将 NLTK 树叶值作为字符串获取答案

【问题标题】：Getting NLTK tree leaf values as a string将 NLTK 树叶值作为字符串获取
【发布时间】：2016-02-13 07:38:39
【问题描述】：

我正在尝试将 Tree 对象中的叶值作为字符串获取。这里的树对象是 Stanford Parser 的输出。

这是我的代码：

from nltk.parse import stanford
Parser = stanford.StanfordParser("path")


example = "Selected variables by univariate/multivariate analysis, constructed logistic regression, calibrated the low defaults portfolio to benchmark ratings, performed back"
sentences = Parser.raw_parse(example)
for line in sentences:
    for sentence in line:
        tree = sentence

这就是我提取 VP（动词短语）叶子的方法。

VP=[]

VP_tree = list(tree.subtrees(filter=lambda x: x.label()=='VP'))

for i in VP_tree:
    VP.append(' '.join(i.flatten()))

这是 i.flatten() 的样子：（它返回已解析的单词列表）

(VP
  constructed
  logistic
  regression
  ,
  calibrated
  the
  low
  defaults
  portfolio
  to
  benchmark
  ratings)

因为我只能将它们作为已解析单词的列表来获取，所以我用 ' ' 加入了它们。因此，“回归”和“，”之间有一个空格。

In [33]: VP
Out [33]: [u'constructed logistic regression , calibrated the low defaults portfolio to benchmark ratings']

我想将动词短语作为字符串（而不是作为已解析单词的列表）获取，而不必通过 ' ' 加入它们。

我查看了 Tree 类 (http://www.nltk.org/_modules/nltk/tree.html) 下的方法，但到目前为止没有运气。

【问题讨论】：

标签： python tree nlp nltk stanford-nlp

【解决方案1】：

要根据输入位置检索字符串，您应该考虑使用https://github.com/smilli/py-corenlp 而不是 NLTK API to Stanford 工具。

首先您必须下载、安装和设置 Stanford CoreNLP，请参阅 http://stanfordnlp.github.io/CoreNLP/corenlp-server.html#getting-started

然后将python包装器安装到CoreNLP，https://github.com/smilli/py-corenlp

然后，启动服务器后（很多人错过了这一步！），在python中，你可以这样做：

>>> from pycorenlp import StanfordCoreNLP
>>> stanford = StanfordCoreNLP('http://localhost:9000')
>>> text = ("Selected variables by univariate/multivariate analysis, constructed logistic regression, calibrated the low defaults portfolio to benchmark ratings, performed back")
>>> output = stanford.annotate(text, properties={'annotators': 'tokenize,ssplit,pos,depparse,parse', 'outputFormat': 'json'})
>>> print(output['sentences'][0]['parse'])
(ROOT
  (SINV
    (VP (VBN Selected)
      (NP (NNS variables))
      (PP (IN by)
        (NP
          (NP (JJ univariate/multivariate) (NN analysis))
          (, ,)
          (VP (VBN constructed)
            (NP (JJ logistic) (NN regression)))
          (, ,))))
    (VP (VBD calibrated))
    (NP
      (NP
        (NP (DT the) (JJ low) (NNS defaults) (NN portfolio))
        (PP (TO to)
          (NP (JJ benchmark) (NNS ratings))))
      (, ,)
      (VP (VBN performed)
        (ADVP (RB back))))))

