agalma / word2vec.py
Mark7549's picture
removed forms for first 2 tabs and used cache to make program faster
cdb0a70
raw
history blame
13.8 kB
from gensim.models import Word2Vec
from collections import defaultdict
import os
import tempfile
import pandas as pd
import xlsxwriter
def load_all_models():
'''
Load all word2vec models
'''
archaic = ('archaic', load_word2vec_model('models/archaic_cbow.model'))
classical = ('classical', load_word2vec_model('models/classical_cbow.model'))
early_roman = ('early_roman', load_word2vec_model('models/early_roman_cbow.model'))
hellen = ('hellen', load_word2vec_model('models/hellen_cbow.model'))
late_roman = ('late_roman', load_word2vec_model('models/late_roman_cbow.model'))
return [archaic, classical, early_roman, hellen, late_roman]
def load_selected_models(selected_models):
'''
Load the selected word2vec models
'''
models = []
for model in selected_models:
if model == "Early Roman":
model = "early_roman"
elif model == "Late Roman":
model = "late_roman"
elif model == "Hellenistic":
model = "hellen"
model_name = model.lower() + "_cbow"
models.append([model_name, load_word2vec_model(f'models/{model_name}.model')])
return models
def load_word2vec_model(model_path):
'''
Load a word2vec model from a file
'''
return Word2Vec.load(model_path)
def get_word_vector(model, word):
'''
Return the word vector of a word
'''
return model.wv[word]
def iterate_over_words(model):
'''
Iterate over all words in the vocabulary and print their vectors
'''
index = 0
for word, index in model.wv.key_to_index.items():
vector = get_word_vector(model, word)
print(f'{index} Word: {word}, Vector: {vector}')
index += 1
def model_dictionary(model):
'''
Return the dictionary of the word2vec model
Key is the word and value is the vector of the word
'''
dict = defaultdict(list)
for word, index in model.wv.key_to_index.items():
vector = get_word_vector(model, word)
dict[word] = vector
return dict
def dot_product(vector_a, vector_b):
'''
Return the dot product of two vectors
'''
return sum(a * b for a, b in zip(vector_a, vector_b))
def magnitude(vector):
'''
Return the magnitude of a vector
'''
return sum(x**2 for x in vector) ** 0.5
def cosine_similarity(vector_a, vector_b):
'''
Return the cosine similarity of two vectors
'''
dot_prod = dot_product(vector_a, vector_b)
mag_a = magnitude(vector_a)
mag_b = magnitude(vector_b)
# Avoid division by zero
if mag_a == 0 or mag_b == 0:
return 0.0
similarity = dot_prod / (mag_a * mag_b)
return "{:.2f}".format(similarity)
def get_cosine_similarity(word1, time_slice_1, word2, time_slice_2):
'''
Return the cosine similarity of two words
'''
# TO DO: MOET NETTER
# Return if path does not exist
time_slice_1 = convert_time_name_to_model(time_slice_1)
time_slice_2 = convert_time_name_to_model(time_slice_2)
if not os.path.exists(f'models/{time_slice_1}.model'):
return
model_1 = load_word2vec_model(f'models/{time_slice_1}.model')
model_2 = load_word2vec_model(f'models/{time_slice_2}.model')
dict_1 = model_dictionary(model_1)
dict_2 = model_dictionary(model_2)
return cosine_similarity(dict_1[word1], dict_2[word2])
def get_cosine_similarity_one_word(word, time_slice1, time_slice2):
'''
Return the cosine similarity of one word in two different time slices
'''
# Return if path does not exist
if not os.path.exists(f'models/{time_slice1}.model') or not os.path.exists(f'models/{time_slice2}.model'):
return
model1 = load_word2vec_model(f'models/{time_slice1}.model')
model2 = load_word2vec_model(f'models/{time_slice2}.model')
dict1 = model_dictionary(model1)
dict2 = model_dictionary(model2)
return cosine_similarity(dict1[word], dict2[word])
def validate_nearest_neighbours(word, n, models):
'''
Validate the input of the nearest neighbours function
'''
if word == '' or n == '' or models == []:
return False
return True
def convert_model_to_time_name(model_name):
'''
Convert the model name to the time slice name
