Word2vec：skip-gram模型+Negative Sampling(负采样)代码实现

算法原理：

算法原理可以参考该链接

超参数

# Training Parameters
learning_rate = 0.1  
batch_size = 128	
num_steps = 3000000
display_step = 10000
eval_step = 200000
# Evaluation Parameters
valid_size = 20  
valid_window = 100 
#从词典的前100个词中随机选取20个词来验证模型
eval_words = np.random.choice(valid_window, valid_size, replace=False)
# Word2Vec Parameters
embedding_size = 200  
max_vocabulary_size = 50000  
min_occurrence = 10  # 词典中词出现的最低次数
skip_window = 3  # 窗口大小
num_skips = 2  # 每个输入中心词在其上下文区间中选取num_skips个词来生成样本
num_sampled = 64  # Number of negative examples

解释：

• 我们模型的验证是：计算于eval_words数组中的词最近似的几个词

词典生成模块

def make_vocabulary(data):
	"""
	data:是一个一维的list，每个元素可以是单个字也可以是切词后的词
	data是我们将句子切词后再拼接生成的（如果以字为单位不用切词直接拼接）
	"""
    word2count = [('UNK', -1)]
    #统计语言库词的次数
    word2count.extend(collections.Counter("".join(data)).most_common(max_vocabulary_size - 1))
	#去掉出现次数比较少的词
    for i in range(len(word2count) - 1, -1, -1):
        if word2count[i][1] < min_occurrence:
            word2count.pop(i)
        else:
            break

    vocabulary_size = len(word2count)
	
    word2id = dict()
    for i, (word, _) in enumerate(word2count):
        word2id[word] = i
	#将data中的词转化为其对应索引ID
    data_id = list()
    unk_count = 0
    for word in data:
        index = word2id.get(word, 0)
        if index == 0:
            unk_count += 1
        data_id.append(index)
    word2count[0] = ('UNK', unk_count)