(use = "pairwise.complete.obs")
# Plot the correlation matrix
ggcorrplot(cor_matrix,
hc.order = FALSE,
method = "square") +
theme(
axis.text.x = element_text(size = 10, angle = 45, hjust = 1),
axis.text.y = element_text(size = 10)
)
```
# Baseline with LS Norms
```{r}
df_with_baseline = read_csv("../../data/processed/sentence_pairs_with_baseline.csv") %>%
drop_na(sensorimotor_distance) %>%
drop_na(baseline_distance)
nrow(df_with_baseline)
```
## Does sensorimotor distance predict relatedness above the baseline?
We also ask whether whether our measure of contextualized sensorimotor distance predicts relatedness above and beyond a baseline that simply considers the decontextualized Lancaster Sensorimotor Norms for the dismabiguating words in a sentence. (We find that it does.)
```{r}
model_bow_sm = lmer(data = df_with_baseline,
mean_relatedness ~ baseline_distance + sensorimotor_distance +
(1| word),
REML = FALSE)
model_just_bow = lmer(data = df_with_baseline,
mean_relatedness ~ baseline_distance +
(1| word),
REML = FALSE)
model_just_sm = lmer(data = df_with_baseline,
mean_relatedness ~ sensorimotor_distance +
(1| word),
REML = FALSE)
summary(model_bow_sm)
anova(model_bow_sm, model_just_bow)
anova(model_bow_sm, model_just_sm)
```
# Predicting Trott & Bergen (2023)
Here, we also ask whether sensorimotor distance can account for variance in RT and accuracy on a primed sensibility judgment task.
```{r}
# ============================================================
# Load T&B 2023 data and merge with norms
# ============================================================
df_s1 <- read_csv("../../data/tb2023/polysemy_s1_final.csv")
df_s2 <- read_csv("../../data/tb2023/polysemy_s2_final.csv")
df_tb2023 <- bind_rows(df_s1, df_s2)
nrow(df_tb2023)
# Item-level aggregates
df_tb2023_agg_acc <- df_tb2023 %>%
group_by(word, same, version) %>%
summarise(accuracy = mean(correct_response), .groups = "drop")
df_tb2023_agg_rt <- df_tb2023 %>%
filter(correct_response == TRUE) %>%
group_by(word, same, version) %>%
summarise(mean_rt = mean(rt), .groups = "drop") %>%
mutate(mean_log_rt = log10(mean_rt))
# Merge with sensorimotor norms (for Part 1)
df_sm_acc <- df_rawc_with_norms %>%
select(word, same, version, sensorimotor_distance,
action_distance, perceptual_distance) %>%
inner_join(df_tb2023_agg_acc, by = c("word", "same", "version"))
df_sm_rt <- df_rawc_with_norms %>%
select(word, same, version, sensorimotor_distance,
action_distance, perceptual_distance) %>%
inner_join(df_tb2023_agg_rt, by = c("word", "same", "version"))
# Merge with full distributional sweep (for Part 2)
df_full_acc <- df_merged %>%
inner_join(df_tb2023_agg_acc, by = c("word", "same", "version"))
df_full_rt <- df_merged %>%
inner_join(df_tb2023_agg_rt, by = c("word", "same", "version"))
# Sanity check
nrow(df_sm_acc)
nrow(df_sm_rt)
nrow(df_full_acc)
nrow(df_full_rt)
```
## Part 1: SM on its own
```{r}
mod_sm_acc = lmer(data = df_sm_acc,
accuracy ~ sensorimotor_distance +
(1 | word))
summary(mod_sm_acc)
mod_sm_rt = lmer(data = df_sm_rt,
mean_log_rt ~ sensorimotor_distance +
(1 | word))
summary(mod_sm_rt)
```
## Part 2: Comopare to distributional distance
Here, we compare sensorimotor distance to the best-performing distributional distance measure from the RAW-C analysis.
