1. Mining Reference Tables for Automatic Text Segmentation E. Agichtein V
Mining Reference Tables for Automatic Text Segmentation E. Agichtein V. Ganti Columbia Univ. Microsoft R. KDD’04
Shui-Lung Chuang
Oct 27, 2004

2. Text Segmentation A (short)-text string N attributes
Conventional approaches
Rule-based — human creates rules
Supervised model-based — human labels data
Mining Ref. Table for Auto Text Segmentation E. Agichtein, V. Ganti, SIGKDD
Null
[ Authors , Title , Conference , Year ]

3. The Approach Utilize the existing (large, clean) reference data
E.g, DBLP  Papers, US Addresses, …
Author
Title
Conference
Year
Mark Steyvers, Padhraic Smyth
Probabilistic Author-Topic Models for
SIGKDD
2004
Lotlikar, S. Roy
A Hierarchical Document Clustering
WWW
Cimiano, S. Handschuh
Towards the Self-Annotating Web …
2003 ……
……. ….
ARM1
ARM2
ARM3
ARM3
s: a sub-string
prob. s is generated
ARM: Attribute Recognition Model

4. Segmentation Model
Mining Ref. Table for Auto Text Segmentation E. Agichtein, V. Ganti, SIGKDD
To find s1 s2 s3 s4
ARM1
ARM2
ARM3
ARM3
s: a sub-string
prob. s is generated
ARM: Attribute Recognition Model

5. Challenges Robust to input error Adaptive to varied attribute orders
The ref. data may be clean, but
Input may contain various errors:
Missing values, spelling error, extraneous or unknown tokens, etc
Adaptive to varied attribute orders
Reference data don’t contain info for attribute order in input
Efficient in training
Reference data is large
Engineer features
Adjust model topology
Determine attribute order from early input strings
Fix model topology
Don’t use advanced learning (e.g., EM)

6. Feature Hierarchy High-level features considered:
Token classes (words, numbers, mixed, delimiters) + Token length

7. Attribute Recognition Model
57th n sixth st
s fifth st
n goodwin ave

8. Model Training … … 57th n sixth st 1010 s fifth st 201 n goodwin ave
Transition:
B  { M, T, END }
M  { M, T, END }
T  { T, END }
Emission:
p(x|e)=(x=e) ? 1 : 0
Mixed
[a-z0-9]{1,-}
… …
[a-z0-9]{1,5}
[a-z0-9]{1,4}
57th

9. Sequential Specificity Relaxation
Token insertion
e.g., 57th 57th n sixth st
Token deletion
e.g., n sixth
Missing attribute value
e.g., <null>

10. Determining Attribute Value Order
Attribute order is usually preserved in the same batch of input strings

11. Determining Attribute Value Order
s = walmart s. randall ave madison wi.
pos v(s,Ai):
[ 0.05, , 0.02, 0.1, , 0.8, , ]  city attr.
[ 0.1, , 0.8, 0.7, 0.9, , , ]  street attr.
(partial order)
(total order)
Search all permutation for the best total order

12. Experiment Data Reference relations
Addresses: 1,000,000 tuples
Schema; [ Name,Number1,Number2,Address, City, State, Zip ]
Media: 280,000 music tracks
Schema: [ ArtistName, AlbumName, TrackName ]
Bibliography: 100,000 records from DBLP
Schema: [ Title, Author, Journal, Volume, Month, Year ]
Test datasets – Naturally concatenated test sets
Addresses: from RISE repository
Media: from Microsoft
Papers: 100 most cited papers from Citeseer

13. Experiment Data (cont.)
Test datasets – Controlled test data sets
Randomly chosen order
Error injection

14. Experiment Results

15. Experiment Results
1-Pos vs BMT vs BMT-robust

16. Comments The idea of using reference tables is good
The approach is well engineered to deal with issues of robustness and efficiency
Experiment is thorough
The approach is somewhat still ad hoc, and every component seems replaceable