MTLD words by usefuleness
defined as adult trigrams from childes
2018-05-08
library(knitr)
opts_chunk$set(echo = T, message = F, warning = F,
error = F, cache = F, tidy = F)
library(tidyverse)
library(langcog)
library(data.table)
library(feather)
library(broom)
theme_set(theme_classic(base_size = 10))Do kids that know more words at T2 (using mtld measure) know words that are more “useful” at T1? In particular, we define useful words to be words that allow you to make more trigrams in adult speech from childes (child-directed). Because high kids have bigger vocabs, we sample n words per group (here, 500).
It turns out, that the pattern of effects depends a lot on what you set the threshold value for knowing a word to be. The results shown here are where the threhold is set at 5 occurences (but it looks different if it’s set to 1.)
FINDINGS:
Min occurences of word in vocab = 1:
- At the group level, the high group produces more frequent trigrams, controling for the frequency of the individual words.
- At the kid level, there are no effect when controlling for word frequency
Min occurences of word in vocab = 5:
- At the group level, the high group produces more frequent trigrams, controling for the frequency of the individual words.
- At the kid level, high kids produce more trigrams controling for frequency of words in kids’ vocab (there are a number of kids missing here though, either because they don’t produce words at least 5 times or their words are not in the sampled words)
- At kid level, low kids produce more frequent trigrams, controling for vocab word frequency
Group level
MINCOUNT <- 5
TOPNWORDS_PER_GROUP <- 500
# Get all unique words for two groups at t1.
groups_info <- read_csv("../1_mtld_measure/data/groups_info_600_900.csv")
target_types <- read_csv("../1_mtld_measure/data/target_types_for_MTLD_kids_600_900.csv") %>%
group_by(target_child_id, tbin, gloss) %>%
summarize(count = sum(count)) %>%
ungroup() %>%
filter(count >= MINCOUNT) %>%
mutate(gloss = tolower(gloss))
t1_words <- target_types %>%
filter(tbin == "t1") %>% # #t1 only
left_join(groups_info %>%
mutate(target_child_id = as.numeric(target_child_id))) %>% # merge in group info
group_by(delta_resid_group, gloss) %>% # get glosses by group
summarize(n = n()) %>% # get gloss counts by group
top_n(TOPNWORDS_PER_GROUP) %>%
group_by(delta_resid_group) %>%
slice(1:TOPNWORDS_PER_GROUP) %>%
arrange(delta_resid_group, -n)childes_trigrams <- read_csv("data/trigrams/adult_childes_trigrams_turns.csv") %>%
data.table()
t1_words_high <- t1_words %>%
ungroup()%>%
filter(delta_resid_group == "high") %>%
select(gloss) %>%
unlist(use.names = F)
t1_words_low <- t1_words %>%
ungroup() %>%
filter(delta_resid_group == "low") %>%
select(gloss) %>%
unlist(use.names = F)
high_trigrams <- childes_trigrams[w1 %in% t1_words_high &
w2 %in% t1_words_high &
w3 %in% t1_words_high] %>%
mutate(vocab_group = "high")
low_trigrams <- childes_trigrams[w1 %in% t1_words_low &
w2 %in% t1_words_low &
w3 %in% t1_words_low] %>%
mutate(vocab_group = "low")
all_trigrams <- bind_rows(high_trigrams, low_trigrams)number of trigrams
count(all_trigrams, vocab_group) %>%
kable()| vocab_group | n |
|---|---|
| high | 100231 |
| low | 25762 |
frequency of trigrams
all_trigrams %>%
mutate(log_freq = log(freq)) %>%
group_by(vocab_group) %>%
multi_boot_standard(col = "log_freq", na.rm = T) %>%
ggplot(aes( x = vocab_group, y = mean, group = vocab_group, fill = vocab_group)) +
geom_bar(position = "dodge", stat = "identity") +
geom_linerange(aes(ymin = ci_lower, ymax = ci_upper),
position=position_dodge(width=0.9)) +
theme_classic()
#summarize(mean_freq= mean(log_freq),
# ci_lower_freq = mean(log_freq) - (1.96 * sd(log_freq)),
# ci_upper_freq = mean(log_freq) + (1.96 * sd(log_freq)))
all_trigrams %>%
group_by(vocab_group) %>%
multi_boot_standard(col = "freq", na.rm = T) %>%
kable()| vocab_group | ci_lower | ci_upper | mean |
|---|---|---|---|
| high | 4.023067 | 4.362018 | 4.16881 |
| low | 3.726431 | 4.612536 | 4.09417 |
freq <- read_tsv("../1_mtld_measure/data/control_variables/SUBTLEXus_corpus.txt") %>%
select(Word, Lg10WF)
freq_tidy <- freq %>%
mutate(Word = tolower(Word)) %>%
rename(w1 = Word,
log_freq1 = Lg10WF) %>%
