溫馨提示×

您好,登錄后才能下訂單哦!

密碼登錄×
登錄注冊(cè)×
其他方式登錄
點(diǎn)擊 登錄注冊(cè) 即表示同意《億速云用戶服務(wù)條款》

nlp計(jì)數(shù)法應(yīng)用于PTB數(shù)據(jù)集的方法

發(fā)布時(shí)間:2022-04-06 10:15:40 來源:億速云 閱讀:146 作者:iii 欄目:開發(fā)技術(shù)

今天小編給大家分享一下nlp計(jì)數(shù)法應(yīng)用于PTB數(shù)據(jù)集的方法的相關(guān)知識(shí)點(diǎn),內(nèi)容詳細(xì),邏輯清晰,相信大部分人都還太了解這方面的知識(shí),所以分享這篇文章給大家參考一下,希望大家閱讀完這篇文章后有所收獲,下面我們一起來了解一下吧。

PTB數(shù)據(jù)集

內(nèi)容如下:

一行保存一個(gè)句子;將稀有單詞替換成特殊字符 < unk > ;將具體的數(shù)字替換 成“N”

 we 're talking about years ago before anyone heard of asbestos having any questionable properties 
 there is no asbestos in our products now 
 neither <unk> nor the researchers who studied the workers were aware of any research on smokers of the kent cigarettes 
 we have no useful information on whether users are at risk said james a. <unk> of boston 's <unk> cancer institute 
 dr. <unk> led a team of researchers from the national cancer institute and the medical schools of harvard university and boston university

ptb.py

使用PTB數(shù)據(jù)集:

由下面這句話,可知用PTB數(shù)據(jù)集時(shí)候,是把所有句子首尾連接了。

words = open(file_path).read().replace('\n', '<eos>').strip().split()

ptb.py起到了下載PTB數(shù)據(jù)集,把數(shù)據(jù)集存到文件夾某個(gè)位置,然后對(duì)數(shù)據(jù)集進(jìn)行提取的功能,提取出corpus, word_to_id, id_to_word。

import sys
import os
sys.path.append('..')
try:
    import urllib.request
except ImportError:
    raise ImportError('Use Python3!')
import pickle
import numpy as np
url_base = 'https://raw.githubusercontent.com/tomsercu/lstm/master/data/'
key_file = {
    'train':'ptb.train.txt',
    'test':'ptb.test.txt',
    'valid':'ptb.valid.txt'
}
save_file = {
    'train':'ptb.train.npy',
    'test':'ptb.test.npy',
    'valid':'ptb.valid.npy'
}
vocab_file = 'ptb.vocab.pkl'
dataset_dir = os.path.dirname(os.path.abspath(__file__))
def _download(file_name):
    file_path = dataset_dir + '/' + file_name
    if os.path.exists(file_path):
        return
    print('Downloading ' + file_name + ' ... ')
    try:
        urllib.request.urlretrieve(url_base + file_name, file_path)
    except urllib.error.URLError:
        import ssl
        ssl._create_default_https_context = ssl._create_unverified_context
        urllib.request.urlretrieve(url_base + file_name, file_path)
    print('Done')
def load_vocab():
    vocab_path = dataset_dir + '/' + vocab_file
    if os.path.exists(vocab_path):
        with open(vocab_path, 'rb') as f:
            word_to_id, id_to_word = pickle.load(f)
        return word_to_id, id_to_word
    word_to_id = {}
    id_to_word = {}
    data_type = 'train'
    file_name = key_file[data_type]
    file_path = dataset_dir + '/' + file_name
    _download(file_name)
    words = open(file_path).read().replace('\n', '<eos>').strip().split()
    for i, word in enumerate(words):
        if word not in word_to_id:
            tmp_id = len(word_to_id)
            word_to_id[word] = tmp_id
            id_to_word[tmp_id] = word
    with open(vocab_path, 'wb') as f:
        pickle.dump((word_to_id, id_to_word), f)
    return word_to_id, id_to_word
def load_data(data_type='train'):
    '''
        :param data_type: 數(shù)據(jù)的種類:'train' or 'test' or 'valid (val)'
        :return:
    '''
    if data_type == 'val': data_type = 'valid'
    save_path = dataset_dir + '/' + save_file[data_type]
    word_to_id, id_to_word = load_vocab()
    if os.path.exists(save_path):
        corpus = np.load(save_path)
        return corpus, word_to_id, id_to_word
    file_name = key_file[data_type]
    file_path = dataset_dir + '/' + file_name
    _download(file_name)
    words = open(file_path).read().replace('\n', '<eos>').strip().split()
    corpus = np.array([word_to_id[w] for w in words])
    np.save(save_path, corpus)
    return corpus, word_to_id, id_to_word
if __name__ == '__main__':
    for data_type in ('train', 'val', 'test'):
        load_data(data_type)

使用ptb.py

corpus保存了單詞ID列表,id_to_word 是將單詞ID轉(zhuǎn)化為單詞的字典,word_to_id 是將單詞轉(zhuǎn)化為單詞ID的字典。

