自然语言处理 专项1 TF-IDF、word2vec、CBOW和Skip-Gram

自然语言处理 专项1 TF-IDF、word2vec、CBOW和Skip-Gram

借鉴：https://blog.csdn.net/weixin_40699243/article/details/109202976

TF-IDF

TF-IDF的主要思想是：如果某个词或短语在一篇文章中出现的频率TF高，并且在其他文章中很少出现，则认为此词或者短语具有很好的类别区分能力，适合用来分类。TF-IDF实际上是：TF×IDF，TF为词频，IDF反文档频率。
（倾向于过滤掉常见的词语，保留重要的词语）
词频（TF） = 某个词在文章中的出现次数 / 文章总词数
逆文档频率（IDF） = log（词料库的文档总数/包含该词的文档数+1）
TF-IDF = 词频（TF) ×逆文档频率（IDF）

import numpy as np
from collections import Counter
import itertools #一系列用来产生不同类型迭代器的函数或类，这些函数的返回都是一个迭代器，我们可以通过 for 循环来遍历取值，也可以使用 next() 来取值
from visual import show_tfidf

docs = [
    "it is a good day, I like to stay here",
    "I am happy to be here",
    "I am bob",
    "it is sunny today",
    "I have a party today",
    "it is a dog and that is a cat",
    "there are dog and cat on the tree",
    "I study hard this morning",
    "today is a good day",
    "tomorrow will be a good day",
    "I like coffee, I like book and I like apple",
    "I do not like it",
    "I am kitty, I like bob",
    "I do not care who like bob, but I like kitty",
    "It is coffee time, bring your cup",
]

docs_words = [d.replace(",", "").split(" ") for d in docs] #得到一个列表
vocab = set(itertools.chain(*docs_words)) #得到一个集合，用chain创建一个迭代器，依次连续的返回每个可迭代对象的元素
v2i = {v: i for i, v in enumerate(vocab)} #将一个可遍历的数据对象(如列表、元组或字符串)组合为一个索引序列，同时列出数据和数据下标
i2v = {i: v for v, i in v2i.items()} #字典形式，键和值；items代表键和值都遍历


def safe_log(x):
    mask = x != 0
    x[mask] = np.log(x[mask])
    return x


tf_methods = {
        "log": lambda x: np.log(1+x),
        "augmented": lambda x: 0.5 + 0.5 * x / np.max(x, axis=1, keepdims=True),
        "boolean": lambda x: np.minimum(x, 1),
        "log_avg": lambda x: (1 + safe_log(x)) / (1 + safe_log(np.mean(x, axis=1, keepdims=True))),
    }
idf_methods = {
        "log": lambda x: 1 + np.log(len(docs) / (x+1)),
        "prob": lambda x: np.maximum(0, np.log((len(docs) - x) / (x+1))),
        "len_norm": lambda x: x / (np.sum(np.square(x))+1),
    }


def get_tf(method="log"):
    # term frequency: how frequent a word appears in a doc
    _tf = np.zeros((len(vocab), len(docs)), dtype=np.float64)    # [n_vocab, n_doc]
    for i, d in enumerate(docs_words):
        counter = Counter(d)
        for v in counter.keys():
            _tf[v2i[v], i] = counter[v] / counter.most_common(1)[0][1] #counter.most_common(1)找到出现频率最高的词

    weighted_tf = tf_methods.get(method, None)
    if weighted_tf is None:
        raise ValueError
    return weighted_tf(_tf)


def get_idf(method="log"):
    # inverse document frequency: low idf for a word appears in more docs, mean less important
    df = np.zeros((len(i2v), 1))
    for i in range(len(i2v)):
        d_count = 0
        for d in docs_words:
            d_count += 1 if i2v[i] in d else 0
        df[i, 0] = d_count

    idf_fn = idf_methods.get(method, None)
    if idf_fn is None:
        raise ValueError
    return idf_fn(df)


def cosine_similarity(q, _tf_idf): #余弦距离
    unit_q = q / np.sqrt(np.sum(np.square(q), axis=0, keepdims=True))
    unit_ds = _tf_idf / np.sqrt(np.sum(np.square(_tf_idf), axis=0, keepdims=True))
    similarity = unit_ds.T.dot(unit_q).ravel()
    return similarity

