-
Notifications
You must be signed in to change notification settings - Fork 0
/
plots_pretrained.Rmd
142 lines (121 loc) · 3.55 KB
/
plots_pretrained.Rmd
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
---
title: "Gender bias of pretrained embeddings"
output: html_document
---
```{r setup, include=FALSE}
knitr::opts_chunk$set(echo = TRUE)
```
```{r, message=F, warning=F}
library(tidyverse)
library(gg.layers)
library(scales)
library(glue)
library(latex2exp)
library(cowplot)
```
## Data
```{r}
files_bias = c(
"glove-wiki-gigaword-300"="results/bias_glove-wiki-gigaword-300_FEMALE_MALE.csv",
"word2vec-google-news-300"="results/bias_word2vec-google-news-300_FEMALE_MALE.csv"
)
dfs = list()
for (n in names(files_bias)) {
dfs[[n]] = read_csv(files_bias[n], show_col_types=F)
}
```
```{r}
# modify frequency bins if needed
add_frequency_bins = function(df) {
log_freq = log10(df[["reverse_freq_rank"]])
max_value = max(log_freq)
cuts = c(0, seq(1.5, max_value, .5), max_value)
# print(cuts)
df = df %>% mutate(bins = cut(log_freq, cuts, include.lowest=T))
return(df)
}
for (n in names(dfs)) {
dfs[[n]] = add_frequency_bins(dfs[[n]])
}
```
### Explore
```{r}
# words in each frequency bin
table(dfs[[1]]$bins, useNA = "always")
table(dfs[[2]]$bins, useNA = "always")
```
## Plots
### Boxplots
```{r}
# boxplots
boxplots_plt = function(df, xlab=NULL, ylab="Female bias", title=NULL,
subtitle=NULL, effect_sizes=F) {
# bin labels
labs = levels(df[["bins"]])
labs = str_replace(labs, ",", "},10^{")
labs = str_replace(labs, "([\\[\\(])", "\\1$10^{")
labs = str_replace(labs, "([\\]\\)])", "}$\\1")
labs = str_replace(labs, "(\\])$", r"(\\])")
labs = str_replace(labs, "^(\\[)", r"(\\[)")
labs = lapply(sprintf(r'(%s)', labs), TeX)
labs = unlist(labs)
# plot
p = ggplot(df, aes(x=bins, y=bias)) +
gg.layers::geom_boxplot2(
width.errorbar=0.2, fill="lightblue", color="black") +
stat_summary(fun="mean", color="navy") +
geom_hline(yintercept=0, color="black", linetype="dashed") +
# facet_wrap(vars(model), scales="free", ncol=1) +
labs(x=xlab, y=ylab, title=title, subtitle=subtitle) +
scale_x_discrete(labels=labs, guide=guide_axis(angle=35)) +
theme_minimal() +
theme(
axis.title.x=element_text(size=15), axis.title.y=element_text(size=16),
axis.text=element_text(size=14), strip.text=element_text(size=15),
plot.subtitle=element_text(size=16)
) +
NULL
if (effect_sizes == T) {
df_effect_sizes = df %>%
group_by(bins) %>%
summarise(mean_bias = mean(bias), sd_bias = sd(bias)) %>%
mutate(ef = mean_bias / sd_bias)
y_limits = layer_scales(p)$y$get_limits()
y_adj = (y_limits[2] - y_limits[1]) * 0.07
y_text = y_limits[2] + y_adj
p = p +
geom_text(
data=df_effect_sizes,
aes(x=bins, label=round(ef, 2)), y=y_text, color="navy", size=5) +
lims(y = c(NA, y_text)) +
NULL
}
return(p)
}
```
```{r, warning=F}
last_name = names(dfs)[length(names(dfs))]
plot_list = list()
for (n in names(dfs)) {
x_label = NULL
if (n == last_name) x_label = "Frequency rank"
p = boxplots_plt(dfs[[n]], xlab=x_label, ylab=glue("Female bias\n({n})"),
effect_sizes=T)
plot_list[[n]] = p
print(p)
}
outname = glue("results/plots/boxplots_pretrained.png")
grid = plot_grid(plotlist=plot_list, ncol=1)
save_plot(outname, grid, base_height=8.5, base_width=7, dpi=300)
# NOTE we dont use facet_wrap because it is hard to use it with geom_text
```
```{r}
# # one DF with all models:
# df_full = dfs %>% bind_rows(.id="model")
#
# # One plot for all models with facet_wrap:
# p = boxplots_plt(df_full)
# outname = glue("results/plots/boxplots_pretrained.png")
# ggsave(outname, p, width=8, height=2*5, dpi=300)
# print(p)
```