def test_fontclos(self):
# loop over SPGC books
results = []
books = get_books()
for spgc_id in books.spgc_id:
wfd = spgc_read(spgc_id)
corpus = Corpus(wfd)
# fit TTR curve for all & compare RMSE
corpus.seed = SEED
TTR = corpus.TTR
m_tokens = TTR.m_tokens.values
n_types = TTR.n_types.values
# log model
lmodel = LogModel()
predictions_log = lmodel.fit_predict(m_tokens, n_types)
rmse_log = RMSE_pct(n_types, predictions_log)
# Font-Clos model
M, N = corpus.M, corpus.N
fmodel = FontClosModel()
predictions_fontclos = fmodel.fit_predict(m_tokens, n_types)
rmse_fontclos = RMSE_pct(n_types, predictions_fontclos)
# append results
results.append((
spgc_id,
corpus.M,
corpus.N,
lmodel.M_z,
lmodel.N_z,
fmodel.gamma,
rmse_log,
rmse_fontclos,
))
# aggregate & analyze results
results = pd.DataFrame(
results,
columns=[
"id",
"m_tokens",
"n_types",
"M_z",
"N_z",
"gamma",
"RMSE_log",
"RMSE_fontclos",
],
).set_index("id")
# save results
results.to_csv("data/fontclos.csv")
# assert both models performance
assert results.RMSE_log.max() < 0.01
assert results.RMSE_fontclos.max() < 0.015
def test_high_dimensions(self):
# retrieve corpus from NLTK
filename = "melville-moby_dick.txt"
words = gutenberg.words(filename)
# initialize class
corpus = Corpus(words)
dim = 64
corpus.dimension = dim
corpus.seed = SEED
TTR = corpus.TTR
# k vector calculated correctly
raw = corpus.k[:dim]
agg = TTR.tail(1).values[0, 4:]
assert all(agg == raw)
# create & fit model
m_tokens, n_types = TTR.m_tokens, TTR.n_types
model = LogModel().fit(m_tokens, n_types)
# predict E(M) & k_n(M)
model.dimension = dim
E_m = model.predict(corpus.M)
k = model.predict_k(corpus.M, dim)
assert len(k) == dim
err = std_err(corpus.k[:dim], k)
# visualize error
if GRAPHICS_ON:
import matplotlib.pyplot as plt
plt.plot(err)
plt.title("Error Analysis")
plt.xlabel("n = k-vector index")
plt.ylabel("std error")
plt.show()
def test_basic(self):
# retrieve corpus from NLTK
filename = "blake-poems.txt"
words = gutenberg.words(filename)
# initialize class
corpus = Corpus(words)
# sanity checks
assert set(corpus.tokens) == set(words)
assert corpus.M == len(words)
assert corpus.N == len(set(words))
assert corpus.fdist.freq.sum() == corpus.M
assert corpus.fdist.shape[0] == corpus.N
assert corpus.k[0] == 0
assert corpus.k[1] == len(corpus.hapax)
assert corpus.k[2] == len(corpus.dis)
assert corpus.k[3] == len(corpus.tris)
assert corpus.k[4] == len(corpus.tetrakis)
assert corpus.k[5] == len(corpus.pentakis)
assert corpus.k[42] == len(corpus.nlegomena(42))
assert sum(corpus.k) == corpus.N
assert len(corpus.types) == corpus.N
assert corpus.WFD.equals(corpus.fdist)
assert corpus.alpha == 0.9196082282619522
assert corpus.gamma == 1.7739943128244318
assert corpus.as_datarow(7) == (
8354,
1820,
corpus.alpha,
corpus.gamma,
0,
1009,
293,
138,
72,
57,
41,
)
# sample()
assert corpus.sample(99).M == 99
assert corpus.sample(x=0).M == 0
assert corpus.sample(x=1).M == corpus.M
with self.assertRaises(Exception) as context:
corpus.sample(x=1.5) # can't over sample with replace=False
def test_random_seed(self):
# retrieve corpus from NLTK
filename = "blake-poems.txt"
words = gutenberg.words(filename)
# sampling process works
corpus = Corpus(words)
corpus.seed = SEED
before = corpus.TTR.n_types
corpus = Corpus(words)
corpus.seed = SEED
after = corpus.TTR.n_types
pd.testing.assert_series_equal(before, after)
# explicit checks
assert list(corpus.TTR.n_types[:9].values) == [
73,
115,
176,
215,
253,
285,
328,
359,
381,
]
assert corpus.sample(9).tokens == [
'"',
"O",
"Tongue",
"a",
"fear",
"go",
"of",
"the",
"the",
]
# model fitting works
TTR = corpus.TTR
model = LogModel()
model.fit(TTR.m_tokens, TTR.n_types)
before = model.params
model.fit(TTR.m_tokens, TTR.n_types)
after = model.params
assert before == after
def test_optimization(self):
# retrieve corpus from NLTK
filename = "melville-moby_dick.txt"
words = gutenberg.words(filename)