要根据输入字符串检索 VP 字符串，您必须使用 characterOffsetBegin 和 characterOffsetEnd 遍历 JSON 输出：

>>> output['sentences'][0]
{u'tokens': [{u'index': 1, u'word': u'Selected', u'after': u' ', u'pos': u'VBN', u'characterOffsetEnd': 8, u'characterOffsetBegin': 0, u'originalText': u'Selected', u'before': u''}, {u'index': 2, u'word': u'variables', u'after': u' ', u'pos': u'NNS', u'characterOffsetEnd': 18, u'characterOffsetBegin': 9, u'originalText': u'variables', u'before': u' '}, {u'index': 3, u'word': u'by', u'after': u' ', u'pos': u'IN', u'characterOffsetEnd': 21, u'characterOffsetBegin': 19, u'originalText': u'by', u'before': u' '}, {u'index': 4, u'word': u'univariate/multivariate', u'after': u' ', u'pos': u'JJ', u'characterOffsetEnd': 45, u'characterOffsetBegin': 22, u'originalText': u'univariate/multivariate', u'before': u' '}, {u'index': 5, u'word': u'analysis', u'after': u'', u'pos': u'NN', u'characterOffsetEnd': 54, u'characterOffsetBegin': 46, u'originalText': u'analysis', u'before': u' '}, {u'index': 6, u'word': u',', u'after': u' ', u'pos': u',', u'characterOffsetEnd': 55, u'characterOffsetBegin': 54, u'originalText': u',', u'before': u''}, {u'index': 7, u'word': u'constructed', u'after': u' ', u'pos': u'VBN', u'characterOffsetEnd': 67, u'characterOffsetBegin': 56, u'originalText': u'constructed', u'before': u' '}, {u'index': 8, u'word': u'logistic', u'after': u' ', u'pos': u'JJ', u'characterOffsetEnd': 76, u'characterOffsetBegin': 68, u'originalText': u'logistic', u'before': u' '}, {u'index': 9, u'word': u'regression', u'after': u'', u'pos': u'NN', u'characterOffsetEnd': 87, u'characterOffsetBegin': 77, u'originalText': u'regression', u'before': u' '}, {u'index': 10, u'word': u',', u'after': u' ', u'pos': u',', u'characterOffsetEnd': 88, u'characterOffsetBegin': 87, u'originalText': u',', u'before': u''}, {u'index': 11, u'word': u'calibrated', u'after': u' ', u'pos': u'VBD', u'characterOffsetEnd': 99, u'characterOffsetBegin': 89, u'originalText': u'calibrated', u'before': u' '}, {u'index': 12, u'word': u'the', u'after': u' ', u'pos': u'DT', u'characterOffsetEnd': 103, u'characterOffsetBegin': 100, u'originalText': u'the', u'before': u' '}, {u'index': 13, u'word': u'low', u'after': u' ', u'pos': u'JJ', u'characterOffsetEnd': 107, u'characterOffsetBegin': 104, u'originalText': u'low', u'before': u' '}, {u'index': 14, u'word': u'defaults', u'after': u' ', u'pos': u'NNS', u'characterOffsetEnd': 116, u'characterOffsetBegin': 108, u'originalText': u'defaults', u'before': u' '}, {u'index': 15, u'word': u'portfolio', u'after': u' ', u'pos': u'NN', u'characterOffsetEnd': 126, u'characterOffsetBegin': 117, u'originalText': u'portfolio', u'before': u' '}, {u'index': 16, u'word': u'to', u'after': u' ', u'pos': u'TO', u'characterOffsetEnd': 129, u'characterOffsetBegin': 127, u'originalText': u'to', u'before': u' '}, {u'index': 17, u'word': u'benchmark', u'after': u' ', u'pos': u'JJ', u'characterOffsetEnd': 139, u'characterOffsetBegin': 130, u'originalText': u'benchmark', u'before': u' '}, {u'index': 18, u'word': u'ratings', u'after': u'', u'pos': u'NNS', u'characterOffsetEnd': 147, u'characterOffsetBegin': 140, u'originalText': u'ratings', u'before': u' '}, {u'index': 19, u'word': u',', u'after': u' ', u'pos': u',', u'characterOffsetEnd': 148, u'characterOffsetBegin': 147, u'originalText': u',', u'before': u''}, {u'index': 20, u'word': u'performed', u'after': u' ', u'pos': u'VBN', u'characterOffsetEnd': 158, u'characterOffsetBegin': 149, u'originalText': u'performed', u'before': u' '}, {u'index': 21, u'word': u'back', u'after': u'', u'pos': u'RB', u'characterOffsetEnd': 163, u'characterOffsetBegin': 159, u'originalText': u'back', u'before': u' '}], u'index': 0, u'basic-dependencies': [{u'dep': u'ROOT', u'dependent': 1, u'governorGloss': u'ROOT', u'governor': 0, u'dependentGloss': u'Selected'}, {u'dep': u'dobj', u'dependent': 