PyTorch入门-词向量

一. 概念
1 . 词的表示
在自然语言处理任务中，首先需要考虑词如何在计算机中表示。通常，有两种表示方式：one-hot representation和distribution representation。
1.1 离散表示（one-hot representation）
传统的基于规则或基于统计的自然语义处理方法将单词看作一个原子符号被称作one-hot representation。one-hot representation把每个词表示为一个长向量。这个向量的维度是词表大小，向量中只有一个维度的值为1，其余维度为0，这个维度就代表了当前的词。例如：苹果 [0，0，0，1，0，0，0，0，0，……]
one-hot representation相当于给每个词分配一个id，这就导致这种表示方式不能展示词与词之间的关系。另外，one-hot representation将会导致特征空间非常大，但也带来一个好处，就是在高维空间中，很多应用任务线性可分。
1.2 分布式表示（distribution representation）
word embedding指的是将词转化成一种分布式表示，又称词向量。分布式表示将词表示成一个定长的连续的稠密向量。
分布式表示优点:

1. 词之间存在相似关系：是词之间存在“距离”概念，这对很多自然语言处理的任务非常有帮助。
2. 包含更多信息：词向量能够包含更多信息，并且每一维都有特定的含义。在采用one-hot特征时，可以对特征向量进行删减，词向量则不能。

二. 常用的词向量技术操作
1. 用Skip-thought模型训练词向量

import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.utils.data as tud
from torch.nn.parameter import Parameter
from collections import Counter
import numpy as np
import random
import math
import pandas as pd
import scipy
import sklearn
from sklearn.metrics.pairwise import cosine_similarity
USE_CUDA = torch.cuda.is_available()
# 为了保证实验结果可以复现，我们经常会把各种random seed固定在某一个值
random.seed(53113)
np.random.seed(53113)
torch.manual_seed(53113)
if USE_CUDA:
    torch.cuda.manual_seed(53113) 
# 设定一些超参数  
K = 100 # number of negative samples
C = 3 # nearby words threshold
NUM_EPOCHS = 2 # The number of epochs of training
MAX_VOCAB_SIZE = 30000 # the vocabulary size
BATCH_SIZE = 128 # the batch size
LEARNING_RATE = 0.2 # the initial learning rate
EMBEDDING_SIZE = 100 
LOG_FILE = "word-embedding.log"
# tokenize函数，把一篇文本转化成一个个单词
def word_tokenize(text):
    return text.split()
with open("text8.train.txt", "r") as fin:
    text = fin.read()
text = [w for w in word_tokenize(text.lower())]
vocab = dict(Counter(text).most_common(MAX_VOCAB_SIZE-1))
vocab["<unk>"] = len(text) - np.sum(list(vocab.values()))
idx_to_word = [word for word in vocab.keys()] 
word_to_idx = {word:i for i, word in enumerate(idx_to_word)}
word_counts = np.array([count for count in vocab.values()], dtype=np.float32)
word_freqs = word_counts / np.sum(word_counts)
word_freqs = word_freqs ** (3./4.)
word_freqs = word_freqs / np.sum(word_freqs) # 用来做 negative sampling
VOCAB_SIZE = len(idx_to_word)
VOCAB_SIZE

结果：

30000

2. 实现Dataloader
一个dataloader需要以下内容：

• 把所有text编码成数字，然后用subsampling预处理这些文字。
• 保存vocabulary，单词count，normalized word frequency
• 每个iteration sample一个中心词
• 根据当前的中心词返回context单词
• 根据中心词sample一些negative单词
• 返回单词的counts

为了使用dataloader，需要定义以下两个function:

• len function需要返回整个数据集中有多少个item
• get 根据给定的index返回一个item

有了dataloader之后，可以轻松随机打乱整个数据集，拿到一个batch的数据等等。

class WordEmbeddingDataset(tud.Dataset):
    def __init__(self, text, word_to_idx, idx_to_word, word_freqs, word_counts):
        ''' text: a list of words, all text from the training dataset
            word_to_idx: the dictionary from word to idx
            idx_to_word: idx to word mapping
            word_freq: the frequency of each word
            word_counts: the word counts
        '''
        super(WordEmbeddingDataset, self).__init__()
        self.text_encoded = [word_to_idx.get(t, VOCAB_SIZE-1) for t in text]
        self.text_encoded = torch.Tensor(self.text_encoded).long()
        self.word_to_idx = word_to_idx
        self.idx_to_word = idx_to_word
        self.word_freqs = torch.Tensor(word_freqs)
        self.word_counts = torch.Tensor(word_counts)
        
    def __len__(self):
        ''' 返回整个数据集（所有单词）的长度
        '''
        return len(self.text_encoded)
    def __getitem__(self, idx):
        ''' 这个function返回以下数据用于训练
            - 中心词
            - 这个单词附近的(positive)单词
            - 随机采样的K个单词作为negative sample
        '''
        center_word = self.text_encoded[idx]
        pos_indices = list(range(idx-C, idx)) + list(range(idx+1, idx+C+1))
        pos_indices = [i%len(self.text_encoded) for i in pos_indices]
        pos_words = self.text_encoded[pos_indices] 
        neg_words = torch.multinomial(self.word_freqs, K * pos_words.shape[0], True)
        return center_word, pos_words, neg_words 

