An Unsupervised Framework for Extracting and Normalizing Product Attributes from Multiple Web Sites

1
An Unsupervised Framework for Extracting and
Normalizing Product Attributes from Multiple Web
Sites

• Tak-Lam Wong
• Dept. of Computer Science and Engineering
• The Chinese University of Hong Kong
• Wai Lam, Tik-Shun Wong
• Dept. of Systems Engineering and Engineering
  Management
• The Chinese University of Hong Kong

_at_ SIGIR 2008 Singapore
2
Presentation Outline

• Introduction
• Problem Definition
• Our Model
• Inference Method
• Experimental Results
• Conclusions

3
Motivation
(Source http//www.crayeon3.com)
(Source http//www.superwarehouse.com)
4
Information Extraction

• To extract product attributes
• Prior knowledge about content
• Effective sensor resolution
• Layout format
• White balance, shutter speed
• Mutual influence
• Light sensitivity

5
Attribute Normalization

• Samples of extracted text fragments from a page
• Cloudy, daylight, etc
• What do they refer to?
• A text fragment extracted from another page
• white balance auto, daylight,cloudy, tungsten,
  
• Attribute normalization
• To cluster text fragments into the same group
• Better indexing for product search
• Easier understanding and interpretation

6
Existing Works

• Supervised wrapper induction
• They need training examples.
• The wrapper learned from a Web site cannot be
  applied to other sites.
• Template-independent extraction (Zhu et al.,
  2007)
• They cannot handle previously unseen attributes.
• Unsupervised wrapper learning (Crescenzi et al,
  2001)
• Extracted data are not normalized.

7
Contributions

• Unsupervised learning framework for jointly
  extracting and normalizing product attributes
  from multiple Web sites.
• Our framework considers page-independent content
  information and page-dependent layout
  information.
• Can extract unlimited number of product
  attributes (Dirichlet process)
• Can visualize the semantic meaning of each
  product attribute

8
Presentation Outline

• Introduction
• Problem Definition
• Our Model
• Inference Method
• Experimental Results
• Conclusions

9
Problem Definition (1)

• A product domain,
• E.g., Digital camera domain
• A set of reference attributes,
• E.g., resolution, white balance, etc.
• A special element, , representing
  not-an-attribute
• A collection of Web pages from any Web sites,
  , each of which contains a single product
• Let be any text fragment from a Web page

10
Problem Definition (2)
Line separator
ltTRgt ltTDgt ltPgt ltSPANgt White
balance lt/SPANgt lt/Pgt lt/TDgt ltTDgt
ltPgt ltSPANgt Auto, daylight, cloudy,
tungstem, fluorescent, fluorescent H,
custom lt/SPANgt lt/Pgt lt/TDgt lt/TRgt ltTRgt
Line separator
11
Problem Definition (3)

• Information extraction
• Attribute normalization
• Joint attribute extraction and normalization

Attribute information
Target information
Layout information
Content information
12
Problem Definition (4)

• White balance Auto, daylight, cloudy, tungstem,
  fluorescent, fluorescent H, custom
• T1
• Awhite balance
• Cloudy, daylight
• T1
• Awhite balance
• View larger image
• T0
• Anot-an-attribute

13
Presentation Outline

• Introduction
• Problem Definition
• Our Model
• Inference Method
• Experimental Results
• Conclusions

14
Our Model
Dirichlet Process Prior(Infinite Mixture Model)
S Different Web Site
N Text Fragment
k-th component proportion
Content info. generation
Target info. generation
15
Generation Process
16
Generation Process

• The joint probability for generating a particular
  text fragment given the parameters, , ,
  , and,
• Inference
• Intractable

17
Presentation Outline

• Introduction
• Problem Definition
• Our Model
• Inference Method
• Experimental Results
• Conclusions

18
Variational Method (1)

• Finding is intractable
• Our goal Design a tractable distribution
  such that should be as close to
  as possible.
• KL divergence

19
Variational Method (2)

• Truncated stick-breaking process (Ishwaran and
  James, 2001)
• Replace infinity with a truncation level K
• Max

20
Variational Method (3)

• One important variational parameters
• How likely does come from the k-th
  component?
• Attribute normalization!

21
Variational Method (4)

• Another important variational parameter
• where
• How likely should be extracted?
• Attribute extraction!

22
Unsupervised Approach

• What should be extracted?
• Make use of the prior knowledge about a domain.
• Only a few terms about the product attributes
• E.g., resolution, light sensitivity, shutter
  speed, etc.
• Can be easily obtained, for example, by just
  highlighting the attributes of a Web page
• Initialization

23
Presentation Outline

• Introduction
• Problem Definition
• Our Model
• Inference Method
• Experimental Results
• Conclusions

24
Experiments

• We have conducted experiments on four different
  domains
• Digital camera 85 Web pages from 41 different
  sites
• MP3 player 96 Web pages from 62 different sites
• Camcorder 111 Web pages from 61 different sites
• Restaurant 29 Web pages from LA-Weekly
  Restaurant Guide
• In each domain, we conducted 10 runs of
  experiments.
• In each run, we randomly selected a Web page and
  use the attributes inside as prior knowledge.

25
Evaluation on Attribute Normalization

• Baseline approach
• Agglomerative clustering
• Edit distance between text fragments
• Evaluation metrics
• Pairwise recall (R)
• Pairwise precision (P)
• Pairwise F1-measure (F)

26
Results of Attribute Normalization
27
Visualize the Normalized Attributes

• The top five weighted terms in the ten largest
  normalized attributes in the digital camera
  domain

28
Evaluation on Attribute Extraction

• Surprisingly, in the restaurant domain, our
  framework achieves a performance (0.95
  F1-measure) which is comparable to the supervised
  method (Muslea et al. 2001)

29
Presentation Outline

• Introduction
• Problem Definition
• Our Model
• Inference Method
• Experimental Results
• Conclusions

30
Conclusions (1)

• We aim at simultaneously extracting and
  normalizing product attributes from Web pages
  collected from different sites.
• Our method considers page-independent content
  information and the page-dependent layout
  information.
• We have developed a graphical model, which
  employs Dirichlet process prior, to model the
  generation of text fragments in Web pages.

31
Conclusions (2)

• An unsupervised inference algorithm based on
  variational method is designed.
• We formally show that content and layout
  information can collaborate and improve both
  extraction and normalization performance under
  our model.
• Experiments on four different domains have been
  conducted to show the robustness and
  effectiveness of our approach.

32
Questions and Answers