```{r sm_dd_tb2023}
# --- Compute AIC for each model/layer: SM, Distributional, Hybrid ---
tb_acc_results <- df_full_acc %>%
group_by(Model, Layer, n_params, Multilingual) %>%
summarize(
r2_sm = summary(lm(accuracy ~ sensorimotor_distance))$r.squared,
aic_sm = AIC(lm(accuracy ~ sensorimotor_distance)),
r2_bert = summary(lm(accuracy ~ Distance))$r.squared,
aic_bert = AIC(lm(accuracy ~ Distance)),
r2_combined = summary(lm(accuracy ~ Distance + sensorimotor_distance))$r.squared,
aic_combined = AIC(lm(accuracy ~ Distance + sensorimotor_distance)),
.groups = "drop"
) %>%
mutate(
aic_bert_vs_sm = aic_bert - aic_sm,
aic_combined_vs_sm = aic_sm - aic_combined,
aic_combined_vs_dist = aic_bert - aic_combined
)
tb_rt_results <- df_full_rt %>%
group_by(Model, Layer, n_params, Multilingual) %>%
summarize(
r2_sm = summary(lm(mean_log_rt ~ sensorimotor_distance))$r.squared,
aic_sm = AIC(lm(mean_log_rt ~ sensorimotor_distance)),
r2_bert = summary(lm(mean_log_rt ~ Distance))$r.squared,
aic_bert = AIC(lm(mean_log_rt ~ Distance)),
r2_combined = summary(lm(mean_log_rt ~ Distance + sensorimotor_distance))$r.squared,
aic_combined = AIC(lm(mean_log_rt ~ Distance + sensorimotor_distance)),
.groups = "drop"
) %>%
mutate(
aic_bert_vs_sm = aic_bert - aic_sm,
aic_combined_vs_sm = aic_sm - aic_combined,
aic_combined_vs_dist = aic_bert - aic_combined
)
# --- Best layer per model (selected by best distributional R²) ---
best_layers_acc <- tb_acc_results %>%
group_by(Model, n_params, Multilingual) %>%
slice_max(r2_bert, n = 1) %>%
ungroup()
best_layers_rt <- tb_rt_results %>%
group_by(Model, n_params, Multilingual) %>%
slice_max(r2_bert, n = 1) %>%
ungroup()
print(best_layers_acc)
print(best_layers_rt)
# --- Summary counts ---
best_layers_acc %>%
mutate(dist_better_than_sm = aic_bert_vs_sm < -4,
hybrid_better_than_sm = aic_combined_vs_sm > 4,
hybrid_better_than_dist = aic_combined_vs_dist > 4) %>%
summarise(across(c(dist_better_than_sm,
hybrid_better_than_sm,
hybrid_better_than_dist), mean))
best_layers_rt %>%
mutate(dist_better_than_sm = aic_bert_vs_sm < -4,
hybrid_better_than_sm = aic_combined_vs_sm > 4,
hybrid_better_than_dist = aic_combined_vs_dist > 4) %>%
summarise(across(c(dist_better_than_sm,
hybrid_better_than_sm,
hybrid_better_than_dist), mean))
# --- LaTeX tables ---
best_layers_acc %>%
select(Model, aic_bert, aic_combined) %>%
mutate(across(starts_with("aic"), round)) %>%
kbl(col.names = c("Model", "Distributional", "Hybrid"),
format = "latex", booktabs = TRUE,
caption = sprintf("AIC values for predicting accuracy (Trott \\& Bergen, 2023). AIC for Sensorimotor Distance alone was %.2f.",
unique(best_layers_acc$aic_sm)),
label = "aic_tb_acc") %>%
kable_styling(latex_options = "hold_position")
best_layers_rt %>%
select(Model, aic_bert, aic_combined) %>%
mutate(across(starts_with("aic"), round)) %>%
kbl(col.names = c("Model", "Distributional", "Hybrid"),
format = "latex", booktabs = TRUE,
caption = sprintf("AIC values for predicting log RT (Trott \\& Bergen, 2023). AIC for Sensorimotor Distance alone was %.2f.",
unique(best_layers_rt$aic_sm)),
label = "aic_tb_rt") %>%
kable_styling(latex_options = "hold_position")
# ============================================================
# Figures: AIC dot plots (matching your existing style)
# ============================================================
# --- Accuracy ---
aic_long_acc <- best_layers_acc %>%
select(Model, Layer, aic_sm, aic_bert, aic_combined) %>%
pivot_longer(cols = c(aic_sm, aic_bert, aic_combined),
names_to = "type", values_to = "aic") %>%
mutate(type = case_when(
type == "aic_sm" ~ "Sensorimotor",
type == "aic_bert" ~ "Distributional",
type == "aic_combined" ~ "Hybrid"
),
type = factor(type, levels = c("Sensorimotor", "Distributional", "Hybrid")))
baseline_acc <- aic_long_acc %>% filter(type == "Sensorimotor") %>% pull(aic) %>% unique()
aic_long_acc_no_sm <- aic_long_acc %>% filter(type != "Sensorimotor")