2, u'governorGloss': u'Selected', u'governor': 1, u'dependentGloss': u'variables'}, {u'dep': u'case', u'dependent': 3, u'governorGloss': u'analysis', u'governor': 5, u'dependentGloss': u'by'}, {u'dep': u'amod', u'dependent': 4, u'governorGloss': u'analysis', u'governor': 5, u'dependentGloss': u'univariate/multivariate'}, {u'dep': u'nmod', u'dependent': 5, u'governorGloss': u'Selected', u'governor': 1, u'dependentGloss': u'analysis'}, {u'dep': u'punct', u'dependent': 6, u'governorGloss': u'analysis', u'governor': 5, u'dependentGloss': u','}, {u'dep': u'acl', u'dependent': 7, u'governorGloss': u'analysis', u'governor': 5, u'dependentGloss': u'constructed'}, {u'dep': u'amod', u'dependent': 8, u'governorGloss': u'regression', u'governor': 9, u'dependentGloss': u'logistic'}, {u'dep': u'dobj', u'dependent': 9, u'governorGloss': u'constructed', u'governor': 7, u'dependentGloss': u'regression'}, {u'dep': u'punct', u'dependent': 10, u'governorGloss': u'analysis', u'governor': 5, u'dependentGloss': u','}, {u'dep': u'dep', u'dependent': 11, u'governorGloss': u'Selected', u'governor': 1, u'dependentGloss': u'calibrated'}, {u'dep': u'det', u'dependent': 12, u'governorGloss': u'portfolio', u'governor': 15, u'dependentGloss': u'the'}, {u'dep': u'amod', u'dependent': 13, u'governorGloss': u'portfolio', u'governor': 15, u'dependentGloss': u'low'}, {u'dep': u'compound', u'dependent': 14, u'governorGloss': u'portfolio', u'governor': 15, u'dependentGloss': u'defaults'}, {u'dep': u'nsubj', u'dependent': 15, u'governorGloss': u'Selected', u'governor': 1, u'dependentGloss': u'portfolio'}, {u'dep': u'case', u'dependent': 16, u'governorGloss': u'ratings', u'governor': 18, u'dependentGloss': u'to'}, {u'dep': u'amod', u'dependent': 17, u'governorGloss': u'ratings', u'governor': 18, u'dependentGloss': u'benchmark'}, {u'dep': u'nmod', u'dependent': 18, u'governorGloss': u'portfolio', u'governor': 15, u'dependentGloss': u'ratings'}, {u'dep': u'punct', u'dependent': 19, u'governorGloss': u'portfolio', u'governor': 15, u'dependentGloss': u','}, {u'dep': u'acl', u'dependent': 20, u'governorGloss': u'portfolio', u'governor': 15, u'dependentGloss': u'performed'}, {u'dep': u'advmod', u'dependent': 21, u'governorGloss': u'performed', u'governor': 20, u'dependentGloss': u'back'}], u'parse': u'(ROOT\n  (SINV\n    (VP (VBN Selected)\n      (NP (NNS variables))\n      (PP (IN by)\n        (NP\n          (NP (JJ univariate/multivariate) (NN analysis))\n          (, ,)\n          (VP (VBN constructed)\n            (NP (JJ logistic) (NN regression)))\n          (, ,))))\n    (VP (VBD calibrated))\n    (NP\n      (NP\n        (NP (DT the) (JJ low) (NNS defaults) (NN portfolio))\n        (PP (TO to)\n          (NP (JJ benchmark) (NNS ratings))))\n      (, ,)\n      (VP (VBN performed)\n        (ADVP (RB back))))))', u'collapsed-dependencies': [{u'dep': u'ROOT', u'dependent': 1, u'governorGloss': u'ROOT', u'governor': 0, u'dependentGloss': u'Selected'}, {u'dep': u'dobj', u'dependent': 2, u'governorGloss': u'Selected', u'governor': 1, u'dependentGloss': u'variables'}, {u'dep': u'case', u'dependent': 3, u'governorGloss': u'analysis', u'governor': 5, u'dependentGloss': u'by'}, {u'dep': u'amod', u'dependent': 4, u'governorGloss': u'analysis', u'governor': 5, u'dependentGloss': u'univariate/multivariate'}, {u'dep': u'nmod:by', u'dependent': 5, u'governorGloss': u'Selected', u'governor': 1, u'dependentGloss': u'analysis'}, {u'dep': u'punct', u'dependent': 6, u'governorGloss': u'analysis', u'governor': 5, u'dependentGloss': u','}, {u'dep': u'acl', u'dependent': 7, u'governorGloss': u'analysis', u'governor': 5, u'dependentGloss': u'constructed'}, {u'dep': u'amod', u'dependent': 8, u'governorGloss': u'regression', u'governor': 9, u'dependentGloss': u'logistic'}, {u'dep': u'dobj', u'dependent': 9, u'governorGloss': u'constructed', u'governor': 