# TODO: Why SPGC model quality suffers relative to NLTK ?
# initialize class
corpus = Corpus(words)
dim = 6
corpus.dimension = dim
corpus.seed = SEED
TTR = corpus.TTR
# infer optimum sample size from observed hapax:type ratio
hapax = corpus.k[1]
model = LogModel().fit_naive(corpus.M, corpus.N, hapax)
m_tokens, n_types = TTR.m_tokens, TTR.n_types
# generate single prediction
E_m = model.predict(corpus.M)
k = model.predict_k(corpus.M, dim)
assert std_err(corpus.N, E_m) < 0.0001
assert std_err(corpus.k[1], k[1]) < 0.001
assert std_err(corpus.k[2], k[2]) < 0.005
assert std_err(corpus.k[3], k[3]) < 0.05
assert std_err(corpus.k[4], k[4]) < 0.1
assert std_err(corpus.k[5], k[5]) < 0.1
# optimized predictions: worse fit at m=M, better fit overall
E_m = model.predict(m_tokens)
RMSE_before = RMSE_pct(n_types, E_m)
E_m = model.fit_predict(m_tokens, n_types)
RMSE_after = RMSE_pct(n_types, E_m)
assert RMSE_after < RMSE_before
# draw pretty pictures
if GRAPHICS_ON:
import matplotlib.pyplot as plt
# predicted hapaxes
predictions = E_m
realization = n_types
plt.scatter(m_tokens, realization)
plt.plot(m_tokens, predictions, color="red")
plt.title("Type-Token Relation (Log Formula)")
plt.xlabel("tokens")
plt.ylabel("types")
plt.show()
# predicted hapax fraction
k = model.predict_k(m_tokens, dim)
predictions = k[:, 1] / E_m
realization = TTR.lego_1 / n_types
plt.scatter(m_tokens, realization)
plt.plot(m_tokens, predictions, color="red")
plt.title("Hapax-Token Relation (Log Formula)")
plt.xlabel("tokens")
plt.ylabel("hapax fraction")
plt.show()
# predicted dis legomena fraction
predictions = k[:, 2] / E_m
realization = TTR.lego_2 / n_types
plt.scatter(m_tokens, realization)
plt.plot(m_tokens, predictions, color="red")
plt.title("Dis-Token Relation (Log Formula)")
plt.xlabel("tokens")
plt.ylabel("dis legomena fraction")
plt.show()
def test_models(self):
# initialize class
wfd = spgc_read("PG2701_counts.txt") # moby dick
corpus = Corpus(wfd)
# build TTR curve
corpus.seed = SEED
TTR = corpus.TTR
m_tokens = TTR.m_tokens.values
n_types = TTR.n_types.values
# fit Heap's Law model to TTR curve
hmodel = HeapsModel().fit(m_tokens, n_types)
predictions_heaps = hmodel.predict(m_tokens)
H = hmodel.predict(1000)
# infinite series
imodel = InfSeriesModel(corpus)
predictions_iseries = imodel.predict(m_tokens)
I = imodel.predict(1000)
# fit logarithmic model to TTR curve
lmodel = LogModel().fit(m_tokens, n_types)
predictions_log = lmodel.predict(m_tokens)
L = lmodel.predict(1000)
# fit Font-Clos model to TTR curve
M, N = corpus.M, corpus.N
fmodel = FontClosModel().fit(m_tokens, n_types)
predictions_fontclos = lmodel.predict(m_tokens)
F = fmodel.predict(1000)