aic_se_acc <- aic_long_acc_no_sm %>%
group_by(type) %>%
summarize(mean_aic = mean(aic),
se_aic = sd(aic) / sqrt(n()),
.groups = "drop")
# Dynamic offset for the label (5% of y-range above baseline)
offset_acc <- diff(range(aic_long_acc_no_sm$aic)) * 0.05
ggplot() +
geom_hline(yintercept = baseline_acc,
linetype = "dashed", color = "coral", linewidth = 1) +
geom_jitter(data = aic_long_acc_no_sm,
aes(x = type, y = aic, color = type),
alpha = 0.3, width = 0.1, size = 2) +
geom_point(data = aic_se_acc,
aes(x = type, y = mean_aic, color = type),
size = 5) +
geom_errorbar(data = aic_se_acc,
aes(x = type, ymin = mean_aic - se_aic,
ymax = mean_aic + se_aic, color = type),
width = 0.15, linewidth = 1.2) +
annotate("text", x = 1.5, y = baseline_acc + offset_acc,
label = "Sensorimotor", color = "coral", size = 5) +
labs(x = NULL, y = "AIC (lower is better)",
title = "Predicting Accuracy") +
scale_color_manual(values = c("Distributional" = "gray60",
"Hybrid" = "steelblue")) +
theme_minimal() +
theme(text = element_text(size = 16), legend.position = "none")
# --- RT ---
aic_long_rt <- best_layers_rt %>%
select(Model, Layer, aic_sm, aic_bert, aic_combined) %>%
pivot_longer(cols = c(aic_sm, aic_bert, aic_combined),
names_to = "type", values_to = "aic") %>%
mutate(type = case_when(
type == "aic_sm" ~ "Sensorimotor",
type == "aic_bert" ~ "Distributional",
type == "aic_combined" ~ "Hybrid"
),
type = factor(type, levels = c("Sensorimotor", "Distributional", "Hybrid")))
baseline_rt <- aic_long_rt %>% filter(type == "Sensorimotor") %>% pull(aic) %>% unique()
aic_long_rt_no_sm <- aic_long_rt %>% filter(type != "Sensorimotor")
aic_se_rt <- aic_long_rt_no_sm %>%
group_by(type) %>%
summarize(mean_aic = mean(aic),
se_aic = sd(aic) / sqrt(n()),
.groups = "drop")
offset_rt <- diff(range(aic_long_rt_no_sm$aic)) * 0.05
ggplot() +
geom_hline(yintercept = baseline_rt,
linetype = "dashed", color = "coral", linewidth = 1) +
geom_jitter(data = aic_long_rt_no_sm,
aes(x = type, y = aic, color = type),
alpha = 0.3, width = 0.1, size = 2) +
geom_point(data = aic_se_rt,
aes(x = type, y = mean_aic, color = type),
size = 5) +
geom_errorbar(data = aic_se_rt,
aes(x = type, ymin = mean_aic - se_aic,
ymax = mean_aic + se_aic, color = type),
width = 0.15, linewidth = 1.2) +
annotate("text", x = 1.5, y = baseline_rt + offset_rt,
label = "Sensorimotor", color = "coral", size = 5) +
labs(x = NULL, y = "AIC (lower is better)",
title = "Predicting RT") +
scale_color_manual(values = c("Distributional" = "gray60",
"Hybrid" = "steelblue")) +
theme_minimal() +
theme(text = element_text(size = 16), legend.position = "none")
```
## Part 3: Does SM account for sense boundaries?
```{r}
# --- Compute AIC for each model/layer: same alone vs. same + continuous ---
tb_acc_q3 <- df_full_acc %>%
group_by(Model, Layer, n_params, Multilingual) %>%
summarize(
aic_same = AIC(lm(accuracy ~ same)),
aic_same_sm = AIC(lm(accuracy ~ same + sensorimotor_distance)),
aic_same_dist = AIC(lm(accuracy ~ same + Distance)),
aic_same_both = AIC(lm(accuracy ~ same + Distance + sensorimotor_distance)),
r2_same_dist = summary(lm(accuracy ~ same + Distance))$r.squared,
.groups = "drop"
) %>%
mutate(
delta_aic_sm = aic_same - aic_same_sm, # positive = SM adds over same
delta_aic_dist = aic_same - aic_same_dist, # positive = Distributional adds over same
delta_aic_both = aic_same - aic_same_both # positive = both add over same
)
tb_rt_q3 <- df_full_rt %>%
group_by(Model, Layer, n_params, Multilingual) %>%
summarize(
aic_same = AIC(lm(mean_log_rt ~ same)),
aic_same_sm = AIC(lm(mean_log_rt ~ same + sensorimotor_distance)),
aic_same_dist = AIC(lm(mean_log_rt ~ same + Distance)),
aic_same_both = AIC(lm(mean_log_rt ~ same + Distance + sensorimotor_distance)),
r2_same_dist = summary(lm(mean_log_rt ~ same + Distance))$r.squared,
.groups = "drop"
) %>%
mutate(
delta_aic_sm = aic_same - aic_same_sm,