mutate(w2 = w1,
log_freq2 = log_freq1,
w3 = w1,
log_freq3 = log_freq1)
all_trigrams_with_freq <- all_trigrams %>%
left_join(freq_tidy %>% select(w1, log_freq1)) %>%
left_join(freq_tidy %>% select(w2, log_freq2)) %>%
left_join(freq_tidy %>% select(w3, log_freq3)) %>%
rowwise()%>%
mutate(mean_word_freq = mean(c(log_freq1, log_freq2, log_freq3),
na.rm = T),
max_word_freq = max(c(log_freq1, log_freq2, log_freq3),
na.rm = T),
log_trigram_freq = log(freq)) %>%
select(vocab_group, log_trigram_freq, max_word_freq, mean_word_freq)
lm(log_trigram_freq ~ vocab_group + mean_word_freq, d = all_trigrams_with_freq) %>%
summary()##
## Call:
## lm(formula = log_trigram_freq ~ vocab_group + mean_word_freq,
## data = all_trigrams_with_freq)
##
## Residuals:
## Min 1Q Median 3Q Max
## -1.5250 -0.5693 -0.2550 0.3161 7.3738
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) -3.941802 0.035122 -112.23 < 2e-16 ***
## vocab_grouplow -0.039697 0.005960 -6.66 2.74e-11 ***
## mean_word_freq 0.866465 0.006744 128.48 < 2e-16 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 0.853 on 125950 degrees of freedom
## (40 observations deleted due to missingness)
## Multiple R-squared: 0.116, Adjusted R-squared: 0.116
## F-statistic: 8266 on 2 and 125950 DF, p-value: < 2.2e-16lm(log_trigram_freq ~ vocab_group + max_word_freq, d = all_trigrams_with_freq %>% filter(is.finite(max_word_freq))) %>%
summary()##
## Call:
## lm(formula = log_trigram_freq ~ vocab_group + max_word_freq,
## data = all_trigrams_with_freq %>% filter(is.finite(max_word_freq)))
##
## Residuals:
## Min 1Q Median 3Q Max
## -0.8122 -0.5972 -0.3792 0.3139 7.7071
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) -2.043505 0.037827 -54.023 < 2e-16 ***
## vocab_grouplow -0.026020 0.006222 -4.182 2.89e-05 ***
## max_word_freq 0.452620 0.006565 68.949 < 2e-16 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 0.8905 on 125950 degrees of freedom
## Multiple R-squared: 0.03654, Adjusted R-squared: 0.03652
## F-statistic: 2388 on 2 and 125950 DF, p-value: < 2.2e-16Kids in the high group produce high frequency trigrams, contorling for the mean frequency of the words/max frequency of the words in the trigrams.
Individual kids
good_types_t1 <- target_types %>%
left_join(groups_info %>%
mutate(target_child_id = as.numeric(target_child_id))) %>%
filter(tbin == "t1") %>%
filter(gloss %in% t1_words$gloss)
get_trigrams_by_kid <- function(df, all_trigrams, measure, freq){
current_trigrams <- all_trigrams[w1 %in% df$gloss &
w2 %in% df$gloss &
w3 %in% df$gloss]
vocab_freqs <- df %>%
left_join(freq, by = c("gloss" = "Word"))
mean_freq <- mean(vocab_freqs$Lg10WF, na.rm = T )
median_freq <- median(vocab_freqs$Lg10WF, na.rm = T )
max_freq <- median(vocab_freqs$Lg10WF, na.rm = T )
if (measure == "num"){
list(log_trigram_num = log(nrow(current_trigrams)),
mean_word_freq = mean_freq,
median_word_freq = median_freq,
max_word_freq = max_freq)
} else if (measure == "freq"){
list(mean_trigram_freq = mean(log(current_trigrams$freq)),
mean_word_freq = mean_freq,
median_word_freq = median_freq,
max_word_freq = max_freq)
}
}number of trigrams
trigram_num_by_kid <- good_types_t1 %>%
nest(-target_child_id, -tbin) %>%
mutate(kid_measures =
map(data, get_trigrams_by_kid,
childes_trigrams, "num", freq),
kid_measure_names = list(c("log_num_trigrams", # this is bad
"mean_word_freq",
"median_word_freq",
"max_word_freq"))) %>%
select(-data) %>%
unnest() %>%
mutate(kid_measures = unlist(kid_measures)) %>%
left_join(groups_info %>% select(target_child_id, delta_resid_group)) %>%
spread(kid_measure_names, kid_measures)
ggplot(trigram_num_by_kid,
aes(x = log_num_trigrams,
fill = delta_resid_group,
group = delta_resid_group)) +
geom_density(alpha = .4) +
theme_classic()
ggplot(trigram_num_by_kid,
aes(x = log_num_trigrams,
fill = delta_resid_group)) +
geom_histogram() +
theme_classic()
trigram_num_by_kid %>%
group_by(delta_resid_group) %>%
multi_boot_standard(col = "log_num_trigrams", na.rm = T) %>%
ggplot(aes( x = delta_resid_group, y = mean,
group = delta_resid_group, fill = delta_resid_group)) +
geom_bar(position = "dodge", stat = "identity") +
geom_linerange(aes(ymin = ci_lower, ymax = ci_upper),
position=position_dodge(width=0.9)) +