# explicit check
assert (H, I, L, F) == (756, 513, 515, 770)
# draw pretty pictures
if GRAPHICS_ON:
import matplotlib.pyplot as plt
# log-log graph of Heap's Model
plt.scatter(m_tokens, n_types)
plt.plot(m_tokens, predictions_heaps, color="red")
plt.title("Heap's Model (K,B) = (%0.4f, %0.4f)" % hmodel.params)
plt.xscale("log")
plt.yscale("log")
plt.xlabel("log(tokens)")
plt.ylabel("log(types)")
plt.show()
# normal graph of Heap's Model
plt.scatter(m_tokens, n_types)
plt.plot(m_tokens, predictions_heaps, color="red")
plt.title("Heap's Model (K,B) = (%0.4f, %0.4f)" % hmodel.params)
plt.xlabel("tokens")
plt.ylabel("types")
plt.show()
# Infinite Series Model
plt.scatter(m_tokens, n_types)
plt.plot(m_tokens, predictions_iseries, color="red")
plt.title("Infinite Series Model")
plt.xlabel("tokens")
plt.ylabel("types")
plt.show()
# Logarithmic Model
plt.scatter(m_tokens, n_types)
plt.plot(m_tokens, predictions_log, color="red")
plt.title("Logarithmic Model (M_z, N_z) = (%s, %s)" %
lmodel.params)
plt.xlabel("tokens")
plt.ylabel("types")
plt.show()
# Font-Clos Model
plt.scatter(m_tokens, n_types)
plt.plot(m_tokens, predictions_fontclos, color="red")
plt.title("Font-Clos Model (γ = %s)" % fmodel.gamma)
plt.xlabel("tokens")
plt.ylabel("types")
plt.show()
def test_spgc_nltk(self):
# NLTK-SPGC lookup
books = get_books()
# build corpi from each source
corpi = {}
for book in books.itertuples():
title = book.nltk_id.split(".")[0]
wfd = spgc_read(book.spgc_id)
corpus = Corpus(wfd)
corpi[book.spgc_id] = ("SPGC", title, corpus)
words = gutenberg.words(book.nltk_id)
corpus = Corpus(words)
corpi[book.nltk_id] = ("NLTK", title, corpus)
# fit TTR curve for all & compare RMSE
results = []
for corpus_id, (source, title, corpus) in corpi.items():
corpus.seed = SEED
TTR = corpus.TTR
m_tokens = TTR.m_tokens.values
n_types = TTR.n_types.values
# log model
model = LogModel()
predictions = model.fit_predict(m_tokens, n_types)
rmse_log = RMSE_pct(n_types, predictions)
# iseries model
predictions = InfSeriesModel(corpus).predict(m_tokens)
rmse_iseries = RMSE_pct(n_types, predictions)
# heaps model
predictions = HeapsModel().fit_predict(m_tokens, n_types)
rmse_heaps = RMSE_pct(n_types, predictions)
# append results
results.append((
corpus_id,
title,
source,
corpus.M,
corpus.N,
model.M_z,
model.N_z,
rmse_log,
rmse_iseries,
rmse_heaps,
))
# aggregate & analyze results
results = pd.DataFrame(
results,
columns=[
"id",
"title",
"source",
"m_tokens",
"n_types",
"M_z",
"N_z",
"RMSE_log",
"RMSE_iseries",
"RMSE_heaps",
],
).set_index("id")
# save results to data/books.csv
results.to_csv("data/books.csv")
# assert log model outperforms heaps
assert results.RMSE_log.max() < 0.01
assert results.RMSE_heaps.min() > 0.008