Parallel Tools for Natural Language Processing

1
Parallel Tools for Natural Language Processing

• Mark Brigham
• Melanie Goetz
• Andrew Hogue

6.338 / 18.337 - March 16, 2004
2
Sentence Parsing

• Consider the sentence
• John ate the cookie on the table
• We want to
• Tag the sentence with parts of speech
• Group the words by phrase

3
Context Free Grammars
S ? NP VP
NP ? Det N
NP ? NP PP
VP ? VP PP
VP ? V NP
N ? cookie
N ? table
Det ? the
V ? ate

• Recursive set of rules
• Defines what syntactic structure can be applied
  to a phrase or word
• Top-level rule S defines the sentence

4
Context Free Grammars

• Applying a CFG to a sentence creates a parse-tree
  for that sentence

5
Context Free Grammars
Top-down parse
6
Context Free Grammars
Bottom-up parse
Parallelizable!
7
Ambiguity
More than one parse for a single sentence!
8
Parallelization

• Bottom-up rule application appropriate for
  parallel processing
• Ambiguous parses also parallelizable
• Long, complex sentences may be most interesting
• Proust?

9
Chart Parsing

• Create a matrix where entries correspond to
  words/phrases
• If there is a valid CFG parse of a phrase i,j,
  add it to that matrix cell
• A cell i,j may only depend on other cells m,n
  where m lt i and n lt j.

10
John ate the cookie on the table
John
ate
the
cookie
on
the
table
11
John ate the cookie on the table
John
ate
the
cookie
on
the
table
12
John ate the cookie on the table
John
ate
the
cookie
on
the
table
13
John ate the cookie on the table
John
ate
the
cookie
on
the
table
14
John ate the cookie on the table
John
ate
the
cookie
on
the
table
15
John ate the cookie on the table
John
ate
the
cookie
on
the
table
16
John ate the cookie on the table
John
ate
the
cookie
on
the
table
17
John ate the cookie on the table
John
ate
the
cookie
on
the
table
18
John ate the cookie on the table
John
ate
the
cookie
on
the
table
19
John ate the cookie on the table
John
ate
the
cookie
on
the
table
20
John ate the cookie on the table
John
ate
the
cookie
on
the
table
21
John ate the cookie on the table
John
ate
the
cookie
on
the
table
22
John ate the cookie on the table
John
ate
the
cookie
on
the
table
23
John ate the cookie on the table
John
ate
the
cookie
on
the
table
24
John ate the cookie on the table
John
ate
the
cookie
on
the
table
25
John ate the cookie on the table
John
ate
the
cookie
on
the
table
26
John ate the cookie on the table
John
ate
the
cookie
on
the
table
27
John ate the cookie on the table
John
ate
the
cookie
on
the
table
28
Other Tools

• Considering parallelizing other NLP tools
• Word-stemming Multiple finite state automata
  applied to a single word in parallel
• Automated part-of-speech recognition on large
  corpora