3.4 无限制依赖成分

考虑下面的对比：

  1. >>> nltk.data.show_cfg('grammars/book_grammars/feat1.fcfg')
  2. % start S
  3. # ###################
  4. # Grammar Productions
  5. # ###################
  6. S[-INV] -> NP VP
  7. S[-INV]/?x -> NP VP/?x
  8. S[-INV] -> NP S/NP
  9. S[-INV] -> Adv[+NEG] S[+INV]
  10. S[+INV] -> V[+AUX] NP VP
  11. S[+INV]/?x -> V[+AUX] NP VP/?x
  12. SBar -> Comp S[-INV]
  13. SBar/?x -> Comp S[-INV]/?x
  14. VP -> V[SUBCAT=intrans, -AUX]
  15. VP -> V[SUBCAT=trans, -AUX] NP
  16. VP/?x -> V[SUBCAT=trans, -AUX] NP/?x
  17. VP -> V[SUBCAT=clause, -AUX] SBar
  18. VP/?x -> V[SUBCAT=clause, -AUX] SBar/?x
  19. VP -> V[+AUX] VP
  20. VP/?x -> V[+AUX] VP/?x
  21. # ###################
  22. # Lexical Productions
  23. # ###################
  24. V[SUBCAT=intrans, -AUX] -> 'walk' | 'sing'
  25. V[SUBCAT=trans, -AUX] -> 'see' | 'like'
  26. V[SUBCAT=clause, -AUX] -> 'say' | 'claim'
  27. V[+AUX] -> 'do' | 'can'
  28. NP[-WH] -> 'you' | 'cats'
  29. NP[+WH] -> 'who'
  30. Adv[+NEG] -> 'rarely' | 'never'
  31. NP/NP ->
  32. Comp -> 'that'

3.1中的语法包含一个“缺口引进”产生式，即S[-INV] -> NP S/NP。为了正确的预填充斜线特征，我们需要为扩展S，VP和NP的产生式中箭头两侧的斜线添加变量值。例如，VP/?x -> V SBar/?x是VP -> V SBar的斜线版本，也就是说，可以为一个成分的父母VP指定斜线值，只要也为孩子SBar指定同样的值。最后，NP/NP ->允许NP上的斜线信息为空字符串。使用3.1中的语法，我们可以分析序列 who do you claim that you like

  1. >>> tokens = 'who do you claim that you like'.split()
  2. >>> from nltk import load_parser
  3. >>> cp = load_parser('grammars/book_grammars/feat1.fcfg')
  4. >>> for tree in cp.parse(tokens):
  5. ...     print(tree)
  6. (S[-INV]
  7.  (NP[+WH] who)
  8.  (S[+INV]/NP[]
  9.  (V[+AUX] do)
  10.  (NP[-WH] you)
  11.  (VP[]/NP[]
  12.  (V[-AUX, SUBCAT='clause'] claim)
  13.  (SBar[]/NP[]
  14.  (Comp[] that)
  15.  (S[-INV]/NP[]
  16.  (NP[-WH] you)
  17.  (VP[]/NP[] (V[-AUX, SUBCAT='trans'] like) (NP[]/NP[] )))))))

这棵树的一个更易读的版本如(52)所示。

  1. >>> tokens = 'you claim that you like cats'.split()
  2. >>> for tree in cp.parse(tokens):
  3. ...     print(tree)
  4. (S[-INV]
  5.  (NP[-WH] you)
  6.  (VP[]
  7.  (V[-AUX, SUBCAT='clause'] claim)
  8.  (SBar[]
  9.  (Comp[] that)
  10.  (S[-INV]
  11.  (NP[-WH] you)
  12.  (VP[] (V[-AUX, SUBCAT='trans'] like) (NP[-WH] cats))))))

此外，它还允许没有 wh 结构的倒装句：

  1. >>> tokens = 'rarely do you sing'.split()
  2. >>> for tree in cp.parse(tokens):
  3. ...     print(tree)
  4. (S[-INV]
  5.  (Adv[+NEG] rarely)
  6.  (S[+INV]
  7.  (V[+AUX] do)
  8.  (NP[-WH] you)
  9.  (VP[] (V[-AUX, SUBCAT='intrans'] sing))))