#误差得分？？？
def docs_score(q, len_norm=False):
    q_words = q.replace(",", "").split(" ")

    # add unknown words
    unknown_v = 0
    for v in set(q_words):
        if v not in v2i:
            v2i[v] = len(v2i)
            i2v[len(v2i)-1] = v
            unknown_v += 1
    if unknown_v > 0:
        _idf = np.concatenate((idf, np.zeros((unknown_v, 1), dtype=np.float)), axis=0)
        _tf_idf = np.concatenate((tf_idf, np.zeros((unknown_v, tf_idf.shape[1]), dtype=np.float)), axis=0)
    else:
        _idf, _tf_idf = idf, tf_idf
    counter = Counter(q_words)
    q_tf = np.zeros((len(_idf), 1), dtype=np.float)     # [n_vocab, 1]
    for v in counter.keys():
        q_tf[v2i[v], 0] = counter[v]

    q_vec = q_tf * _idf            # [n_vocab, 1]

    q_scores = cosine_similarity(q_vec, _tf_idf)
    if len_norm:
        len_docs = [len(d) for d in docs_words]
        q_scores = q_scores / np.array(len_docs)
    return q_scores


def get_keywords(n=2):
    for c in range(3):
        col = tf_idf[:, c]
        idx = np.argsort(col)[-n:]
        print("doc{}, top{} keywords {}".format(c, n, [i2v[i] for i in idx]))


tf = get_tf()           # [n_vocab, n_doc]
idf = get_idf()         # [n_vocab, 1]
tf_idf = tf * idf       # [n_vocab, n_doc]
print("tf shape(vecb in each docs): ", tf.shape)
print("\ntf samples:\n", tf[:2])
print("\nidf shape(vecb in all docs): ", idf.shape)
print("\nidf samples:\n", idf[:2])
print("\ntf_idf shape: ", tf_idf.shape)
print("\ntf_idf sample:\n", tf_idf[:2])


# test
get_keywords()
q = "I get a coffee cup"
scores = docs_score(q)
d_ids = scores.argsort()[-3:][::-1]
print("\ntop 3 docs for '{}':\n{}".format(q, [docs[i] for i in d_ids]))

show_tfidf(tf_idf.T, [i2v[i] for i in range(len(i2v))], "tfidf_matrix")

Word2vec

Word2vec 是 Word Embedding 方式之一，属于 NLP 领域。他是将词转化为「可计算」「结构化」的向量的过程。

运用神经网络将V维的词向量，通过权重矩阵Vd转换为 d维词向量。
输入为word bag中V个词中的一个词，已经转换为V维 one-hot编码， 通过权重矩阵W, 得到输出1d维的H, 再经过softmax d*V维转换成为V维的词向量。
输出含义：条件概率
还可以用训练好的H权重，将V维的one-hot词向量转化为d维的词向量。

CBOW和Skip-Gram 比较优劣

在cbow方法中，是用周围词预测中心词，从而利用中心词的预测结果情况，使用GradientDesent方法，不断的去调整周围词的向量。
当训练完成之后，每个词都会作为中心词，把周围词的词向量进行了调整，这样也就获得了整个文本里面所有词的词向量。
要注意的是， cbow的对周围词的调整是统一的：求出的gradient的值会同样的作用到每个周围词的词向量当中去。
可以看到，cbow预测行为的次数跟整个文本的词数几乎是相等的
（每次预测行为才会进行一次backpropgation, 而往往这也是最耗时的部分），复杂度大概是O(V)

skip-gram是用中心词来预测周围的词，利用周围的词的预测结果情况使用GradientDecent来不断的调整中心词的词向量
最终所有的文本遍历完毕之后，也就得到了文本所有词的词向量。
可以看出，skip-gram进行预测的次数是要多于cbow的：因为每个词在作为中心词时，都要使用周围词进行预测一次。
这样相当于比cbow的方法多进行了K次（假设K为窗口大小），因此时间的复杂度为O(KV)，训练时间要比cbow要长。