7, u'dependentGloss': u'regression'}, {u'dep': u'punct', u'dependent': 10, u'governorGloss': u'analysis', u'governor': 5, u'dependentGloss': u','}, {u'dep': u'dep', u'dependent': 11, u'governorGloss': u'Selected', u'governor': 1, u'dependentGloss': u'calibrated'}, {u'dep': u'det', u'dependent': 12, u'governorGloss': u'portfolio', u'governor': 15, u'dependentGloss': u'the'}, {u'dep': u'amod', u'dependent': 13, u'governorGloss': u'portfolio', u'governor': 15, u'dependentGloss': u'low'}, {u'dep': u'compound', u'dependent': 14, u'governorGloss': u'portfolio', u'governor': 15, u'dependentGloss': u'defaults'}, {u'dep': u'nsubj', u'dependent': 15, u'governorGloss': u'Selected', u'governor': 1, u'dependentGloss': u'portfolio'}, {u'dep': u'case', u'dependent': 16, u'governorGloss': u'ratings', u'governor': 18, u'dependentGloss': u'to'}, {u'dep': u'amod', u'dependent': 17, u'governorGloss': u'ratings', u'governor': 18, u'dependentGloss': u'benchmark'}, {u'dep': u'nmod:to', u'dependent': 18, u'governorGloss': u'portfolio', u'governor': 15, u'dependentGloss': u'ratings'}, {u'dep': u'punct', u'dependent': 19, u'governorGloss': u'portfolio', u'governor': 15, u'dependentGloss': u','}, {u'dep': u'acl', u'dependent': 20, u'governorGloss': u'portfolio', u'governor': 15, u'dependentGloss': u'performed'}, {u'dep': u'advmod', u'dependent': 21, u'governorGloss': u'performed', u'governor': 20, u'dependentGloss': u'back'}], u'collapsed-ccprocessed-dependencies': [{u'dep': u'ROOT', u'dependent': 1, u'governorGloss': u'ROOT', u'governor': 0, u'dependentGloss': u'Selected'}, {u'dep': u'dobj', u'dependent': 2, u'governorGloss': u'Selected', u'governor': 1, u'dependentGloss': u'variables'}, {u'dep': u'case', u'dependent': 3, u'governorGloss': u'analysis', u'governor': 5, u'dependentGloss': u'by'}, {u'dep': u'amod', u'dependent': 4, u'governorGloss': u'analysis', u'governor': 5, u'dependentGloss': u'univariate/multivariate'}, {u'dep': u'nmod:by', u'dependent': 5, u'governorGloss': u'Selected', u'governor': 1, u'dependentGloss': u'analysis'}, {u'dep': u'punct', u'dependent': 6, u'governorGloss': u'analysis', u'governor': 5, u'dependentGloss': u','}, {u'dep': u'acl', u'dependent': 7, u'governorGloss': u'analysis', u'governor': 5, u'dependentGloss': u'constructed'}, {u'dep': u'amod', u'dependent': 8, u'governorGloss': u'regression', u'governor': 9, u'dependentGloss': u'logistic'}, {u'dep': u'dobj', u'dependent': 9, u'governorGloss': u'constructed', u'governor': 7, u'dependentGloss': u'regression'}, {u'dep': u'punct', u'dependent': 10, u'governorGloss': u'analysis', u'governor': 5, u'dependentGloss': u','}, {u'dep': u'dep', u'dependent': 11, u'governorGloss': u'Selected', u'governor': 1, u'dependentGloss': u'calibrated'}, {u'dep': u'det', u'dependent': 12, u'governorGloss': u'portfolio', u'governor': 15, u'dependentGloss': u'the'}, {u'dep': u'amod', u'dependent': 13, u'governorGloss': u'portfolio', u'governor': 15, u'dependentGloss': u'low'}, {u'dep': u'compound', u'dependent': 14, u'governorGloss': u'portfolio', u'governor': 15, u'dependentGloss': u'defaults'}, {u'dep': u'nsubj', u'dependent': 15, u'governorGloss': u'Selected', u'governor': 1, u'dependentGloss': u'portfolio'}, {u'dep': u'case', u'dependent': 16, u'governorGloss': u'ratings', u'governor': 18, u'dependentGloss': u'to'}, {u'dep': u'amod', u'dependent': 17, u'governorGloss': u'ratings', u'governor': 18, u'dependentGloss': u'benchmark'}, {u'dep': u'nmod:to', u'dependent': 18, u'governorGloss': u'portfolio', u'governor': 15, u'dependentGloss': u'ratings'}, {u'dep': u'punct', u'dependent': 19, u'governorGloss': u'portfolio', u'governor': 15, u'dependentGloss': u','}, {u'dep': u'acl', u'dependent': 20, u'governorGloss': u'portfolio', u'governor': 15, u'dependentGloss': u'performed'}, {u'dep': u'advmod', u'dependent': 21, u'governorGloss': u'performed', u'governor': 20, u'dependentGloss': u'back'}]}