创建dataset和dataloader

dataset = WordEmbeddingDataset(text, word_to_idx, idx_to_word, word_freqs, word_counts)
dataloader = tud.DataLoader(dataset, batch_size=BATCH_SIZE, shuffle=True, num_workers=4)    

3. 定义PyTorch模型

class EmbeddingModel(nn.Module):
    def __init__(self, vocab_size, embed_size):
        ''' 初始化输出和输出embedding
        '''
        super(EmbeddingModel, self).__init__()
        self.vocab_size = vocab_size
        self.embed_size = embed_size
        initrange = 0.5 / self.embed_size
        self.out_embed = nn.Embedding(self.vocab_size, self.embed_size, sparse=False)
        self.out_embed.weight.data.uniform_(-initrange, initrange)
        self.in_embed = nn.Embedding(self.vocab_size, self.embed_size, sparse=False)
        self.in_embed.weight.data.uniform_(-initrange, initrange)
    def forward(self, input_labels, pos_labels, neg_labels):
        '''
        input_labels: 中心词, [batch_size]
        pos_labels: 中心词周围 context window 出现过的单词 [batch_size * (window_size * 2)]
        neg_labelss: 中心词周围没有出现过的单词，从 negative sampling 得到 [batch_size, (window_size * 2 * K)]
        
        return: loss, [batch_size]
        '''
        batch_size = input_labels.size(0)
        input_embedding = self.in_embed(input_labels) # B * embed_size
        pos_embedding = self.out_embed(pos_labels) # B * (2*C) * embed_size
        neg_embedding = self.out_embed(neg_labels) # B * (2*C * K) * embed_size
        log_pos = torch.bmm(pos_embedding, input_embedding.unsqueeze(2)).squeeze() # B * (2*C)
        log_neg = torch.bmm(neg_embedding, -input_embedding.unsqueeze(2)).squeeze() # B * (2*C*K)
        log_pos = F.logsigmoid(log_pos).sum(1)
        log_neg = F.logsigmoid(log_neg).sum(1) # batch_size
        loss = log_pos + log_neg
        return -loss
    def input_embeddings(self):
        return self.in_embed.weight.data.cpu().numpy()

4. 训练模型

• 模型一般需要训练若干个epoch
• 每个epoch我们都把所有的数据分成若干个batch
• 把每个batch的输入和输出都包装成cuda tensor
• forward pass，通过输入的句子预测每个单词的下一个单词
• 用模型的预测和正确的下一个单词计算cross entropy loss
• 清空模型当前gradient
• backward pass
• 更新模型参数
• 每隔一定的iteration输出模型在当前iteration的loss，以及在验证数据集上做模型的评估

def evaluate(filename, embedding_weights): 
    if filename.endswith(".csv"):
        data = pd.read_csv(filename, sep=",")
    else:
        data = pd.read_csv(filename, sep="\t")
    human_similarity = []
    model_similarity = []
    for i in data.iloc[:, 0:2].index:
        word1, word2 = data.iloc[i, 0], data.iloc[i, 1]
        if word1 not in word_to_idx or word2 not in word_to_idx:
            continue
        else:
            word1_idx, word2_idx = word_to_idx[word1], word_to_idx[word2]
            word1_embed, word2_embed = embedding_weights[[word1_idx]], embedding_weights[[word2_idx]]
            model_similarity.append(float(sklearn.metrics.pairwise.cosine_similarity(word1_embed, word2_embed)))
            human_similarity.append(float(data.iloc[i, 2]))

    return scipy.stats.spearmanr(human_similarity, model_similarity)# , model_similarity