但是在skip-gram当中，每个词都要收到周围的词的影响，每个词在作为中心词的时候，都要进行K次的预测、调整。
因此， 当数据量较少，或者词为生僻词出现次数较少时， 这种多次的调整会使得词向量相对的更加准确。

因为尽管cbow从另外一个角度来说，某个词也是会受到多次周围词的影响（多次将其包含在内的窗口移动），进行词向量的跳帧
但是他的调整是跟周围的词一起调整的，grad的值会平均分到该词上， 相当于该生僻词没有收到专门的训练，它只是沾了周围词的光而已。

skip-gram：1个中心词受K个周围词的影响，中心词的“能力”（向量结果）相对就会扎实（准确）一些，但是这样肯定会使用更长的时间；

cbow：1个中心词影响K个周围词，但单轮训练不保证结果，还要看下一轮（假如还在窗口内），或者以后的某一轮再次出现作为周围词，然后进步一点，效率高，速度更快，但可能结果不那么准确

代码分析：CBOW

from tensorflow import keras
import tensorflow as tf
from utils import process_w2v_data
from visual import show_w2v_word_embedding
tf.enable_eager_execution()
corpus = [
    # numbers
    "5 2 4 8 6 2 3 6 4",
    "4 8 5 6 9 5 5 6",
    "1 1 5 2 3 3 8",
    "3 6 9 6 8 7 4 6 3",
    "8 9 9 6 1 4 3 4",
    "1 0 2 0 2 1 3 3 3 3 3",
    "9 3 3 0 1 4 7 8",
    "9 9 8 5 6 7 1 2 3 0 1 0",

    # alphabets, expecting that 9 is close to letters
    "a t g q e h 9 u f",
    "e q y u o i p s",
    "q o 9 p l k j o k k o p",
    "h g y i u t t a e q",
    "i k d q r e 9 e a d",
    "o p d g 9 s a f g a",
    "i u y g h k l a s w",
    "o l u y a o g f s",
    "o p i u y g d a s j d l",
    "u k i l o 9 l j s",
    "y g i s h k j l f r f",
    "i o h n 9 9 d 9 f a 9",
]


class CBOW(keras.Model):
    def __init__(self, v_dim, emb_dim):
        super().__init__()
        self.v_dim = v_dim
        self.embeddings = keras.layers.Embedding(
            input_dim=v_dim, output_dim=emb_dim,  # [n_vocab, emb_dim]
            embeddings_initializer=keras.initializers.RandomNormal(0., 0.1),
        )

        # noise-contrastive estimation
        self.nce_w = self.add_weight(
            name="nce_w", shape=[v_dim, emb_dim],
            initializer=keras.initializers.TruncatedNormal(0., 1.))  # [n_vocab, emb_dim]
        self.nce_b = self.add_weight(
            name="nce_b", shape=(v_dim,),
            initializer=keras.initializers.Constant(0.1))  # [n_vocab, ]

        self.opt = keras.optimizers.Adam(0.01)

    def call(self, x, training=None, mask=None):
        # x.shape = [n, skip_window*2]
        o = self.embeddings(x)          # [n, skip_window*2, emb_dim]
        o = tf.reduce_mean(o, axis=1)   # [n, emb_dim]
        return o

