File size: 4,082 Bytes
ac1fe46 | 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 | """Headline plot for exp0: the within-posts karma effect.
Comparing AF posts only to AF posts, denser rationalist dialect still tracks
modestly higher karma. This isolates the real effect from the posts-vs-comments
confound that inflates the pooled correlation.
Output: figures/headline_within_posts_karma.png
"""
import json
import os
import sys
from pathlib import Path
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
from scipy import stats
PROJECT_DIR = os.path.dirname(os.path.abspath(__file__))
sys.path.insert(0, PROJECT_DIR)
from config import COMPILED, count_all_terms, word_count
DATA = Path(PROJECT_DIR) / "data"
FIG = Path(PROJECT_DIR) / "figures"
FIG.mkdir(parents=True, exist_ok=True)
START_YEAR = 2022
def load_posts():
rows = []
with open(DATA / "alignmentforum_posts.jsonl") as f:
for line in f:
p = json.loads(line)
text = p.get("body") or ""
wc = max(1, word_count(text))
counts = count_all_terms(text)
rate = sum((c / wc) * 10000 for c in counts.values())
rows.append({"id": p["id"], "posted_at": p.get("posted_at"),
"base_score": p.get("base_score"), "dialect_rate": rate})
df = pd.DataFrame(rows)
df["posted_at"] = pd.to_datetime(df["posted_at"], errors="coerce", utc=True)
df["year"] = df["posted_at"].dt.year
df["base_score"] = pd.to_numeric(df["base_score"], errors="coerce")
# posts only, 2022-2025 (drop the partial-2026 crawl artifact), valid karma
df = df[(df["year"] >= START_YEAR) & (df["year"] <= 2025)]
df = df.dropna(subset=["base_score"]).reset_index(drop=True)
return df
def main():
df = load_posts()
n = len(df)
rho, p = stats.spearmanr(df["dialect_rate"], df["base_score"])
pear_r, pear_p = stats.pearsonr(df["dialect_rate"], df["base_score"])
print(f"AF posts only, 2022-2025: n={n}")
print(f" Spearman rho={rho:+.3f} (p={p:.2g})")
print(f" Pearson r ={pear_r:+.3f} (p={pear_p:.2g})")
bins = [-0.001, 0, 5, 15, 30, 1e6]
labels = ["0", "(0, 5]", "(5, 15]", "(15, 30]", "> 30"]
df["bin"] = pd.cut(df["dialect_rate"], bins=bins, labels=labels)
g = df.groupby("bin", observed=True).agg(
mean_karma=("base_score", "mean"),
median_karma=("base_score", "median"),
std=("base_score", "std"),
n=("base_score", "count"),
).reindex(labels)
ci = 1.96 * g["std"] / np.sqrt(g["n"].clip(lower=1))
fig, ax = plt.subplots(figsize=(9, 5.5))
x = np.arange(len(labels))
ax.bar(x, g["mean_karma"], 0.62, yerr=ci, capsize=5,
color="tab:blue", alpha=0.88, label="mean karma (±95% CI)",
error_kw={"elinewidth": 1.4, "ecolor": "black"})
ax.plot(x, g["median_karma"], "D--", color="tab:orange", markersize=8,
linewidth=1.8, label="median karma")
for i in x:
top = g["mean_karma"].iloc[i] + (ci.iloc[i] if not np.isnan(ci.iloc[i]) else 0)
ax.text(i, top + 2, f"n={int(g['n'].iloc[i])}", ha="center",
va="bottom", fontsize=9, color="dimgray")
ax.set_xticks(x)
ax.set_xticklabels(labels)
ax.set_xlabel("rationalist-dialect density (matches / 10k words, binned)", fontsize=11)
ax.set_ylabel("post karma (base_score)", fontsize=11)
ax.set_title("Within Alignment Forum posts, denser rationalist dialect\n"
"tracks modestly higher karma", fontsize=13, fontweight="bold")
ax.legend(fontsize=10, loc="upper left")
ax.grid(alpha=0.3, axis="y")
ax.margins(y=0.34)
ax.text(0.975, 0.965,
f"posts only, 2022–2025 n = {n:,}\n"
f"Spearman ρ = {rho:+.2f} (p = {p:.1g})",
transform=ax.transAxes, ha="right", va="top", fontsize=10,
bbox=dict(boxstyle="round,pad=0.45", facecolor="whitesmoke",
edgecolor="gray"))
fig.tight_layout()
out = FIG / "headline_within_posts_karma.png"
fig.savefig(out, dpi=140)
plt.close(fig)
print(f"wrote {out}")
if __name__ == "__main__":
main()
|