下面我们从语料中抽取出pairs。这个步骤超参很多，比如窗口大小，高频词低频词怎么处理，提取什么样的中心词上下文特征都是在这个步骤中确定的。这个步骤也决定了整个后面模型的信息源。这个过程会调用多线程，因为提取pairs的过程每行之间都是独立的。先看看main函数。主要功能就是开线程，把参数传递给每个线程。
def main(): args = docopt(""" Usage: corpus2pairs.py [options] <corpus> <vocab> <pairs> Options: --win NUM Window size [default: 2] --sub NUM Subsampling threshold [default: 0] --ngram_word NUM (Center) word vocabulary includes grams of 1st to nth order [default: 1] --ngram_context NUM Context vocabulary includes grams of 1st to nth order [default: 1] --threads_num NUM The number of threads [default: 8] --overlap Whether overlaping pairs are allowed or not """) print "**********************" print "corpus2pairs" threads_num = int(args['--threads_num']) threads_list = [] for i in xrange(0, threads_num): #extract pairs from corpus through multipule threads thread = multiprocessing.Process(target=c2p, args=(args, i))//开启多线程，把参数和线程id都传到线程中去 thread.start()//开启线程 threads_list.append(thread) for thread in threads_list: thread.join()//等待线程结束 print "corpus2pairs finished"
然后看一下每个线程都做什么。每个线程首先要根据当前处理的行号判断一下是不是这个线程需要处理的。如果是的就开始处理，从一行中抽取pairs，抽取的过程中会进行一系列的逻辑判断，确认是否抽取这个pair。
def c2p(args, tid): pairs_file = open(args['<pairs>']+"_"+str(tid), 'w')//每个线程都会把抽取的pairs写到自己的文件中，最后再合并。 win = int(args['--win']) subsample = float(args['--sub']) sub = subsample != 0 ngram_word = int(args['--ngram_word'])//中心词考虑到几阶的ngram ngram_context = int(args['--ngram_context'])//上下文考虑到几阶的ngram overlap = args['--overlap']//pair中的ngram如果重叠，这个pair是否还抽取出来 threads_num = int(args['--threads_num']) vocab = load_count_vocabulary(args['<vocab>']) #load vocabulary (generated in corpus2vocab stage)//读取corpus2vocab生成的词典，只有单词在词典中才会被抽取出来 train_uni_num = 0 #number of (unigram) tokens in corpus//计算语料中token的数量，忽略了低频词，所以不准 for w, c in vocab.iteritems(): if '@$' not in w: train_uni_num += c train_num = sum(vocab.values()) #number of (ngram) tokens in corpus if tid == 0: print 'vocabulary size: ' + str(len(vocab)) print 'number of training words (uni-grams): ' + str(train_uni_num) print 'number of training n-grams: ' + str(train_num) subsample *= train_uni_num//做subsampling，和hyperwords中的代码类似 if sub: subsampler = dict([(word, 1 - sqrt(subsample / count)) for word, count in vocab.iteritems() if count > subsample]) #subsampling technique rnd = Random(17) with open(args['<corpus>']) as f: line_num = 0 if tid == 0: print str(line_num/1000**1) + "K lines processed." for line in f://遍历文件的每一行 line_num += 1 if ((line_num) % 1000) == 0 and tid == 0: print "\x1b[1A" + str(line_num/1000) + "K lines processed." if line_num % threads_num != tid://每个线程处理指定的行 continue tokens = line.strip().split()//得到每行的单词列表 for i in xrange(len(tokens)): #loop for each position in a line//对单词列表的每一个位置循环 for gram_word in xrange(1, ngram_word+1): #loop for grams of different orders in (center) word //在每个位置上面考虑中心词ngram word = getNgram(tokens, i, gram_word)//中心词 word = check_word(word, vocab, sub, subsampler, rnd)//用一系列逻辑判断这个中心词是否合格会被留下，不合格的话以其为中心词的pair也可以被扔掉 if word is None: continue for gram_context in xrange(1, ngram_context+1): #loop for grams of different orders in context//对于一个中心词考虑期上下文的ngram start = i - win + gram_word - 1//根据窗口大小确定上下文ngram的起点和终点 end = i + win - gram_context + 1 for j in xrange(start, end + 1)://遍历上下文窗口的所有ngram if overlap: if i == j and gram_word == gram_context://判断中心词上下文是不是一个ngram continue else: if len(set(range(i, i + gram_word)) & set(range(j, j + gram_context))) > 0://判断中心词上下文是否有重叠 continue context = getNgram(tokens, j, gram_context)//得到上下文特征 context = check_word(context, vocab, sub, subsampler, rnd)//通过一系列逻辑判断要不要这个上下文，不要的话pair也就不要了 if context is None: continue pairs_file.write(word + ' ' + context + "\n") #write pairs to the file//最终版满足条件的pair写到文件中。 pairs_file.close()
最后看看我们用什么样的逻辑判断我们到底要不要一个单词。
def check_word(t, vocab, sub, subsampler, rnd): #discard tokens if t is None://首先单词不是None return None if sub://然后不被subsampling掉，过滤高频词的手段 t = t if t not in subsampler or rnd.random() > subsampler[t] else None if t is None: return None t = t if t in vocab else None//单词要在之前从corpus2vocab中生成的词典 return t
