python - 关系抽取中如何获取有方向的实体？

Question

我从事关系提取工作已经一周了。但我需要的是两个实体之间的方向，例如 Company_x 被 Company_y 收购。因此，模型应该预测像 Company_y->bought-> Company_X 这样的实体。你们认为有什么模型对此有帮助吗？

Answer 1

被动语态通常可以很好地指示关系的方向。

您可以从两个实体之间的上下文中提取以动词开头的模式，然后检测被动语态是否存在。

一些简单的概念验证代码(使用 NLTK 的 RegexpParser 实际上可以更简单)

from nltk import pos_tag
from nltk import word_tokenize
from nltk.stem.wordnet import WordNetLemmatizer

lmtzr = WordNetLemmatizer()
aux_verbs = ['be']

def detect_passive_voice(pattern):
    passive_voice = False

    if len(pattern) >= 3:
        if pattern[0][1].startswith('V'):
            verb = lmtzr.lemmatize(pattern[0][0], 'v')
            if verb in aux_verbs:
                if (pattern[1][1] == 'VBN' or pattern[1][1] == 'VBD') and pattern[-1][0] == 'by':
                    passive_voice = True

            # past verb + by
            elif (pattern[-2][1] == 'VBN' or pattern[-2][1] == 'VBD') and pattern[-1][0] == 'by':
                passive_voice = True

        # past verb + by
        elif (pattern[-2][1] == 'VBN' or pattern[-2][1] == 'VBD') and pattern[-1][0] == 'by':
                passive_voice = True

    # past verb + by
    elif len(pattern) >= 2:
        if (pattern[-2][1] == 'VBN' or pattern[-2][1] == 'VBD') and pattern[-1][0] == 'by':
            passive_voice = True

return passive_voice

运行一些示例:

In [4]: tokens = word_tokenize("was bought by")
   ...: tags = pos_tag(tokens)
   ...: detect_passive_voice(tags)
Out[4]: True

In [5]: tokens = word_tokenize("mailed the letter")
   ...: tags = pos_tag(tokens)
   ...: detect_passive_voice(tags)
Out[5]: False

In [7]: tokens = word_tokenize("was mailed by")
   ...: tags = pos_tag(tokens)
   ...: detect_passive_voice(tags)
Out[7]: True

您可以添加更多助动词，也可以允许中间存在副词或形容词。