使用ptb.load_data()加載數(shù)據(jù)。里面的參數(shù) &lsquo;train&rsquo;、&lsquo;test&rsquo;、&lsquo;valid&rsquo; 分別對(duì)應(yīng)訓(xùn)練用數(shù)據(jù)、測(cè)試用數(shù)據(jù)、驗(yàn)證用數(shù)據(jù)。

import sys
sys.path.append('..')
from dataset import ptb
corpus, word_to_id, id_to_word = ptb.load_data('train')
print('corpus size:', len(corpus))
print('corpus[:30]:', corpus[:30])
print()
print('id_to_word[0]:', id_to_word[0])
print('id_to_word[1]:', id_to_word[1])
print('id_to_word[2]:', id_to_word[2])
print()
print("word_to_id['car']:", word_to_id['car'])
print("word_to_id['happy']:", word_to_id['happy'])
print("word_to_id['lexus']:", word_to_id['lexus'])

結(jié)果:

corpus size: 929589
corpus[:30]: [ 0  1  2  3  4  5  6  7  8  9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
 24 25 26 27 28 29]
id_to_word[0]: aer
id_to_word[1]: banknote
id_to_word[2]: berlitz
word_to_id['car']: 3856
word_to_id['happy']: 4428
word_to_id['lexus']: 7426
Process finished with exit code 0

計(jì)數(shù)方法應(yīng)用于PTB數(shù)據(jù)集

其實(shí)和不用PTB數(shù)據(jù)集的區(qū)別就在于這句話。

corpus, word_to_id, id_to_word = ptb.load_data('train')

下面這句話起降維的效果

word_vecs = U[:, :wordvec_size]

整個(gè)代碼其實(shí)耗時(shí)最大的是在下面這個(gè)函數(shù)上:

W = ppmi(C, verbose=True)

完整代碼:

import sys
sys.path.append('..')
import numpy as np
from common.util import most_similar, create_co_matrix, ppmi
from dataset import ptb
window_size = 2
wordvec_size = 100
corpus, word_to_id, id_to_word = ptb.load_data('train')
vocab_size = len(word_to_id)
print('counting  co-occurrence ...')
C = create_co_matrix(corpus, vocab_size, window_size)
print('calculating PPMI ...')
W = ppmi(C, verbose=True)
print('calculating SVD ...')
#try:
    # truncated SVD (fast!)
print("ok")
from sklearn.utils.extmath import randomized_svd
U, S, V = randomized_svd(W, n_components=wordvec_size, n_iter=5,
                             random_state=None)
#except ImportError:
    # SVD (slow)
    # U, S, V = np.linalg.svd(W)
word_vecs = U[:, :wordvec_size]
querys = ['you', 'year', 'car', 'toyota']
for query in querys:
    most_similar(query, word_to_id, id_to_word, word_vecs, top=5)

下面這個(gè)是用普通的np.linalg.svd(W)做出的結(jié)果。

[query] you
 i: 0.7016294002532959
 we: 0.6388039588928223
 anybody: 0.5868048667907715
 do: 0.5612815618515015
 'll: 0.512611985206604
[query] year
 month: 0.6957005262374878
 quarter: 0.691483736038208
 earlier: 0.6661213636398315
 last: 0.6327787041664124
 third: 0.6230476498603821
[query] car
 luxury: 0.6767407655715942
 auto: 0.6339930295944214
 vehicle: 0.5972712635993958
 cars: 0.5888376235961914
 truck: 0.5693157315254211
[query] toyota
 motor: 0.7481387853622437
 nissan: 0.7147319316864014
 motors: 0.6946366429328918
 lexus: 0.6553674340248108
 honda: 0.6343469619750977

下面結(jié)果,是用了sklearn模塊里面的randomized_svd方法,使用了隨機(jī)數(shù)的 Truncated SVD,僅對(duì)奇異值較大的部分進(jìn)行計(jì)算,計(jì)算速度比常規(guī)的 SVD 快。

calculating SVD ...
ok
[query] you
 i: 0.6678948998451233
 we: 0.6213737726211548
 something: 0.560122013092041
 do: 0.5594725608825684
 someone: 0.5490139126777649
[query] year
 month: 0.6444296836853027
 quarter: 0.6192560791969299
 next: 0.6152222156524658
 fiscal: 0.5712860226631165
 earlier: 0.5641934871673584
[query] car
 luxury: 0.6612467765808105
 auto: 0.6166062355041504
 corsica: 0.5270425081253052
 cars: 0.5142025947570801
 truck: 0.5030257105827332
[query] toyota
 motor: 0.7747215628623962
 motors: 0.6871038675308228
 lexus: 0.6786072850227356
 nissan: 0.6618651151657104
 mazda: 0.6237337589263916
Process finished with exit code 0

以上就是“nlp計(jì)數(shù)法應(yīng)用于PTB數(shù)據(jù)集的方法”這篇文章的所有內(nèi)容,感謝各位的閱讀!相信大家閱讀完這篇文章都有很大的收獲,小編每天都會(huì)為大家更新不同的知識(shí),如果還想學(xué)習(xí)更多的知識(shí),請(qǐng)關(guān)注億速云行業(yè)資訊頻道。

向AI問一下細(xì)節(jié)

免責(zé)聲明:本站發(fā)布的內(nèi)容(圖片、視頻和文字)以原創(chuàng)、轉(zhuǎn)載和分享為主,文章觀點(diǎn)不代表本網(wǎng)站立場(chǎng),如果涉及侵權(quán)請(qǐng)聯(lián)系站長(zhǎng)郵箱:is@yisu.com進(jìn)行舉報(bào),并提供相關(guān)證據(jù),一經(jīng)查實(shí),將立刻刪除涉嫌侵權(quán)內(nèi)容。

nlp
AI