但它似乎不是一个容易解析以获得字符偏移量的输出，因为解析树没有直接链接到偏移量。只有依赖三元组包含指向偏移量的单词 ID 的链接。

要访问output['sentences'][0]['tokens'] 中的标记和'after' 和'before' 键（但遗憾的是没有直接链接到解析树）：

>>> tokens = output['sentences'][0]['tokens']
>>> tokens
[{u'index': 1, u'word': u'Selected', u'after': u' ', u'pos': u'VBN', u'characterOffsetEnd': 8, u'characterOffsetBegin': 0, u'originalText': u'Selected', u'before': u''}, {u'index': 2, u'word': u'variables', u'after': u' ', u'pos': u'NNS', u'characterOffsetEnd': 18, u'characterOffsetBegin': 9, u'originalText': u'variables', u'before': u' '}, {u'index': 3, u'word': u'by', u'after': u' ', u'pos': u'IN', u'characterOffsetEnd': 21, u'characterOffsetBegin': 19, u'originalText': u'by', u'before': u' '}, {u'index': 4, u'word': u'univariate/multivariate', u'after': u' ', u'pos': u'JJ', u'characterOffsetEnd': 45, u'characterOffsetBegin': 22, u'originalText': u'univariate/multivariate', u'before': u' '}, {u'index': 5, u'word': u'analysis', u'after': u'', u'pos': u'NN', u'characterOffsetEnd': 54, u'characterOffsetBegin': 46, u'originalText': u'analysis', u'before': u' '}, {u'index': 6, u'word': u',', u'after': u' ', u'pos': u',', u'characterOffsetEnd': 55, u'characterOffsetBegin': 54, u'originalText': u',', u'before': u''}, {u'index': 7, u'word': u'constructed', u'after': u' ', u'pos': u'VBN', u'characterOffsetEnd': 67, u'characterOffsetBegin': 56, u'originalText': u'constructed', u'before': u' '}, {u'index': 8, u'word': u'logistic', u'after': u' ', u'pos': u'JJ', u'characterOffsetEnd': 76, u'characterOffsetBegin': 68, u'originalText': u'logistic', u'before': u' '}, {u'index': 9, u'word': u'regression', u'after': u'', u'pos': u'NN', u'characterOffsetEnd': 87, u'characterOffsetBegin': 77, u'originalText': u'regression', u'before': u' '}, {u'index': 10, u'word': u',', u'after': u' ', u'pos': u',', u'characterOffsetEnd': 88, u'characterOffsetBegin': 87, u'originalText': u',', u'before': u''}, {u'index': 11, u'word': u'calibrated', u'after': u' ', u'pos': u'VBD', u'characterOffsetEnd': 99, u'characterOffsetBegin': 89, u'originalText': u'calibrated', u'before': u' '}, {u'index': 12, u'word': u'the', u'after': u' ', u'pos': u'DT', u'characterOffsetEnd': 103, u'characterOffsetBegin': 100, u'originalText': u'the', u'before': u' '}, {u'index': 13, u'word': u'low', u'after': u' ', u'pos': u'JJ', u'characterOffsetEnd': 107, u'characterOffsetBegin': 104, u'originalText': u'low', u'before': u' '}, {u'index': 14, u'word': u'defaults', u'after': u' ', u'pos': u'NNS', u'characterOffsetEnd': 116, u'characterOffsetBegin': 108, u'originalText': u'defaults', u'before': u' '}, {u'index': 15, u'word': u'portfolio', u'after': u' ', u'pos': u'NN', u'characterOffsetEnd': 126, u'characterOffsetBegin': 117, u'originalText': u'portfolio', u'before': u' '}, {u'index': 16, u'word': u'to', u'after': u' ', u'pos': u'TO', u'characterOffsetEnd': 129, u'characterOffsetBegin': 127, u'originalText': u'to', u'before': u' '}, {u'index': 17, u'word': u'benchmark', u'after': u' ', u'pos': u'JJ', u'characterOffsetEnd': 139, u'characterOffsetBegin': 130, u'originalText': u'benchmark', u'before': u' '}, {u'index': 18, u'word': u'ratings', u'after': u'', u'pos': u'NNS', u'characterOffsetEnd': 147, u'characterOffsetBegin': 140, u'originalText': u'ratings', u'before': u' '}, {u'index': 19, u'word': u',', u'after': u' ', u'pos': u',', u'characterOffsetEnd': 148, u'characterOffsetBegin': 147, u'originalText': u',', u'before': u''}, {u'index': 20, u'word': u'performed', u'after': u' ', u'pos': u'VBN', u'characterOffsetEnd': 158, u'characterOffsetBegin': 149, u'originalText': u'performed', u'before': u' '}, {u'index': 21, u'word': u'back', u'after': u'', u'pos': u'RB', u'characterOffsetEnd': 163, u'characterOffsetBegin': 159, u'originalText': u'back', u'before': u' '}]