delta_aic_dist = aic_same - aic_same_dist,
delta_aic_both = aic_same - aic_same_both
)
# --- Best layer per model (selected by best same + Distributional R²) ---
best_layers_acc_q3 <- tb_acc_q3 %>%
group_by(Model, n_params, Multilingual) %>%
slice_max(r2_same_dist, n = 1) %>%
ungroup()
best_layers_rt_q3 <- tb_rt_q3 %>%
group_by(Model, n_params, Multilingual) %>%
slice_max(r2_same_dist, n = 1) %>%
ungroup()
print(best_layers_acc_q3)
print(best_layers_rt_q3)
# --- Summary counts: how often does each augmentation add over same alone ---
best_layers_acc_q3 %>%
mutate(sm_adds = delta_aic_sm > 4,
dist_adds = delta_aic_dist > 4,
both_add = delta_aic_both > 4) %>%
summarise(across(c(sm_adds, dist_adds, both_add), mean))
best_layers_rt_q3 %>%
mutate(sm_adds = delta_aic_sm > 4,
dist_adds = delta_aic_dist > 4,
both_add = delta_aic_both > 4) %>%
summarise(across(c(sm_adds, dist_adds, both_add), mean))
make_aic_plot_q3 <- function(best_layers_df, title) {
aic_long <- best_layers_df %>%
select(Model, Layer, aic_same, aic_same_sm, aic_same_dist, aic_same_both) %>%
pivot_longer(cols = c(aic_same, aic_same_sm, aic_same_dist, aic_same_both),
names_to = "type", values_to = "aic") %>%
mutate(type = case_when(
type == "aic_same" ~ "Same only",
type == "aic_same_sm" ~ "Same + SM",
type == "aic_same_dist" ~ "Same + Distributional",
type == "aic_same_both" ~ "Same + Both"
),
type = factor(type, levels = c("Same only", "Same + SM",
"Same + Distributional", "Same + Both")))
baseline <- aic_long %>% filter(type == "Same only") %>% pull(aic) %>% unique()
aic_long_no_base <- aic_long %>% filter(type != "Same only")
aic_se <- aic_long_no_base %>%
group_by(type) %>%
summarize(mean_aic = mean(aic),
se_aic = sd(aic) / sqrt(n()),
.groups = "drop")
offset <- diff(range(aic_long_no_base$aic)) * 0.05
ggplot() +
geom_hline(yintercept = baseline,
linetype = "dashed", color = "coral", linewidth = 1) +
geom_jitter(data = aic_long_no_base,
aes(x = type, y = aic, color = type),
alpha = 0.3, width = 0.1, size = 2) +
geom_point(data = aic_se,
aes(x = type, y = mean_aic, color = type),
size = 5) +
geom_errorbar(data = aic_se,
aes(x = type, ymin = mean_aic - se_aic,
ymax = mean_aic + se_aic, color = type),
width = 0.15, linewidth = 1.2) +
annotate("text", x = 2, y = baseline + offset,
label = "Same only", color = "coral", size = 5) +
labs(x = NULL, y = "AIC (lower is better)", title = title) +
scale_color_manual(values = c("Same + SM" = "coral",
"Same + Distributional" = "gray60",
"Same + Both" = "steelblue")) +
theme_minimal() +
theme(text = element_text(size = 16), legend.position = "none",
axis.text.x = element_text(angle = 20, hjust = 1))
}
make_aic_plot_q3(best_layers_acc_q3, "Predicting Accuracy")
make_aic_plot_q3(best_layers_rt_q3, "Predicting RT")
```
# Comparison to Trott (2024) LLM Norms
Here, we conduct a more in-depth comparison to the GPT-4-generated norms from Trott (2024).
## Load data
```{r}
df_gpt_perception = read_csv("../../data/trott2024/cs_norms_perception_gpt-4.csv")
df_gpt_action = read_csv("../../data/trott2024/cs_norms_action_gpt-4.csv")
df_both_gpt = df_gpt_perception %>%
inner_join(df_gpt_action)
df_gpt_with_human = df_both_gpt %>%
inner_join(df_contextualized_meanings)
### RAW-C With GPT and human SM distance
df_rawc_with_sm = read_csv("../../data/trott2024/rawc_with_sm.csv")
```
## Dimension-level correlation
```{r gpt_corrs}
df_summ = df_gpt_with_human %>%
summarise(Vision = cor(Vision.M, Vision, method = "spearman"),
Hearing = cor(Hearing.M, Hearing, method = "spearman"),
Touch = cor(Touch.M, Touch, method = "spearman"),
Olfaction = cor(Olfaction.M, Olfaction, method = "spearman"),
Taste = cor(Taste.M, Taste, method = "spearman"),
Interoception = cor(Interoception.M, Interoception, method = "spearman"),
Hand_arm = cor(Hand_arm.M, Hand_arm, method = "spearman"),
Foot_leg = cor(Foot_leg.M, Foot_leg, method = "spearman"),
Head = cor(Head.M, Head, method = "spearman"),
