Spaces:
Building
Building
import pandas as pd | |
import spacy | |
import numpy as np | |
from sklearn.cluster import DBSCAN | |
from sklearn.metrics.pairwise import cosine_distances | |
import matplotlib.pyplot as plt | |
import nltk | |
from nltk.corpus import wordnet | |
def load_data(file_path): | |
""" | |
This function loads the data from a given file_path | |
parameter: str the file path | |
Returns: the unique words in gloss column | |
""" | |
data = pd.read_csv(file_path, delimiter=";") | |
return data["gloss"].unique() | |
def initialize_spacy_model(model_name="en_core_web_md"): | |
return spacy.load(model_name) | |
def download_wordnet(): | |
""" | |
This function downloads a dictionary that will be used to find antonyms | |
""" | |
nltk.download('wordnet') | |
def generate_word_vectors(words, model): | |
return np.array([model(word).vector for word in words]) | |
def plot_k_distance_graph(distances, k): | |
k_distances = np.sort(distances, axis=1)[:, k] | |
k_distances = np.sort(k_distances) | |
plt.figure(figsize=(10, 5)) | |
plt.plot(k_distances) | |
plt.xlabel('Points sorted by distance') | |
plt.ylabel(f'{k}-th Nearest Neighbor Distance') | |
plt.title(f'k-distance Graph for k={k}') | |
plt.grid(True) | |
plt.show() | |
def perform_dbscan_clustering(word_vectors, eps, min_samples=5): | |
dbscan = DBSCAN(metric='cosine', eps=eps, min_samples=min_samples) | |
dbscan.fit(word_vectors) | |
return dbscan | |
def create_cluster_mapping(words, dbscan_labels): | |
cluster_to_words = {} | |
for word, cluster in zip(words, dbscan_labels): | |
if cluster not in cluster_to_words: | |
cluster_to_words[cluster] = [] | |
cluster_to_words[cluster].append(word) | |
return cluster_to_words | |
def find_antonyms(word): | |
antonyms = set() | |
for syn in wordnet.synsets(word): | |
for lemma in syn.lemmas(): | |
if lemma.antonyms(): | |
antonyms.add(lemma.antonyms()[0].name()) | |
return antonyms | |
def find_synonyms_in_cluster(word, model, cluster_to_words, dbscan_model): | |
""" | |
This function finds the most similar word in the same cluster, and excludes antonyms | |
""" | |
word_vector = model(word).vector | |
cluster_label = dbscan_model.fit_predict([word_vector])[0] | |
cluster_words = cluster_to_words.get(cluster_label, []) | |
if not cluster_words: | |
return None | |
antonyms = find_antonyms(word) | |
similarities = [(dict_word, model(dict_word).similarity(model(word))) for dict_word in cluster_words if dict_word != word and dict_word not in antonyms] | |
if not similarities: | |
return None | |
most_similar_word = sorted(similarities, key=lambda item: -item[1])[0][0] | |
return most_similar_word | |
def display_clusters(cluster_to_words): | |
for cluster_label, words in cluster_to_words.items(): | |
if cluster_label != -1: # Exclude noise points | |
print(f"Cluster {cluster_label}: {words}") | |
else: | |
print(f"Noise: {words}") | |
def main(file_path, model_name="en_core_web_md", eps=0.23, min_samples=5, k=5): | |
global nlp, cluster_to_words, dbscan | |
dict_2000 = load_data(file_path) | |
nlp = initialize_spacy_model(model_name) | |
download_wordnet() | |
word_vectors = generate_word_vectors(dict_2000, nlp) | |
# distances = cosine_distances(word_vectors) | |
# plot_k_distance_graph(distances, k) | |
dbscan = perform_dbscan_clustering(word_vectors, eps, min_samples) | |
cluster_to_words = create_cluster_mapping(dict_2000, dbscan.labels_) | |
if __name__ == "__main__": | |
main("filtered_WLASL.csv") | |
##TEST## | |
#target_word = "unhappy" | |
#synonym = find_synonyms_in_cluster(target_word, nlp, cluster_to_words, dbscan) | |
#print(f"The most similar word to '{target_word}' is '{synonym}'") | |
##If you want to see clusters## | |
#num_clusters = len(set(dbscan.labels_)) - (1 if -1 in dbscan.labels_ else 0) | |
#print(f"Number of clusters: {num_clusters}") | |
#cluster_label = dbscan.fit_predict([nlp("unhappy").vector])[0] | |
#same_cluster_words = cluster_to_words.get(cluster_label, []) | |
#print(f"Words in the same cluster as 'unhappy': {same_cluster_words}") | |
#display_clusters(cluster_to_words) | |