Tool for exploring types, tokens, and n-legomena relationships in text. Based on Davis 2019 research paper.
pip install legomena
This package may be driven by any data source, but the author has tested two: the Natural Language ToolKit and the Standard Project Gutenberg Corpus. The former being the gold standard of python NLP applications, but having a rather measly 18-book gutenberg corpus. The latter containing the full 55,000+ book gutenberg corpus, already tokenized and counted. NOTE: The overlap of the two datasets do not agree in their exact type/token counts, their methodology differing, but this package takes type/token counts as raw data and is therefore methodology-agnostic.
# moby dick from NLTK
import nltk
nltk.download("gutenberg")
from nltk.corpus import gutenberg
words = gutenberg.words("melville-moby_dick.txt")
corpus = Corpus(words)
assert corpus.M, corpus.N == (260819, 19317)
# moby dick from SPGC
# NOTE: download and unzip https://zenodo.org/record/2422561/files/SPGC-counts-2018-07-18.zip into DATA_FOLDER
import pandas as pd
fname = "%s/SPGC-counts-2018-07-18/PG2701_counts.txt" % DATA_FOLDER
with open(fname) as f:
df = pd.read_csv(f, delimiter="\t", header=None, names=["word", "freq"])
f.close()
wfd = {str(row.word): int(row.freq) for row in df.itertuples()}
corpus = Corpus(wfd)
assert corpus.M, corpus.N == (210258, 16402)
Demo notebooks may be found here. Unit tests may be found here.
# basic properties
corpus.tokens # list of tokens
corpus.types # list of types
corpus.fdist # word frequency distribution dataframe
corpus.WFD # alias for corpus.fdist
corpus.M # number of tokens
corpus.N # number of types
corpus.k # n-legomena vector
corpus.k[n] # n-legomena count (n=1 -> number of hapaxes)
corpus.hapax # list of hapax legomena, alias for corpus.nlegomena(1)
corpus.dis # list of dis legomena, alias for corpus.nlegomena(2)
corpus.tris # list of tris legomena, alias for corpus.nlegomena(3)
corpus.tetrakis # list of tetrakis legomena, alias for corpus.nlegomena(4)
corpus.pentakis # list of pentakis legomena, alias for corpus.nlegomena(5)
# advanced properties
corpus.options # tuple of optional settings
corpus.resolution # number of samples to take to calculate TTR curve
corpus.dimension # n-legomena vector length to pre-compute (max 6)
corpus.seed # random number seed for sampling TTR data
corpus.TTR # type-token ratio dataframe
# basic functions
corpus.nlegomena(n:int) # list of types occurring exactly n times
corpus.sample(m:int) # samples m tokens from corpus *without replacement*
corpus.sample(x:float) # samples proportion x of corpus *without replacement*
There are a variety of models in the literature predicting number of types as a function of tokens, the most well-known being Heap's Law. Here are a few implemented, overlaid by the Corpus class.
# three models
model = HeapsModel() # Heap's Law
model = InfSeriesModel(corpus) # Infinite Series Model [Davis 2019]
model = LogModel() # Logarithmic Model [Davis 2019]
model = FontClosModel # Font-Clos Model [Font-Clos 2018]
# model fitting
m_tokens = corpus.TTR.m_tokens
n_types = corpus.TTR.n_types
model.fit(m_tokens, n_types)
predictions = model.fit_predict(m_tokens, n_types)
# model parameters
model.params
# model predictions
predictions = model.predict(m_tokens)
# log model only
dim = corpus.dimension
predicted_k = model.predict_k(m_tokens, dim)
Check out the demo app to explore the type-token and n-legomena counts of a few Project Gutenberg books.