'''
if model_name == 'archaic_cbow' or model_name == 'archaic':
return 'Archaic'
elif model_name == 'classical_cbow' or model_name == 'classical':
return 'Classical'
elif model_name == 'early_roman_cbow' or model_name == 'early_roman':
return 'Early Roman'
elif model_name == 'hellen_cbow' or model_name == 'hellen':
return 'Hellenistic'
elif model_name == 'late_roman_cbow' or model_name == 'late_roman':
return 'Late Roman'
def convert_time_name_to_model(time_name):
'''
Convert the time slice name to the model name
'''
if time_name == 'Archaic':
return 'archaic_cbow'
elif time_name == 'Classical':
return 'classical_cbow'
elif time_name == 'Early Roman':
return 'early_roman_cbow'
elif time_name == 'Hellenistic':
return 'hellen_cbow'
elif time_name == 'Late Roman':
return 'late_roman_cbow'
elif time_name == 'classical':
return 'Classical'
elif time_name == 'early_roman':
return 'Early Roman'
elif time_name == 'hellen':
return 'Hellenistic'
elif time_name == 'late_roman':
return 'Late Roman'
elif time_name == 'archaic':
return 'Archaic'
def get_nearest_neighbours2(word, n=10, models=load_all_models()):
'''
Return the nearest neighbours of a word
word: the word for which the nearest neighbours are calculated
time_slice_model: the word2vec model of the time slice of the input word
models: list of tuples with the name of the time slice and the word2vec model (default: all in ./models)
n: the number of nearest neighbours to return (default: 10)
Return: list of tuples with the word, the time slice and
the cosine similarity of the nearest neighbours
'''
time_slice_model = load_word2vec_model(f'models/{time_slice_model}.model')
vector_1 = get_word_vector(time_slice_model, word)
nearest_neighbours = []
# Iterate over all models
for model in models:
model_name = model[0]
time_name = convert_model_to_time_name(model_name)
model = model[1]
# Iterate over all words of the model
for word, index in model.wv.key_to_index.items():
# Vector of the current word
vector_2 = get_word_vector(model, word)
# Calculate the cosine similarity between current word and input word
cosine_similarity_vectors = cosine_similarity(vector_1, vector_2)
# If the list of nearest neighbours is not full yet, add the current word
if len(nearest_neighbours) < n:
nearest_neighbours.append((word, time_name, cosine_similarity_vectors))
# If the list of nearest neighbours is full, replace the word with the smallest cosine similarity
else:
smallest_neighbour = min(nearest_neighbours, key=lambda x: x[2])
if cosine_similarity_vectors > smallest_neighbour[2]:
nearest_neighbours.remove(smallest_neighbour)
nearest_neighbours.append((word, time_name, cosine_similarity_vectors))
return sorted(nearest_neighbours, key=lambda x: x[2], reverse=True)
def get_nearest_neighbours(target_word, n=10, models=load_all_models()):
"""
Return the nearest neighbours of a word for the given models
word: the word for which the nearest neighbours are calculated
n: the number of nearest neighbours to return (default: 10)
models: list of tuples with the name of the time slice and the word2vec model (default: all in ./models)
Return: { 'model_name': [(word, cosine_similarity), ...], ... }
"""
nearest_neighbours = {}
# Iterate over models and compute nearest neighbours
for model in models:
model_neighbours = []
model_name = convert_model_to_time_name(model[0])
model = model[1]
vector_1 = get_word_vector(model, target_word)
# Iterate over all words of the model
for word, index in model.wv.key_to_index.items():
vector_2 = get_word_vector(model, word)
cosine_sim = cosine_similarity(vector_1, vector_2)
# If the list of nearest neighbours is not full yet, add the current word
if len(model_neighbours) < n:
model_neighbours.append((word, cosine_sim))
else:
# If the list of nearest neighbours is full, replace the word with the smallest cosine similarity
smallest_neighbour = min(model_neighbours, key=lambda x: x[1])