Torso = cor(Torso.M, Torso, method = "spearman"),
Mouth_throat = cor(Mouth_throat.M, Mouth_throat, method = "spearman"))
df_summ
df_long = df_summ %>%
pivot_longer(everything(), names_to = "Factor", values_to = "Correlation")
df_long %>%
ggplot(aes(x = reorder(Factor, Correlation), y = Correlation)) +
geom_bar(stat = "identity") +
labs(x = "Factor", y = "Correlation") +
theme_minimal()
```
## Compare correlation matrices
```{r gpt_corrs_comparison}
library(tidyverse)
library(ggcorrplot)
library(patchwork)
dim_order <- c("Vision", "Hearing", "Olfaction", "Taste", "Interoception",
"Touch", "Mouth_throat", "Head", "Torso", "Hand_arm", "Foot_leg")
# Human matrix: .M columns
human_cols <- df_gpt_with_human %>%
select(all_of(paste0(dim_order, ".M"))) %>%
rename_with(~ str_remove(., "\\.M"))
cors_human <- cor(human_cols, use = "pairwise.complete.obs")
# GPT matrix: bare-named columns
gpt_cols <- df_gpt_with_human %>%
select(all_of(dim_order))
cors_gpt <- cor(gpt_cols, use = "pairwise.complete.obs")
# Difference (GPT - Human)
cors_diff <- cors_gpt - cors_human
ggcorrplot(cors_human, hc.order = FALSE, type = "upper",
lab = TRUE, lab_size = 2.5) +
ggtitle("Human") +
theme(axis.text.x = element_text(size = 10, angle = 45, hjust = 1))
ggcorrplot(cors_gpt, hc.order = FALSE, type = "upper",
lab = TRUE, lab_size = 2.5) +
ggtitle("GPT-4") +
theme(axis.text.x = element_text(size = 10, angle = 45, hjust = 1))
ggcorrplot(cors_diff, hc.order = FALSE, type = "upper",
lab = TRUE, lab_size = 2.5,
colors = c("#2166AC", "white", "#B2182B")) +
ggtitle("GPT − Human") +
theme(axis.text.x = element_text(size = 10, angle = 45, hjust = 1))
### Statistical comparison
library(vegan)
# Mantel test
mantel_result <- mantel(cors_human, cors_gpt,
method = "pearson", # or "spearman"
permutations = 9999)
mantel_result
mantel_result <- mantel(cors_human, cors_gpt,
method = "spearman", # or "spearman"
permutations = 9999)
mantel_result
## Or just straight-up correlation
upper_human <- cors_human[upper.tri(cors_human)]
upper_gpt <- cors_gpt[upper.tri(cors_gpt)]
# Correlation between the two
cor(upper_human, upper_gpt, method = "pearson")
cor(upper_human, upper_gpt, method = "spearman")
```
## Substitution analyses
### Homonymy vs. Polysemy
Here, we ask whether homonyms show bigger sensorimotor distance for GPT-generated norms as well.
```{r}
model_full = lmer(data = df_rawc_with_sm,
sm_gpt ~ same +
(1 + same | word),
control=lmerControl(optimizer="bobyqa"),
REML = FALSE)
model_reduced = lmer(data = df_rawc_with_sm,
sm_gpt ~ # same +
(1 + same | word),
control=lmerControl(optimizer="bobyqa"),
REML = FALSE)
summary(model_full)
anova(model_full, model_reduced)
model_full_at = lmer(data = df_rawc_with_sm,
sm_gpt ~ same * ambiguity_type +
(1 + same | word),
control=lmerControl(optimizer="bobyqa"),
REML = FALSE)
model_reduced_at = lmer(data = df_rawc_with_sm,
sm_gpt ~ same + ambiguity_type +
(1 + same | word),
control=lmerControl(optimizer="bobyqa"),
REML = FALSE)
summary(model_full_at)
anova(model_full_at, model_reduced_at)
df_rawc_with_sm %>%
group_by(same, ambiguity_type) %>%
summarise(mean_gpt = mean(sm_gpt),
sd_gpt = sd(sm_gpt),
mean_human = mean(sm_human),
sd_human = sd(sm_human))
df_rawc_with_sm = df_rawc_with_sm %>%
mutate(Same = case_when(
same == TRUE ~ "Same Sense",
same == FALSE ~ "Different Sense"
))
df_rawc_with_sm %>%
ggplot(aes(x = sm_gpt,
y = ambiguity_type,
fill = Same)) +
geom_density_ridges2(aes(height = ..density..),
color = NA,
alpha = 0.5,
scale=0.85,
stat="density") +
labs(x = "Sensorimotor Distance (GPT-4)",
y = NULL,
fill = NULL) +
scale_fill_manual(
values = c("Same Sense" = viridisLite::viridis(2, option = "mako", begin = 0.8, end = 0.15)[1],
"Different Sense" = viridisLite::viridis(2, option = "mako", begin = 0.8, end = 0.15)[2])
) +
theme_minimal() +
theme(text = element_text(size = 20))
```
### Dominance
We also replicate the dominance analysis.
