1. Sparse Compact Directed Acyclic Word Graphs
Shunsuke Inenaga
(Japan Society for the Promotion of Science & Kyushu University)
Masayuki Takeda
(Kyushu University & Japan Science and Technology Agency)

2. Traditional Pattern Matching Problem
Given：text T in S* and pattern P in S*
Return：whether or not P appears in T
S: alphabet (set of characters)
S*: set of strings
A text indexing structure for T enables you to solve the above problem in O(m) time (for fixed S).
m: the length of P

3. Suffix Trie A trie representing all suffixes of T $ T = aacb$ a b c

4. Unwanted Matching
pattern s
pace runner
text

5. Unwanted Matching
pattern s
pace runner
text

6. Unwanted Matching
pattern s
pace runner
text

7. Introducing Word Separator #
# : word separator - special symbol not in S
D = S* # : dictionary of words
Text T : an element of D+ (T is a sequence T1T2…Tk of k words in D)
e.g., T = This#is#a#pen#
S = {A,…,z}
D = {...,This#,...,a#,...is#,...pen#,...}

8. Word-level Pattern Matching Problem
Given: text T in D+ and pattern P in D+
Return: whether or not P appears at the beginning of any word in T
e.g.
T = The#space#runner#is#not#your#good#pace#runner#
P = pace#runner#

9. Word-level Pattern Matching Problem
Given: text T in D+ and pattern P in D+
Return: whether or not P appears at the beginning of any word in T
e.g.
T = The#space#runner#is#not#your#good#pace#runner#
P = pace#runner#

10. Word Suffix Trie
A trie representing the suffixes of T which begin at a word.
T = aa#b# a b
aa#b#
a#b#
#b# b# # a # # b #

11. Normal and Word Suffix Tries
T = aa#b# a a b b # a # a # # b # # b # b b # # #
Normal Suffix Trie
Word Suffix Trie

12. Normal and Word Suffix Trees
T = aa#b# a b b # a # # a a b # # # # b b b # # #
Normal Suffix Tree
Word Suffix Tree

13. Sizes of Word Suffix Tries and Trees
For text T = T1T2…Tk of length n,
the word suffix trie of T requires O(nk) space, but
the word suffix tree of T requires O(k) space!!
because the word suffix tree has only k leaves and has only branching internal nodes.

14. Construction of Word Suffix Trees
Algorithm by Andersson et al.（1996）
for text T = T1T2…Tk of length n, constructs word suffix trees in O(n) expected time with O(k) space.
Our algorithm (CPM’06)
builds word suffix trees in O(n) time in the worst case, with O(k) space.

15. Our Construction Algorithm
We modify Ukkonen’s on-line normal suffix tree construction algorithm by using minimum DFA accepting dictionary D
We replace the root node of the suffix tree with the final state of the DFA.

16. Minimum DFA
The minimum DFA accepting D = S* # clearly requires constant space (for fixed S). S #

17. On-line Construction of Word Suffix Trees
T = aa#b#
a,b # a

18. On-line Construction of Word Suffix Trees
T = aa#b#
a,b # a

19. On-line Construction of Word Suffix Trees
T = aa#b#
a,b # a a

20. On-line Construction of Word Suffix Trees
T = aa#b#
a,b # a a

21. On-line Construction of Word Suffix Trees
T = aa#b#
a,b # a a #

22. On-line Construction of Word Suffix Trees
T = aa#b#
a,b # a a #

23. On-line Construction of Word Suffix Trees
T = aa#b#
a,b # b a a # b

24. On-line Construction of Word Suffix Trees
T = aa#b#
a,b # b a a # b

25. On-line Construction of Word Suffix Trees
T = aa#b#
a,b # b a # a # b #

26. Pseudo-Code
Just change here

27. Like Dress-up Doll...

28. Like Dress-up Doll... [cont.]a # b
a,b
S,#
Different looking!!!!
Just change hair!!
Body is the same!!

29. Compact Directed Acyclic Word Graphs
T = aa#b# a a # b b # # a b # #
minimization # b b # #
Compact Directed Acyclic
Word Graph (CDAWG)
Suffix Tree

30. Sparse CDAWGs T = aa#b# a # b b a # a # b # Word Suffix Tree
minimization #
Sparse Compact Directed Acyclic
Word Graph (SCDAWG)
Word Suffix Tree

31. Sparse CDAWGs [cont.] T = a#b#a#bab# a # b a # b b # a # a a b a # b #
minimization a # a a b a # b # # b b # a a b b # #
Word Suffix Tree
SCDAWG

32. SCDAWG Construction
SCDAWGs can be constructed by minimizing word suffix trees in O(k) time.
using Revuz’s DAG minimization algorithm (1992)

33. SCDAWG Construction [cont.]
Question : Direct construction for SCDAWGs?
Answer : YES! Using minimal DFA accepting dictionary D, we can directly build SCDAWGs in O(n) time and O(k) space.
We modify the CDAWG on-line construction algorithm (Inenaga et al. 05) by using the above DFA.

34. Pseudo-Code Different looking!!!! Just change here Body is the same!!# b
Different looking!!!!
a,b
S,#
Just change here
Body is the same!!

35. Some Events
Basically on-line construction of SCDAWGs is similar to that of word suffix trees.
Except for the two following unique events:
Edge merging
Node splitting

36. Edge Merging
a,b
T = a#b#a#bab#bc... # a # b b # # a a # # b b

37. Edge Merging
a,b
T = a#b#a#bab#bc... # a # b b # # a a # # b b a a

38. Edge Merging
a,b
T = a#b#a#bab#bc... # a # b b # # a a a # # b b a a

39. Edge Merging
a,b
T = a#b#a#bab#bc... # a # b b # a # a b a

40. Node Splitting
a,b
T = a#b#a#bab#bc... # a # b b # a a # b b a # b #

41. Node Splitting a,b T = a#b#a#bab#bc... # a # b b # a a # b b # a b b #

42. Node Splitting a,b T = a#b#a#bab#bc... # a # b b # # a a a # b # a b #

43. Conclusion
We introduced new text indexing structure sparse compact directed acyclic word graphs (SCDAWGs) for word-level pattern matching.
We presented an on-line algorithm to construct SCDAWGs directly, in O(n) time with O(k) space.
The key is the use of minimum DFA accepting dictionary D.

44. Related Work
“Sparse Directed Acyclic Word Graphs” by Shunsuke Inenaga and Masayuki Takeda Accepted to SPIRE’06