def find_nearest(word):
    index = word_to_idx[word]
    embedding = embedding_weights[index]
    cos_dis = np.array([scipy.spatial.distance.cosine(e, embedding) for e in embedding_weights])
    return [idx_to_word[i] for i in cos_dis.argsort()[:10]]
optimizer = torch.optim.SGD(model.parameters(), lr=LEARNING_RATE)
for e in range(NUM_EPOCHS):
    for i, (input_labels, pos_labels, neg_labels) in enumerate(dataloader):   
        # TODO
        input_labels = input_labels.long()
        pos_labels = pos_labels.long()
        neg_labels = neg_labels.long()
        if USE_CUDA:
            input_labels = input_labels.cuda()
            pos_labels = pos_labels.cuda()
            neg_labels = neg_labels.cuda()    
        optimizer.zero_grad()
        loss = model(input_labels, pos_labels, neg_labels).mean()
        loss.backward()
        optimizer.step()
        if i % 100 == 0:
            with open(LOG_FILE, "a") as fout:
                fout.write("epoch: {}, iter: {}, loss: {}\n".format(e, i, loss.item()))
                print("epoch: {}, iter: {}, loss: {}".format(e, i, loss.item()))
        if i % 2000 == 0:
            embedding_weights = model.input_embeddings()
            sim_simlex = evaluate("simlex-999.txt", embedding_weights)
            sim_men = evaluate("men.txt", embedding_weights)
            sim_353 = evaluate("wordsim353.csv", embedding_weights)
            with open(LOG_FILE, "a") as fout:
                print("epoch: {}, iteration: {}, simlex-999: {}, men: {}, sim353: {}, nearest to monster: {}\n".format(
                    e, i, sim_simlex, sim_men, sim_353, find_nearest("monster")))
                fout.write("epoch: {}, iteration: {}, simlex-999: {}, men: {}, sim353: {}, nearest to monster: {}\n".format(
                    e, i, sim_simlex, sim_men, sim_353, find_nearest("monster")))
                
    embedding_weights = model.input_embeddings()
    np.save("embedding-{}".format(EMBEDDING_SIZE), embedding_weights)
    torch.save(model.state_dict(), "embedding-{}.th".format(EMBEDDING_SIZE))

5. 在 MEN 和 Simplex-999 数据集上做评估

embedding_weights = model.input_embeddings()
print("simlex-999", evaluate("simlex-999.txt", embedding_weights))
print("men", evaluate("men.txt", embedding_weights))
print("wordsim353", evaluate("wordsim353.csv", embedding_weights))

结果:

simlex-999 SpearmanrResult(correlation=0.17251697429101504, pvalue=7.863946056740345e-08)
men SpearmanrResult(correlation=0.1778096817088841, pvalue=7.565661657312768e-20)
wordsim353 SpearmanrResult(correlation=0.27153702278146635, pvalue=8.842165885381714e-07)

6. 寻找nearest neighbors

for word in ["good", "fresh", "monster", "green", "like", "america", "chicago", "work", "computer", "language"]:
    print(word, find_nearest(word))

结果：

good ['good', 'bad', 'perfect', 'hard', 'questions', 'alone', 'money', 'false', 'truth', 'experience']
fresh ['fresh', 'grain', 'waste', 'cooling', 'lighter', 'dense', 'mild', 'sized', 'warm', 'steel']
monster ['monster', 'giant', 'robot', 'hammer', 'clown', 'bull', 'demon', 'triangle', 'storyline', 'slogan']
green ['green', 'blue', 'yellow', 'white', 'cross', 'orange', 'black', 'red', 'mountain', 'gold']
like ['like', 'unlike', 'etc', 'whereas', 'animals', 'soft', 'amongst', 'similarly', 'bear', 'drink']
america ['america', 'africa', 'korea', 'india', 'australia', 'turkey', 'pakistan', 'mexico', 'argentina', 'carolina']
chicago ['chicago', 'boston', 'illinois', 'texas', 'london', 'indiana', 'massachusetts', 'florida', 'berkeley', 'michigan']
work ['work', 'writing', 'job', 'marx', 'solo', 'label', 'recording', 'nietzsche', 'appearance', 'stage']
computer ['computer', 'digital', 'electronic', 'audio', 'video', 'graphics', 'hardware', 'software', 'computers', 'program']
language ['language', 'languages', 'alphabet', 'arabic', 'grammar', 'pronunciation', 'dialect', 'programming', 'chinese', 'spelling']

7. 单词之间的关系

man_idx = word_to_idx["man"] 
king_idx = word_to_idx["king"] 
woman_idx = word_to_idx["woman"]
embedding = embedding_weights[woman_idx] - embedding_weights[man_idx] + embedding_weights[king_idx]
cos_dis = np.array([scipy.spatial.distance.cosine(e, embedding) for e in embedding_weights])
for i in cos_dis.argsort()[:20]:
    print(idx_to_word[i])

结果：

king
henry
charles
pope
queen
iii
prince
elizabeth
alexander
constantine
edward
son
iv
louis
emperor
mary
james
joseph
frederick
francis