plot_XNLI_correlations.py

import constants
import numpy as np
import matplotlib.pyplot as plt
import matplotlib.patches as mpl_patches
from matplotlib.cm import get_cmap
from scipy.stats import pearsonr, spearmanr
from numpy.polynomial.polynomial import polyfit

corr_func = spearmanr

plt.rcParams["font.family"] = "sans-serif"
plt.rcParams["font.sans-serif"] = "Avant Garde"

name = "Accent"
cmap = get_cmap(name)  # type: matplotlib.colors.ListedColormap
colors = cmap.colors * 200  # type: list


languages = constants.get_xnli_languages()
languages2 = list(languages)

print(languages2)
demand = constants.read_gdp(languages2)
all_accs = constants.read_xnli_acc()
all_populations = constants.read_xnli_populations()

populations = [all_populations[l] for l in languages]
NS = np.sum(populations)

accuracy = [all_accs[l] for l in languages]
# Normalized utility
M = max(accuracy)
utility = [(all_accs[l] / M) for l in languages]

ec_dem1 = [demand[l] for l in languages]


for i, l in enumerate(languages):
    print(l, populations[i], accuracy[i], ec_dem1[i])

rho1 = corr_func(populations, utility)
print(f"rho (pop,utility) = {rho1}")
rho2 = corr_func(ec_dem1, utility)
print(f"rho (ec_dem1,utility) = {rho2}")


fig, ax = plt.subplots(1, 2, figsize=(10, 2))
ax[0].scatter(
    populations, utility, marker="o", s=10, facecolors="none", edgecolors="blue"
)
ax[0].set_xlabel("Population (millions)", fontsize=11)
ax[0].set_ylabel("Per-Person Utility", fontsize=11)
mmax = max(populations)
for i in range(len(populations)):
    lang = languages[i]
    # Only print names for some languages
    if populations[i] > 0.3 * mmax or utility[i] == 1:
        if lang in "slv,hrv".split(","):
            ax[0].annotate(
                lang,
                (populations[i], utility[i]),
                textcoords="offset points",
                ha="center",
                fontweight="bold",
                xytext=(0, -10),
            )
        elif lang in "ron".split(","):
            ax[0].annotate(
                lang,
                (populations[i], utility[i]),
                textcoords="offset points",
                ha="center",
                fontweight="bold",
                xytext=(11, -3),
            )
        elif lang in "hin".split(","):
            ax[0].annotate(
                lang,
                (populations[i], utility[i]),
                textcoords="offset points",
                ha="center",
                fontweight="bold",
                xytext=(11, -4),
            )
        elif lang in "deu".split(","):
            ax[0].annotate(
                lang,
                (populations[i], utility[i]),
                textcoords="offset points",
                ha="center",
                fontweight="bold",
                xytext=(13, -5),
            )
        elif lang in "cmn".split(","):
            ax[0].annotate(
                lang,
                (populations[i], utility[i]),
                textcoords="offset points",
                ha="center",
                fontweight="bold",
                xytext=(-6, 4),
            )
        else:
            ax[0].annotate(
                lang,
                (populations[i], utility[i]),
                textcoords="offset points",
                ha="center",
                fontweight="bold",
                xytext=(0, 4),
            )
p3 = np.poly1d(np.polyfit(populations, utility, 3))
x = np.linspace(0, max(populations), 1000)
ax[0].plot(x, p3(x), "-", c="blue", alpha=0.3, linewidth=10)
handles = [mpl_patches.Rectangle((0, 0), 1, 1, fc="white", ec="white", lw=0, alpha=0)]
labels = [f"ρ={rho1[0]:.2f}"]
ax[0].legend(
    handles,
    labels,
    loc="best",
    fontsize=11,
    fancybox=True,
    framealpha=0.7,
    handlelength=0,
    handletextpad=0,
)
ax[0].set_xscale("log")

ax[1].scatter(ec_dem1, utility, marker="o", s=10, facecolors="none", edgecolors="blue")
ax[1].set_xlabel(f"Economic Indicators:\nGDP ($1M) ", fontsize=11)
mmax = max(ec_dem1) * 0.3
for i in range(len(ec_dem1)):
    lang = languages[i]
    if ec_dem1[i] > mmax:
        if lang in "slv,hrv".split(","):
            ax[1].annotate(
                lang,
                (ec_dem1[i], utility[i]),
                textcoords="offset points",
                ha="center",
                fontweight="bold",
                xytext=(0, -10),
            )
        elif lang in "deu,jpn".split(","):
            ax[1].annotate(
                lang,
                (ec_dem1[i], utility[i]),
                textcoords="offset points",
                ha="center",
                fontweight="bold",
                xytext=(13, -5),
            )
        elif lang in "cmn".split(","):
            ax[1].annotate(
                lang,
                (ec_dem1[i], utility[i]),
                textcoords="offset points",
                ha="center",
                fontweight="bold",
                xytext=(-4, 4),
            )
        elif lang in "spa".split(","):
            ax[1].annotate(
                lang,
                (ec_dem1[i], utility[i]),
                textcoords="offset points",
                ha="center",
                fontweight="bold",
                xytext=(-4, 4),
            )
        elif lang in "ron".split(","):
            ax[1].annotate(
                lang,
                (ec_dem1[i], utility[i]),
                textcoords="offset points",
                ha="center",
                fontweight="bold",
                xytext=(2, 4),
            )
        elif lang in "nld".split(","):
            ax[1].annotate(
                lang,
                (ec_dem1[i], utility[i]),
                textcoords="offset points",
                ha="center",
                fontweight="bold",
                xytext=(-2, 2),
            )
        else:
            ax[1].annotate(
                lang,
                (ec_dem1[i], utility[i]),
                textcoords="offset points",
                ha="center",
                fontweight="bold",
                xytext=(0, 4),
            )
p3 = np.poly1d(np.polyfit(ec_dem1, utility, 3))
x = np.linspace(0, max(ec_dem1), 1000)
ax[1].plot(x, p3(x), "-", c="blue", alpha=0.3, linewidth=10)
handles = [mpl_patches.Rectangle((0, 0), 1, 1, fc="white", ec="white", lw=0, alpha=0)]
labels = [f"ρ={rho2[0]:.2f}"]
ax[1].legend(
    handles,
    labels,
    loc="best",
    fontsize=11,
    fancybox=True,
    framealpha=0.7,
    handlelength=0,
    handletextpad=0,
)
ax[1].set_xscale("log")


plt.tight_layout()
plt.subplots_adjust(wspace=0.2)
plt.show()

fig.savefig(f"figs/xnli_correlations.pdf", quality=100, format="pdf")