```{r gpt_sm_strength}
df_dom_gpt = df_gpt_with_human %>%
inner_join(df_dominance_individual)
df_dom_gpt <- df_dom_gpt %>%
inner_join(df_lancaster)
nrow(df_dom_gpt)
df_dom_gpt = df_dom_gpt %>%
rowwise() %>%
mutate(max_strength_gpt = max(
c(
## Modalities
Vision,
Hearing,
Olfaction,
Touch,
Taste,
Interoception,
## Effectors
Head,
Mouth_throat,
Torso,
Hand_arm,
Foot_leg
)
),
minkowski3_strength_gpt = sum(c(Vision,
Hearing,
Olfaction,
Touch,
Taste,
Interoception,
## Effectors
Head,
Mouth_throat,
Torso,
Hand_arm,
Foot_leg)^3)^(1/3),
max_strength = max(
c(
## Modalities
Vision.M,
Hearing.M,
Olfaction.M,
Touch.M,
Taste.M,
Interoception.M,
## Effectors
Head.M,
Mouth_throat.M,
Torso.M,
Hand_arm.M,
Foot_leg.M
)
),
minkowski3_strength = sum(c(Vision.M, Hearing.M, Olfaction.M, Touch.M, Taste.M,
Interoception.M, Head.M, Mouth_throat.M, Torso.M,
Hand_arm.M, Foot_leg.M)^3)^(1/3)
) %>%
ungroup()
cor.test(df_dom_gpt$max_strength_gpt, df_dom_gpt$max_strength)
cor.test(df_dom_gpt$minkowski3_strength_gpt, df_dom_gpt$minkowski3_strength)
df_dom_gpt %>%
ggplot(aes(x = minkowski3_strength_gpt,
y = minkowski3_strength)) +
geom_point(alpha = .5) +
geom_smooth(method = "lm") +
labs(x = "Minkowski-3 Strength (GPT-4)",
y = "Minkowski-3 Strength (Human)") +
theme_minimal() +
theme(text = element_text(size = 16), legend.position = "none")
```
### Does GPT sensorimotor strength predict dominance?
```{r gpt_dominance}
df_dom_gpt %>%
ggplot(aes(x = minkowski3_strength,
y = dominance)) +
geom_point(alpha = .4) +
geom_smooth(method = "lm") +
labs(x = "Minkowski-3 Sensorimotor Strength (GPT-4)",
y = "Dominance") +
theme_minimal() +
theme(text = element_text(size=16))
model_full = lmer(data = df_dom_gpt,
dominance ~
max_strength_gpt +
Max_strength.sensorimotor + Minkowski3.sensorimotor +
(1 | word),
REML = FALSE)
model_reduced = lmer(data = df_dom_gpt,
dominance ~
# max_strength_gpt +
Max_strength.sensorimotor + Minkowski3.sensorimotor +
(1 | word),
REML = FALSE)
summary(model_full)
anova(model_full, model_reduced)
model_full = lmer(data = df_dom_gpt,
dominance ~
minkowski3_strength_gpt +
Max_strength.sensorimotor + Minkowski3.sensorimotor +
(1 | word),
REML = FALSE)
model_reduced = lmer(data = df_dom_gpt,
dominance ~
# minkowski3_strength_gpt +
Max_strength.sensorimotor + Minkowski3.sensorimotor +
(1 | word),
REML = FALSE)
summary(model_full)
anova(model_full, model_reduced)
```

contextualized_norms_analysis.Rmd at commit 7f87ad2, no license · at the source

Overview

Authors: Sean Trott^1,2, Benjamin Bergen²

ORCID iDs: Sean Trott

Institutions: Rutgers, The State University of New Jersey (United States); University of California San Diego (United States)

Journal: Behavior research methods, volume 58, issue 9, article 264

Dates: received 7 February 2026; accepted 6 July 2026; published online 12 August 2026; in print 2026

Type: Research article · Language: English

License: CC BY

Identifiers: DOI 10.3758/s13428-026-03134-6 · PMID 42587187 · PMCID PMC13469423 · OpenAlex W4221144368

Open access: hybrid, a free copy (OpenAlex)

Preprint: arxiv:2203.05648

Status: code verified

Categories: human (organism)

Methods: Statistics, Machine learning, Connectivity

MeSH: Language*, Linguistics*, Psycholinguistics*, Female, Humans, Judgment, Semantics (* major topic)

Topic: Action Observation and Synchronization (Social Psychology, Psychology), according to OpenAlex

Citations: not cited yet (Europe PMC); 111 references in the paper

Abstract

Embodied theories of language emphasize the role of sensorimotor experience in linguistic knowledge. Central to testing these theories is the creation of large datasets of linguistic norms, which contain judgments about a word’s sensorimotor associations and can be used to predict human behavioral or brain data – sometimes in contrast to competing variables, such as those derived from distributional language models. Yet many of these datasets contain judgments about words in isolation, despite the fact that most words are ambiguous, making it difficult to determine which meaning of a word is characterized by its rating (e.g., “wooden table” vs. “data table”). In the current work, we introduce a new lexical resource (directly inspired by the Lancaster sensorimotor for 112 English words, each rated in four different contexts (448 sentences total). We demonstrate: first, that these ratings encode overlapping but distinct information from the Lancaster sensorimotor norms; second, that decontextualized ratings likely reflect the more dominant meaning of ambiguous words; third, that homonyms have more distinct sensorimotor profiles than polysemes; fourth, that the contextualized sensorimotor distance between two uses of an ambiguous word predicts human judgments about semantic relatedness; and fifth, that ratings derived from GPT-4 align reasonably well with human judgments. We conclude by suggesting that contextualized ratings like these can be used both to inform competing theories of semantic representations and also to evaluate or “probe” the ability of LMs to recover sensorimotor information.