    # negative sampling: take one positive label and num_sampled negative labels to compute the loss
    # in order to reduce the computation of full softmax

    def loss(self, x, y, training=None):
        embedded = self.call(x, training)
        return tf.reduce_mean(
            tf.nn.nce_loss(
                weights=self.nce_w, biases=self.nce_b, labels=tf.expand_dims(y, axis=1),
                inputs=embedded, num_sampled=5, num_classes=self.v_dim))

    def step(self, x, y):
        with tf.GradientTape() as tape:
            loss = self.loss(x, y, True)
            grads = tape.gradient(loss, self.trainable_variables)
        self.opt.apply_gradients(zip(grads, self.trainable_variables))
        return loss.numpy()


def train(model, data):
    for t in range(2500):
        bx, by = data.sample(8)
        loss = model.step(bx, by)
        if t % 200 == 0:
            print("step: {} | loss: {}".format(t, loss))


if __name__ == "__main__":
    d = process_w2v_data(corpus, skip_window=2, method="cbow")
    m = CBOW(d.num_word, 2)
    train(m, d)
    # plotting
    show_w2v_word_embedding(m, d, "./visual/results/cbow.png")

代码分析：skip-gram

import tensorflow as tf
from tensorflow import keras
from utils import process_w2v_data
from visual import show_w2v_word_embedding
tf.enable_eager_execution()

#输入语料库，包含数字和字母，很明显9混入了字母区，那么在向量图中会有所体现
corpus = [
    # numbers
    "5 2 4 8 6 2 3 6 4",
    "4 8 5 6 9 5 5 6",
    "1 1 5 2 3 3 8",
    "3 6 9 6 8 7 4 6 3",
    "8 9 9 6 1 4 3 4",
    "1 0 2 0 2 1 3 3 3 3 3",
    "9 3 3 0 1 4 7 8",
    "9 9 8 5 6 7 1 2 3 0 1 0",

    # alphabets, expecting that 9 is close to letters
    "a t g q e h 9 u f",
    "e q y u o i p s",
    "q o 9 p l k j o k k o p",
    "h g y i u t t a e q",
    "i k d q r e 9 e a d",
    "o p d g 9 s a f g a",
    "i u y g h k l a s w",
    "o l u y a o g f s",
    "o p i u y g d a s j d l",
    "u k i l o 9 l j s",
    "y g i s h k j l f r f",
    "i o h n 9 9 d 9 f a 9",
]

class SkipGram(keras.Model):
    def __init__(self, v_dim, emb_dim):
        super().__init__()
        self.v_dim = v_dim
        self.embeddings = keras.layers.Embedding(
            input_dim=v_dim, output_dim=emb_dim,       # [n_vocab, emb_dim]
            embeddings_initializer=keras.initializers.RandomNormal(0., 0.1),
        )

        # noise-contrastive estimation
        self.nce_w = self.add_weight(
            name="nce_w", shape=[v_dim, emb_dim],
            initializer=keras.initializers.TruncatedNormal(0., 1.))  # [n_vocab, emb_dim]
        self.nce_b = self.add_weight(
            name="nce_b", shape=(v_dim,),
            initializer=keras.initializers.Constant(0.1))  # [n_vocab, ]

        self.opt = keras.optimizers.Adam(0.01)

    def call(self, x, training=None, mask=None):
        # x.shape = [n, ]
        o = self.embeddings(x)      # [n, emb_dim]
        return o

    # negative sampling: take one positive label and num_sampled negative labels to compute the loss
    # in order to reduce the computation of full softmax
    def loss(self, x, y, training=None):
        embedded = self.call(x, training)
        return tf.reduce_mean(
            tf.nn.nce_loss(
                weights=self.nce_w, biases=self.nce_b, labels=tf.expand_dims(y, axis=1),
                inputs=embedded, num_sampled=5, num_classes=self.v_dim))

    def step(self, x, y):
        with tf.GradientTape() as tape:
            loss = self.loss(x, y, True)
            grads = tape.gradient(loss, self.trainable_variables)
        self.opt.apply_gradients(zip(grads, self.trainable_variables))
        return loss.numpy()


def train(model, data):
    for t in range(2500):
        bx, by = data.sample(8)
        loss = model.step(bx, by)
        if t % 200 == 0:
            print("step: {} | loss: {}".format(t, loss))


if __name__ == "__main__":
    d = process_w2v_data(corpus, skip_window=2, method="skip_gram")
    m = SkipGram(d.num_word, 2)
    train(m, d)

    # plotting
    show_w2v_word_embedding(m, d, "./visual/results/skipgram.png")