adaptation_specialization_material/plots/compare_corpora.py

import pandas as pd 
import numpy as np 

from os.path import join as opj

corpora = {
    "output/etm_20_pretrained": "High-energy physics",
    "output/acl_2002_2022": "Computational Linguistics"
}

time_periods = {
    "output/etm_20_pretrained": ("2000--2009", "2015--2019"),
    "output/acl_2002_2022": ("2002--2011", "2012--2022")
}

author_disambiguation = {
    "output/etm_20_pretrained": "Yes",
    "output/acl_2002_2022": "No"
}

publication_requirement = {
    "output/etm_20_pretrained": 5,
    "output/acl_2002_2022": 3    
}

df = []
for corpus in corpora:
    N = len(pd.read_csv(opj(corpus, "aggregate.csv")))
    scores = np.load(opj(corpus, "scores.npy"))
    scores = scores.mean(axis=0)
    model = scores[1]
    baseline = scores[2]

    df.append({
        "Cohort": f"{corpora[corpus]}\\newline ($N={N}$)",
        "Model\\newline $\\mu(d_{{\\mathrm{{TV}}}}(\\bm{{y_{{a}}}}, \\bm{{y_{{a}}}}^{{\\text{{pred}}}}))$": f"\\textbf{{{model:.3f}}}" if model<baseline else f"{model:.3f}",
        "Baseline\\newline $\\mu(d_{{\\mathrm{{TV}}}}(\\bm{{y_{{a}}}}, \\bm{{x_{{a}}}}))$": f"\\textbf{{{baseline:.3f}}}" if baseline<model else f"{baseline:.3f}"
    })

df = pd.DataFrame(df).set_index("Cohort").transpose()

pd.set_option('display.max_colwidth', None)

latex = df.to_latex(
    escape=False,
    multirow=True,
    column_format="|>{\\centering\\arraybackslash}m{4cm}|>{\\centering\\arraybackslash}m{5cm}|>{\\centering\\arraybackslash}m{5cm}|",
    caption="Performance of the actual model versus that of the baseline model, for i) the cohort of high-energy physicists and ii) a cohort from the ACL Anthology corpus of Computation Linguistics papers.",
    label="table:performance"
)

print(latex)

with open(opj("output", "scores.tex"), "w+") as fp:
    fp.write(latex)

df = []
for corpus in corpora:
    N = len(pd.read_csv(opj(corpus, "aggregate.csv")))
    V = len(pd.read_csv(opj(corpus, "ngrams.csv")))
    D = len(pd.read_csv(opj(corpus, "articles.csv")))

    topics = pd.read_csv(opj(corpus, "topics.csv"))
    K = len(topics[~topics["label"].str.contains("Junk")])

    df.append({
        "Cohort": f"{corpora[corpus]}",
        "Cohort size ($N$)": f"${N:,}$".replace(",",r"\,"),
        "Number of abstracts ($D$)": f"${D:,}$".replace(",",r"\,"),
        "Vocabulary size ($V$)": f"${V:,}$".replace(",",r"\,"),
        "Valid topics ($K$)": K,
        "Initial time period": time_periods[corpus][0],
        "Late time period": time_periods[corpus][1],
        "Author disambiguation": author_disambiguation[corpus],
        "Minimum publications": f"${publication_requirement[corpus]}$"
    })

df = pd.DataFrame(df).set_index("Cohort").transpose()

latex = df.to_latex(
    escape=False,
    multirow=True,
    caption="Comparison of the high-energy physics corpus and the computational linguistics corpus.",
    label="table:corpora",
    column_format="|c|c|c|"
)

print(latex)

with open(opj("output", "hep_vs_acl.tex"), "w+") as fp:
    fp.write(latex)