Reproduced under the paper's license (CC BY), from the paper cited above.

Repository

Its files are read in the Code ↔ Paper reader above, with 6 matches between paragraphs and lines of code.

seantrott/cs_norms

License: none: the authors keep all their rights

State: the link answers, verified on 27 September 2026

Evidence: files inventoried

Commit: 7f87ad29ff1280087f995f7e691a4bb0ab12ffcb, 1 June 2026

Languages: Python (2), R (1)

Size: 42 files, 3 scripts

Software Heritage: archived

Found in: “Code Availability”

Holds: README, 1 notebook

Not found: license file, CITATION.cff, environment file, tests, continuous integration, documentation

Tools: NumPy (2 files), pandas (2 files), SciPy (2 files), broom (1 file), lme4 (1 file), lmerTest (1 file), patchwork (1 file), tidyverse (1 file)

Availability: 1 check, the latest on 27 September 2026: the link answers

27 September 2026: the link answers

4 files

src/analysis/contextualized_norms_analysis.Rmd, R, 2,607 lines, 6 matches
src/data/bow.py, Python, 76 lines
src/data/distance_to_ls.py, Python, 45 lines
README.md, Text, 43 lines

At the source: github.com/seantrott/cs_norms

Code Availability

Analysis code can be found on GitHub: https://github.com/seantrott/cs_norms.

Reproduced under the paper's license (CC BY), from the paper cited above.

Tracing map

Proposed by the machine: these links were found in the paper and verified at the source, without human review. The map will receive a Zenodo DOI once one of the paper's authors has validated it with their ORCID.

What the map holds:

1 repository of the authors' code, each at its verified commit, with its license and how the link was found in the paper;
3 scripts, each with its path and the digest of its content;
6 matches between paragraphs of the paper and lines of the code (method lexical-v1);
neither the text of the paper nor the code itself.

Its JSON (tracing-map.json) is deposited on Zenodo with its DOI once the map is validated.

Data

Datasets cited

huggingface.co/datasets/seantrott/cs_norms, at Hugging Face; found in “Data Availability”

Data Availability Statement

The contextualized sensorimotor norms are publicly available on GitHub (https://github.com/seantrott/cs_norms) and HuggingFace (https://huggingface.co/datasets/seantrott/cs_norms).

Analysis code can be found on GitHub: https://github.com/seantrott/cs_norms.

Reproduced under the paper's license (CC BY), from the paper cited above.

Open Practices Statement

All code and data can be found in a publicly available repository (see above).

Reproduced under the paper's license (CC BY), from the paper cited above.

Versions

The history of this record: each version stored by the harvester or made by a correction of its authors or of the maintainers of its code, and what changed in its facts. The texts of the paper (its abstract, its availability statements) are not part of it; versions that changed only those are not listed.

Version 3, 28 September 2026

Publisher: n/a → Springer Science+Business Media

Version 1, 27 September 2026: the first record

Recorded: type, language, journal, volume, issue, pages, dates, 2 authors, 7 MeSH terms, 59 references.

Cite

This paper

Trott, S., & Bergen, B. (2026). Contextualized sensorimotor norms: Multi-dimensional measures of sensorimotor strength for ambiguous English words, in context. Behavior research methods, 58(9), 264. https://doi.org/10.3758/s13428-026-03134-6

BibTeX

@article{trott2026contextualized, author = {Trott, Sean and Bergen, Benjamin}, title = {{Contextualized sensorimotor norms: Multi-dimensional measures of sensorimotor strength for ambiguous English words, in context}}, journal = {Behavior research methods}, year = {2026}, month = aug, volume = {58}, number = {9}, pages = {264}, publisher = {Springer Science+Business Media}, issn = {1554-351X}, doi = {10.3758/s13428-026-03134-6}, url = {https://doi.org/10.3758/s13428-026-03134-6}, pmid = {42587187}, pmcid = {PMC13469423} }

RIS

TY - JOUR AU - Trott, Sean AU - Bergen, Benjamin TI - Contextualized sensorimotor norms: Multi-dimensional measures of sensorimotor strength for ambiguous English words, in context T2 - Behavior research methods J2 - Behav Res Methods PY - 2026 DA - 2026/08/12 VL - 58 IS - 9 SP - 264 SN - 1554-351X PB - Springer Science+Business Media DO - 10.3758/s13428-026-03134-6 UR - https://doi.org/10.3758/s13428-026-03134-6 LA - en ER -

CSL-JSON

{ "id": "10.3758/s13428-026-03134-6", "type": "article-journal", "title": "Contextualized sensorimotor norms: Multi-dimensional measures of sensorimotor strength for ambiguous English words, in context", "container-title": "Behavior research methods", "author": [ { "family": "Trott", "given": "Sean" }, { "family": "Bergen", "given": "Benjamin" } ], "container-title-short": "Behav Res Methods", "volume": "58", "issue": "9", "page": "264", "DOI": "10.3758/s13428-026-03134-6", "PMID": "42587187", "PMCID": "PMC13469423", "ISSN": "1554-351X", "publisher": "Springer Science+Business Media", "URL": "https://doi.org/10.3758/s13428-026-03134-6", "language": "en", "issued": { "date-parts": [ [ 2026, 8, 12 ] ] } }

The tracing map gets a citation of its own once an author has validated it and it has a DOI.

