Taxonomy Generation from Documents

1
Taxonomy Generation from Documents

• ???
• ?????? ?????

2
Taxonomy ?

• Definition
• Represent a form of information classes
• As elucidated by Bailey (1994), taxonomy is an
  orderly classification of information according
  to presumed natural relationships. A vocabulary
  is the simplest form of taxonomy. The typical
  form of taxonomy is a hierarchy. At the top
  level, general terms or descriptive phrases are
  used. Each of the general terms has beneath it a
  set of terms that provide more refinement of the
  top-level term. Each of these second level terms
  may have a set of refining terms beneath it.
• Subject Taxonomy of Documents
• Organize documents according to subject classes
  of document content
• EX Web directory trees (Web portal or Web
  enterprise portal), or library classification
  schemes

3
Using Subject Taxonomy for Search

• Complement search engines and provide
  high-quality searches in a Web portal or an
  enterprise information portal (Sullivan02).
• Users will have an easier time finding
  information.
• users do not have to come up with keywords to
  find information.
• can categorize all content in a portal or
  intranet. This will improve the chance users will
  find what they are looking for in their searches.

4
Automatic Taxonomy Generation

• Taxonomy-generation tools create a Yahoo!-like
  directory structure for navigating content in a
  portal or intranet.
• Some tools, such as Verity's K2 Enterprise,
  Semio, Stratify, Quiver and SmartLogik provide
  automated methods for initial taxonomy
  construction
• These approaches based on hierarchical document
  clustering techniques.
• Rely on term extraction and document clustering
  techniques for analyzing the content similarity
  of the indexing documents.
• Similar documents can be grouped together and
  form a category.
• The subject terms of the category are extracted
  from that represent the key concepts of the
  documents..

5
Difficulties

• Suffer from some difficulties.
• Document clustering varies in accuracy
• Scalability
• Whether the generated categories are
  comprehensive and adaptable with users
  preferences and usages

6
Clustering of Search Result Pages

• Taxonomy Generation from Terms

7
Examples of Document Sources

• 45???????
• ??????

8
Examples of Documents Sources

• Multimedia Sources
• Family photos
• News videos

9
(No Transcript)
10
(No Transcript)
11
(No Transcript)
12
(No Transcript)
13
(No Transcript)
14
Related Techniques

• Extraction
• Word segmentation term extraction
• Word sense disambiguation
• Named entity extraction
• Key pattern extraction
• Semantic Categorization Clustering
• Term classification
• Relation Identification
• Relation extraction, synonym extraction, hyponym
  identification
• Taxonomy Generation
• Ontology Generation

15
Motivation Example I
Text segment Document Query term Search-result
snippets Clustering HAC Single-linkage Complete-li
nkage
Averageg-linkage K-means Cluster
naming Cohesiveness Isolation Precision Recall F1-
measure
Key terms

16
Motivation Example I
Text segment Document Query term Search-result
snippets Clustering HAC Single-linkage Complete-li
nkage
Averageg-linkage K-means Cluster
naming Cohesiveness Isolation Precision Recall F1-
measure
Key terms

Cohesiveness Isolation Precision Recall F1-measure
Text segment Document Query term Search-result
snippets
Clustering HAC K-Means Single-linkage Complete-lin
kage Min-max cut Cluster naming
Cohesiveness Isolation
Precision Recall F1-measure
Clustering HAC K-means
Single-linkage Complete-linkage Min-max cut
Cluster naming
17
Motivation Example II

• Organizing Queries/Questions

users

Call center
mail
email
phone
Web interface

Ex. questions

18
Motivation Example II

• Organizing Queries/Questions

users

Call center
mail
email
phone
Web interface
Ex. questions
19
The Intuition

• How does a human do when facing an unknown term?
• Send it to search engines
• Refer to the context it appears
• Infer its meaning(s)

Academia Sinica ... The top academic institution
in Taiwan. It is engaged in conducting scientific
research in its... Learn Chinese with Clavis
Sinica Chinese language reading and Clavis
Sinica Chinese language learning software
Chinese characters. Learn Chinese with Clavis
Sinica. ... Institute of Information Science
Academia Sinica Homepage Information about
research groups, papers, and projects
Sinica ?
20
Overview of the Approach
21
Experimental Results Structure
HAC
22
Experimental Results Structure
HACP
23
Data Representation

• Highly-ranked search-result snippets are used to
  enrich the representation of the text segments
• Huge pages have been indexed and continuously
  refreshed
• Features are extracted from neighbor contents
  (context)
• Vector space model is adopted to represent each
  text segment. The weight of each feature term is

24
Hierarchical Clustering Algorithm

• Hierarchical agglomerative clustering (HAC)
• Treat each input object as a singleton cluster
• Compute the similarities between all pairs of
  objects
• Merge the most similar two clusters
• Update the similarity matrix with new cluster
• Repeat step 3 and 4 until only a single cluster
  remains

25
In-cluster Similarity

• Inter-object similarity
• Inter-cluster similarity

26
Hierarchical Cluster Partitioning

• To get a more natural multi-way tree from the
  binary tree generated by HAC algorithm

27
Hierarchical Cluster Partitioning

• To get a more natural multi-way tree from the
  binary tree generated by HAC algorithm

28
Hierarchical Cluster Partitioning

• To get a more natural multi-way tree from the
  binary tree generated by HAC algorithm

Cut level l
C9
C9
1
C8
2
C7
C7
3
C6
C6
4
C1
C2
C3
C4
C5
C1
C2
C3
C4
C5
29
Hierarchical Cluster Partitioning

• To get a more natural multi-way tree from the
  binary tree generated by HAC algorithm

Cut level l
C9
C9
1
C8
2
C7
C7
3
C6
C6
4
C1
C2
C3
C4
C5
C1
C2
C3
C4
C5
30
Hierarchical Cluster Partitioning

• To get a more natural multi-way tree from the
  binary tree generated by HAC algorithm

Cut level l
C9
C9
1
C8
2
C7
3
C6
4
T1
T2
T3
T4
T5
C1
C2
C3
C4
C5
31
Cluster Quality Measure

• Min-max cut
• Minimize the inter-similarities
• Maximize the intra-similarities

32
Experiment Data

• Yahoo! CS directory
• 36, 177, 278 category names in 1st, 2nd, 3rd
  levels
• Yahoo! People/Scientists
• 250 famous people in 9 science fields
• Paper
• QuizNLQ
• 163 quiz question in 7 categories

33
Evaluation Metric F-measure
34
Experimental Results Accuracy

• Yahoo! CS category names with various
  inter-cluster similarities
• Average-linkage in various data sets

35
Experimental Results Structure
HAC
36
Experimental Results Structure
HACP
37
Experimental Results Structure

• Structure measure based on
• Depth of the hierarchy
• Total clusters
• Average number of child clusters

38
User Evaluation Comprehension

• Five qualitative measures
• Cohesiveness
• Isolation
• Hierarchy
• Navigation balance
• Readability

39
User Evaluation Usability

• Compare with the hierarchies constructed by human
• Group I construct hierarchy from scratch
• Group II use auto-generated hierarchy as
  reference

40
Conclusion

• Current contributions
• Study the feasibility of using search-result
  snippets and clustering technique to general
  text-segment taxonomy generation
• Conduct extensive experiment are performed on
  several typical data sets and the results are
  promising
• Future work
• Apply the approach on more types of text segments
• Investigate good cluster naming to enhance its
  usability
• Explore techniques to deal with longer text
  segments with less search-result snippets