【讨论】：

由于 OP 想要获取正确间隔的字符串，而不是尝试使用偏移量，您可以只使用记录标记之间的空格的 'before' 和 'after' 属性。您可以看到逗号有一个 '' 的“之前” - 并且该分析有一个 '' 的“之后”。
是否可以将令牌索引链接到 JSON 的“解析”键？这将有很大帮助，因为我们可以从解析输出中轻松访问标记的索引。还是更容易解析 'collapsed-dependencies'/'basic-dependencies' 键上的子 JSON？

【解决方案2】：

简而言之：

使用Tree.leaves()函数访问解析语句中子树的字符串，即：

VPs_str = [" ".join(vp.leaves()) for vp in list(parsed_sent.subtrees(filter=lambda x: x.label()=='VP'))]

没有正确的方法可以访问输入中的真正 VP 字符串，因为斯坦福解析器在解析过程之前对文本进行了标记，并且 NLTK API 没有保留字符串的偏移量 =(

长篇大论：

这个长答案使得其他 NLTK 用户可以使用斯坦福解析器的 NLTK API 访问Tree 对象，它可能不像问题中显示的那么简单 =)

首先为 NLTK 设置环境变量以访问斯坦福工具，请参阅：

TL;DR：

$ cd
$ wget http://nlp.stanford.edu/software/stanford-parser-full-2015-12-09.zip
$ unzip stanford-parser-full-2015-12-09.zip
$ export STANFORDTOOLSDIR=$HOME
$ export CLASSPATH=$STANFORDTOOLSDIR/stanford-parser-full-2015-12-09/stanford-parser.jar:$STANFORDTOOLSDIR/stanford-parser-full-2015-12-09/stanford-parser-3.6.0-models.jar

为 2015 年 12 月 9 日编译的 Stanford Parser 应用 hack（此 hack 将在带有 https://github.com/nltk/nltk/pull/1280/files 的前沿版本中过时）：

>>> from nltk.internals import find_jars_within_path
>>> from nltk.parse.stanford import StanfordParser
>>> parser=StanfordParser(model_path="edu/stanford/nlp/models/lexparser/englishPCFG.ser.gz")
>>> stanford_dir = parser._classpath[0].rpartition('/')[0]
>>> parser._classpath = tuple(find_jars_within_path(stanford_dir))