Similar papers

The papers with a page that share the most with this one: the tools found in their code, their categories, datasets, cited references and authors, the rarest counting most.

[1] doi:10.1073/pnas.2603114123 [code]: The human hippocampus can pattern separate memories by meaning.
Journal: Proceedings of the National Academy of Sciences of the United States of America
In common: broom, lmerTest, lme4, 5 other tools
[2] doi:10.1016/j.dcn.2026.101765 [code]: Fusiform face area development correlates with development in higher-order social brain regions.
Journal: Developmental cognitive neuroscience
In common: broom, lmerTest, lme4, 5 other tools
[3] doi:10.1162/imag.a.1235 [code]: Intracranial volume: To adjust or not to adjust? It is not a matter of if, but how.
Journal: Imaging neuroscience (Cambridge, Mass.)
In common: broom, lmerTest, lme4, 5 other tools
[4] doi:10.64898/2026.03.09.710596 [code]: Infant gut microbiomes contribute to metabolic states that impact brain function
Journal: bioRxiv (preprint)
In common: broom, lmerTest, lme4, 4 other tools, 1 reference
[5] doi:10.1162/imag.a.1321 [code]: Phase similarity between similar objects indicates representational merging across retrieval training but not sleep.
Journal: Imaging neuroscience (Cambridge, Mass.)
In common: broom, lmerTest, lme4, 4 other tools, 1 reference
[6] doi:10.1093/cercor/bhag132 [code]: Spatiotemporal white-matter development across early childhood.
Journal: Cerebral cortex (New York, N.Y. : 1991)
In common: broom, lmerTest, lme4, 4 other tools
[7] doi:10.1038/s41467-026-71264-8 [code]: Habitual coffee intake shapes the gut microbiome and modifies host physiology and cognition.
Journal: Nature communications
In common: broom, lmerTest, lme4, 4 other tools
[8] doi:10.1016/j.celrep.2026.117505 [code]: Impaired spatial coding and neuronal hyperactivity in the medial entorhinal cortex of aged APP knock-in mice.
Journal: Cell reports
In common: broom, lmerTest, lme4, 4 other tools
[9] doi:10.3390/ijms27135713 [code]: Chronic Administration of Marinobufagenin in Mice Causes Hyperlocomotion and Decrease in Anxiety by Altering Monoamine Turnover Unaccompanied by Motor Deficits or Oxidative Stress.
Journal: International journal of molecular sciences
In common: broom, lmerTest, lme4, 4 other tools
[10] doi:10.1093/ageing/afag263 [code]: Cardiometabolic medication exposures and cognitive outcomes in Alzheimer's disease.
Journal: Age and ageing
In common: broom, lmerTest, lme4, 4 other tools

Contribute

The authors of this paper can claim it, correct its record and validate its tracing map, and the maintainers of its code (its owner, or a public member of its organization) correct what it says of their repository; anyone signed in can ask for its removal. Every request goes to OSCR's own machine, which answers it; your account page follows them.

Sign in with ORCID to claim this paper as one of its authors, correct its record or validate its tracing map: when the paper's metadata lists your ORCID iD, you are recognized at once. Maintainers of its code: sign in with GitHub, then claim the repository on your account page.

Request its removal

To ask OSCR to remove this record, the copies of its authors' scripts or its tracing map, use the removal request page: signed in, you say who you are, what to remove and why, then review and confirm the request. Published rules decide every request (how).

Discussion, reproductions, activity

Discussion: questions and error reports about this paper and its code, from signed-in readers and its authors. It opens with sign-in.

Reproductions: reports from readers who ran the authors' code: what they reproduced, with which environment, commit and data. It opens with sign-in.

Activity: what happens around this paper: new versions of its record, its map's validation, discussions and reproductions. It opens with sign-in.

Contextualized sensorimotor norms: Multi-dimensional measures of sensorimotor strength for ambiguous English words, in context.

Code ↔ Paper

Paper

The authors' code

Overview

Abstract

Repository

seantrott/cs_norms

Code Availability

Tracing map

Data

Datasets cited

Data Availability Statement

Open Practices Statement

Versions

Version 3, 28 September 2026

Version 1, 27 September 2026: the first record

Cite

This paper

Similar papers

Contribute

Claim this paper

Correct its record

Validate its tracing map

Add the badge to its README

Request its removal

Discussion, reproductions, activity