theme_classic()
lm(log_num_trigrams ~ delta_resid_group + mean_word_freq,
d = trigram_num_by_kid %>% filter(is.finite(log_num_trigrams))) %>%
summary()##
## Call:
## lm(formula = log_num_trigrams ~ delta_resid_group + mean_word_freq,
## data = trigram_num_by_kid %>% filter(is.finite(log_num_trigrams)))
##
## Residuals:
## Min 1Q Median 3Q Max
## -6.5340 -1.7205 0.6778 2.1102 5.7103
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 16.1982 2.6749 6.056 5.34e-08 ***
## delta_resid_grouplow -2.6669 0.6923 -3.852 0.000247 ***
## mean_word_freq -2.0165 0.5970 -3.378 0.001168 **
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 2.979 on 74 degrees of freedom
## (3 observations deleted due to missingness)
## Multiple R-squared: 0.3053, Adjusted R-squared: 0.2865
## F-statistic: 16.26 on 2 and 74 DF, p-value: 1.403e-06lm(log_num_trigrams ~ delta_resid_group + median_word_freq,
d = trigram_num_by_kid %>% filter(is.finite(log_num_trigrams))) %>%
summary()##
## Call:
## lm(formula = log_num_trigrams ~ delta_resid_group + median_word_freq,
## data = trigram_num_by_kid %>% filter(is.finite(log_num_trigrams)))
##
## Residuals:
## Min 1Q Median 3Q Max
## -6.5658 -2.0368 0.4806 2.0022 5.5137
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 14.3034 2.2911 6.243 2.44e-08 ***
## delta_resid_grouplow -2.6411 0.7028 -3.758 0.000339 ***
## median_word_freq -1.5834 0.5070 -3.123 0.002552 **
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 3.009 on 74 degrees of freedom
## (3 observations deleted due to missingness)
## Multiple R-squared: 0.2916, Adjusted R-squared: 0.2724
## F-statistic: 15.23 on 2 and 74 DF, p-value: 2.892e-06frequency of trigrams
trigram_freqs_by_kid <- good_types_t1 %>%
nest(-target_child_id, -tbin) %>%
mutate(kid_measures =
map(data, get_trigrams_by_kid,
childes_trigrams, "freq", freq),
kid_measure_names = list(c("log_mean_trigram_frequency", # this is bad
"mean_word_freq",
"median_word_freq",
"max_word_freq"))) %>%
select(-data) %>%
unnest() %>%
mutate(kid_measures = unlist(kid_measures)) %>%
left_join(groups_info %>% select(target_child_id, delta_resid_group)) %>%
spread(kid_measure_names, kid_measures)
ggplot(trigram_freqs_by_kid,
aes(x = log_mean_trigram_frequency,
fill = delta_resid_group,
group = delta_resid_group)) +
geom_density(alpha = .4) +
theme_classic()
trigram_freqs_by_kid %>%
group_by(delta_resid_group) %>%
multi_boot_standard(col = "log_mean_trigram_frequency", na.rm = T) %>%
ggplot(aes( x = delta_resid_group, y = mean, group = delta_resid_group, fill = delta_resid_group)) +
geom_bar(position = "dodge", stat = "identity") +
geom_linerange(aes(ymin = ci_lower, ymax = ci_upper),
position=position_dodge(width=0.9)) +
theme_classic()
lm(log_mean_trigram_frequency ~ delta_resid_group + mean_word_freq, d = trigram_freqs_by_kid) %>%
summary()##
## Call:
## lm(formula = log_mean_trigram_frequency ~ delta_resid_group +
## mean_word_freq, data = trigram_freqs_by_kid)
##
## Residuals:
## Min 1Q Median 3Q Max
## -1.3618 -0.4136 -0.1613 0.1153 3.1254
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) -2.4181 0.7778 -3.109 0.00267 **
## delta_resid_grouplow 0.5240 0.2013 2.603 0.01116 *
## mean_word_freq 0.7703 0.1736 4.437 3.12e-05 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 0.8664 on 74 degrees of freedom
## (9 observations deleted due to missingness)
## Multiple R-squared: 0.3029, Adjusted R-squared: 0.284
## F-statistic: 16.08 on 2 and 74 DF, p-value: 1.593e-06lm(log_mean_trigram_frequency ~ delta_resid_group + median_word_freq, d = trigram_freqs_by_kid) %>%
summary()##
## Call:
## lm(formula = log_mean_trigram_frequency ~ delta_resid_group +
## median_word_freq, data = trigram_freqs_by_kid)
##
## Residuals:
## Min 1Q Median 3Q Max
## -1.3375 -0.4262 -0.1738 0.1008 3.1990
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) -1.4487 0.6840 -2.118 0.03752 *
## delta_resid_grouplow 0.5309 0.2098 2.531 0.01351 *
## median_word_freq 0.5493 0.1513 3.629 0.00052 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 0.8982 on 74 degrees of freedom
## (9 observations deleted due to missingness)
## Multiple R-squared: 0.2508, Adjusted R-squared: 0.2305
## F-statistic: 12.38 on 2 and 74 DF, p-value: 2.296e-05