现在进行短语提取。

首先，我们解析句子：

>>> sent = "Selected variables by univariate/multivariate analysis, constructed logistic regression, calibrated the low defaults portfolio to benchmark ratings, performed back"
>>> parsed_sent = list(parser.raw_parse(sent))[0]
>>> parsed_sent
Tree('ROOT', [Tree('S', [Tree('NP', [Tree('NP', [Tree('JJ', ['Selected']), Tree('NNS', ['variables'])]), Tree('PP', [Tree('IN', ['by']), Tree('NP', [Tree('JJ', ['univariate/multivariate']), Tree('NN', ['analysis'])])]), Tree(',', [',']), Tree('VP', [Tree('VBN', ['constructed']), Tree('NP', [Tree('NP', [Tree('JJ', ['logistic']), Tree('NN', ['regression'])]), Tree(',', [',']), Tree('ADJP', [Tree('VBN', ['calibrated']), Tree('NP', [Tree('NP', [Tree('DT', ['the']), Tree('JJ', ['low']), Tree('NNS', ['defaults']), Tree('NN', ['portfolio'])]), Tree('PP', [Tree('TO', ['to']), Tree('NP', [Tree('JJ', ['benchmark']), Tree('NNS', ['ratings'])])])])])])]), Tree(',', [','])]), Tree('VP', [Tree('VBD', ['performed']), Tree('ADVP', [Tree('RB', ['back'])])])])])

然后我们遍历树并检查 VP，就像您所做的那样：

>>> VP_tree = list(tree.subtrees(filter=lambda x: x.label()=='VP'))

之后，我们只需使用子树的叶子来获取 VPs

>>> for vp in VPs:
...     print " ".join(vp.leaves())
... 
constructed logistic regression , calibrated the low defaults portfolio to benchmark ratings
performed back

所以要获取 VP 字符串：

>>> VPs_str = [" ".join(vp.leaves()) for vp in list(parsed_sent.subtrees(filter=lambda x: x.label()=='VP'))]
>>> VPs_str
[u'constructed logistic regression , calibrated the low defaults portfolio to benchmark ratings', u'performed back']

另外，我个人更喜欢使用分块器而不是成熟的解析器来提取短语。

使用nltk_cli 工具 (https://github.com/alvations/nltk_cli)：

alvas@ubi:~/git/nltk_cli$ echo "Selected variables by univariate/multivariate analysis, constructed logistic regression, calibrated the low defaults portfolio to benchmark ratings, performed back" > input-doneyo.txt
alvas@ubi:~/git/nltk_cli$ python senna.py --chunk VP input-doneyo.txt calibrated|to benchmark|performed
alvas@ubi:~/git/nltk_cli$ python senna.py --vp input-doneyo.txt 
calibrated|to benchmark|performed
alvas@ubi:~/git/nltk_cli$ python senna.py --chunk2 VP+NP input-doneyo.txt 
calibrated  the low defaults portfolio|to benchmark ratings

VP标签的输出由|分隔，即

输出：

calibrated|to benchmark|performed

代表：

校准
基准测试
已执行

而且VP+NP块输出也用|隔开，VP和NP用\t隔开，即

输出：

calibrated  the low defaults portfolio|to benchmark ratings

代表（VP + NP）：

校准 + 低默认组合
基准+评级

【讨论】：

【解决方案3】：

与NLTK 或StanfordParser 无关，获得正常阅读文本的一种方法是使用来自 Moses SMT (https://github.com/moses-smt/mosesdecoder) 的脚本“detokenize”输出，例如：

alvas@ubi:~$ wget https://raw.githubusercontent.com/moses-smt/mosesdecoder/master/scripts/tokenizer/detokenizer.perl
--2016-02-13 21:27:12--  https://raw.githubusercontent.com/moses-smt/mosesdecoder/master/scripts/tokenizer/detokenizer.perl
Resolving raw.githubusercontent.com (raw.githubusercontent.com)... 23.235.43.133
Connecting to raw.githubusercontent.com (raw.githubusercontent.com)|23.235.43.133|:443... connected.
HTTP request sent, awaiting response... 200 OK
Length: 12473 (12K) [text/plain]
Saving to: ‘detokenizer.perl’

100%[===============================================================================================================================>] 12,473      --.-K/s   in 0s      

2016-02-13 21:27:12 (150 MB/s) - ‘detokenizer.perl’ saved [12473/12473]

alvas@ubi:~$ echo "constructed logistic regression , calibrated the low defaults portfolio to benchmark ratings" 2> /tmp/null
constructed logistic regression , calibrated the low defaults portfolio to benchmark ratings

请注意，输出可能与输入不同，但在大多数情况下，对于英语，它将被转换为我们读/写的普通文本。

在 NLTK 中有一个 detokenizer 正在筹备中，但我们需要一段时间来对其进行编码、测试并将其推送到存储库，请您耐心等待（请参阅 https://github.com/nltk/nltk/issues/1214）

【讨论】：