看了很多关于word2vec的算法原理的介绍文章，看明白了，但依然有点不深刻。

以下是python直接实现的word2vec的算法，简单明了，读完就懂了

import numpy as npdef tokenize(text):return text.lower().split()def generate_word_pairs(sentences, window_size):word_pairs = []for sentence in sentences:for i, center_word in enumerate(sentence):for j in range(i - window_size, i + window_size + 1):if j >= 0 and j < len(sentence) and j != i:context_word = sentence[j]word_pairs.append((center_word, context_word))return word_pairsdef create_word_index(sentences):word_set = set(word for sentence in sentences for word in sentence)return {word: i for i, word in enumerate(word_set)}def one_hot_encoding(word, word_index):one_hot = np.zeros(len(word_index))one_hot[word_index[word]] = 1return one_hotdef train_word2vec(sentences, vector_size, window_size, learning_rate, epochs):word_index = create_word_index(sentences)W1 = np.random.rand(len(word_index), vector_size)W2 = np.random.rand(vector_size, len(word_index))word_pairs = generate_word_pairs(sentences, window_size)for epoch in range(epochs):loss = 0for center_word, context_word in word_pairs:center_word_encoded = one_hot_encoding(center_word, word_index)context_word_encoded = one_hot_encoding(context_word, word_index)hidden_layer = np.dot(center_word_encoded, W1)output_layer = np.dot(hidden_layer, W2)exp_output = np.exp(output_layer)softmax_output = exp_output / np.sum(exp_output)error = softmax_output - context_word_encodeddW2 = np.outer(hidden_layer, error)dW1 = np.outer(center_word_encoded, np.dot(W2, error))W1 -= learning_rate * dW1W2 -= learning_rate * dW2loss += -np.sum(output_layer * context_word_encoded) + np.log(np.sum(exp_output))print(f"Epoch: {epoch + 1}, Loss: {loss}")return W1, word_indexsentences = [tokenize("This is a sample sentence"),tokenize("Another example sentence"),tokenize("One more example")
]vector_size = 100
window_size = 2
learning_rate = 0.01
epochs = 100W1, word_index = train_word2vec(sentences, vector_size, window_size, learning_rate, epochs)for word, index in word_index.items():print(f"{word}: {W1[index]}")