if cosine_sim > smallest_neighbour[1]:
model_neighbours.remove(smallest_neighbour)
model_neighbours.append((word, cosine_sim))
# Sort the nearest neighbours by cosine similarity
model_neighbours = sorted(model_neighbours, key=lambda x: x[1], reverse=True)
# Add the model name and the nearest neighbours to the dictionary
nearest_neighbours[model_name] = model_neighbours
return nearest_neighbours
def get_nearest_neighbours_vectors(word, time_slice_model, n=15):
"""
Returns the vectors of the nearest neighbours of a word
"""
model_name = convert_model_to_time_name(time_slice_model)
time_slice_model = load_word2vec_model(f'models/{time_slice_model}.model')
vector_1 = get_word_vector(time_slice_model, word)
nearest_neighbours = []
for word, index in time_slice_model.wv.key_to_index.items():
vector_2 = get_word_vector(time_slice_model, word)
cosine_sim = cosine_similarity(vector_1, vector_2)
if len(nearest_neighbours) < n:
nearest_neighbours.append((word, model_name, vector_2, cosine_sim))
else:
smallest_neighbour = min(nearest_neighbours, key=lambda x: x[3])
if cosine_sim > smallest_neighbour[3]:
nearest_neighbours.remove(smallest_neighbour)
nearest_neighbours.append((word, model_name, vector_2, cosine_sim))
return sorted(nearest_neighbours, key=lambda x: x[3], reverse=True)
def write_to_file(data):
'''
Write the data to a file
'''
# Create random tmp file name
temp_file_descriptor, temp_file_path = tempfile.mkstemp(prefix="temp_", suffix=".txt", dir="/tmp")
os.close(temp_file_descriptor)
# Write data to the temporary file
with open(temp_file_path, 'w') as temp_file:
temp_file.write(str(data))
return temp_file_path
def store_df_in_temp_file(all_dfs):
'''
Store the dataframe in a temporary file
'''
# Define directory for temporary files
temp_dir = "./downloads/nn"
# Create the directory if it doesn't exist
os.makedirs(temp_dir, exist_ok=True)
# Create random tmp file name
_, temp_file_path = tempfile.mkstemp(prefix="temp_", suffix=".xlsx", dir=temp_dir)
# Concatenate all dataframes
df = pd.concat([df for _, df in all_dfs], axis=1, keys=[model for model, _ in all_dfs])
# Create an ExcelWriter object
with pd.ExcelWriter(temp_file_path, engine='xlsxwriter') as writer:
# Create a new sheet
worksheet = writer.book.add_worksheet('Results')
# Write text before DataFrames
start_row = 0
for model, df in all_dfs:
# Write model name as text
worksheet.write(start_row, 0, f"Model: {model}")
# Write DataFrame
df.to_excel(writer, sheet_name='Results', index=False, startrow=start_row + 1, startcol=0)
# Update start_row for the next model
start_row += df.shape[0] + 3 # Add some space between models
return temp_file_path
def check_word_in_models(word):
"""
Check in which models a word occurs.
"""
all_models = load_all_models()
eligible_models = []
for model in all_models:
model_name = convert_time_name_to_model(model[0])
model = model[1]
if word in model.wv.key_to_index:
eligible_models.append(model_name)
return eligible_models
def main():
# model = load_word2vec_model('models/archaic_cbow.model')
# archaic_cbow_dict = model_dictionary(model)
# score = cosine_similarity(archaic_cbow_dict['Πελοπόννησος'], archaic_cbow_dict['σπάργανον'])
# print(score)
archaic = ('archaic', load_word2vec_model('models/archaic_cbow.model'))
classical = ('classical', load_word2vec_model('models/classical_cbow.model'))
early_roman = ('early_roman', load_word2vec_model('models/early_roman_cbow.model'))
hellen = ('hellen', load_word2vec_model('models/hellen_cbow.model'))
late_roman = ('late_roman', load_word2vec_model('models/late_roman_cbow.model'))
models = [archaic, classical, early_roman, hellen, late_roman]
nearest_neighbours = get_nearest_neighbours('πατήρ', 'archaic_cbow', n=5)
print(nearest_neighbours)
# vector = get_word_vector(model, 'ἀνήρ')
# print(vector)
# Iterate over all words and print their vectors
# iterate_over_words(model)
if __name__ == "__main__":
main()