�ݺ�ߣ

Integrating Product Data from
Websites offering Microdata
Markup
School of Business Informatics and Mathematics
Petar Petrovski, Volha Bryl, Christian Bizer
Data and Web Science Research Group
University of Mannheim, Germany

Outline
1. HTML-embedded Data on the Web
2. The Data Integration Pipeline
1. Microdata extraction
2. Classification
3. Feature extraction
4. Identity resolution
5. Data Fusion
3. Conclusions
2Integrating Product Data from Microdata Markup. Petar Petrovski, Volha Bryl, Chris Bizer

HTML-embedded Data
More and more Websites semantically markup the
content of their HTML pages.
Microformats
Microdata
RDFa

Schema.org
• ask site owners to embed
data to enrich search results.
• 200+ Classes: Product, Review, LocalBusiness, Person, Place, Event, …
• Encoding: Microdata or RDFa

Usage of Schema.org Data @ Google
Data snippets
within
search results
Data snippets
within
info boxes

Websites Containing Structured Data
(November 2013)
1.7 million websites (PLDs) out of 12.8 million
provide Microformat, Microdata or RDFa data (13%)
585 million of the 2.2 billion pages contain
Microformat, Microdata or RDFa data (26%).
http://webdatacommons.org/structureddata/
Google, October 2013:
15% of all websites provide structured data.

Top Classes, Microdata (2013)
• schema = Schema.org
• datavoc = Google‘s
Rich Snippet Vocabulary

Example: Microdata, Local Business

Example: Microdata, Product
School of Business Informatics and Mathematics

The Data Integration Pipeline
• Objective: integrate all data found on the web
describing a specific entity (e.g. product or organization)
• Motivation: enables creation of powerful applications, e.g.
comparison shopping portals
• Use case: product data
• Implemented Pipeline:

Outline
2. Classification
5. Data Fusion
3. Conclusions

Web Data Commons Extraction
Framework
• Web Data Commons project: extracts structured data from
the Common Crawl
– http://webdatacommons.org/
– http://commoncrawl.org/
• Code available at:
– https://subversion.assembla.com/svn/commondata/
– Based on Anything To Triples (any23) library for extracting
structured data: http://any23.apache.org
• Common Crawl 2012
– 3 billion HTML pages, 40.6 million websites
– 7.3 billion statements describing 1.15 billion things
– 9.4 million product offers from 9240 e-shops
Integrating Product Data from Microdata Markup. Petar Petrovski, Volha Bryl, Chris Bizer

Looking Deeper into E-Commerce Data
Microdata Product (2013)

Looking Deeper into E-Commerce Data
Microdata Product (2012)

Example: Title and Description
Title
Description
AppleMacBook Air MC968/A 11.6-Inch Laptop
Faster Flash Storage with 64 GB Solid State Drive and USB 3.0. 720p FaceTime HD
Camera. The new 1.6 GHz Intel Core i5 Processor with Intel HD Graphics 3000
enabling beautiful rendering and 4GB DDR3 RAM. 11.6” LED display with the best
resolution…
Title
Description
The MacBook Air MC 968/A powered by Intel Core i5(1.6GHz, 3MB L3). 64 GB SSD
and 4096 MB of DDR3 RAM. 29.464cm (11.6”) TFT 1366x768, Intel HD Graphics,
IEEE 802.11a/b/g, Bluetooth 4.0, FaceTme camera, OS X LIon
Apple MacBook Air 11-in, Intel Core i5 1.60GHz, 4
GB, 64 GB, Mac OS X Lion 10.7
Various abbreviations can be
found describing same features Often imprecise values due to rounding
in numeric values can be found
Different descriptions follow
different levels of detail
Integrating Product Data from Microdata Markup. Petar Petrovski, Volha Bryl, Chris Bizer

Outline
2. Classification
5. Data Fusion
3. Conclusions

Product Classification
• Starting from 9.4 million products:
• Products with English descriptions with length grater than 20 words
=> 1,986,359 products from 9,240 e-shops
• Training set
– 18,000 labeled products, 9 classes
• Training the model
– Naïve Bayes Classifier
• Features generation
– 4 step process – tokenizing and removing stop words, pruning,
n-grams, TF-IDF
– ~3600 features

Classification Performance
Category Precision % Recall % #
Books 86.58 87.95 233,249
Movies, Music & Games 89.81 70.63 186,832
Electronics & Computers 92.98 88.00 219,118
Home, Garden & Tools 73.81 60.78 186,495
Grocery, Health & Beauty 70.20 72.86 120,573
Toys, Kids, Baby & Pets 75.00 64.85 114,236
Clothing, Shoes & Jewelry 88.56 89.93 206,315
Sports & Outdoors 72.83 67.90 143,156
Automotive & Industrial 73.06 65.50 168,567
Average 80.31 74.26 1,578,541
The offers originate from 9,240 e-shops

Outline
2. Classification
5. Data Fusion
3. Conclusions

Product Feature Extraction
• Low precision (69%) for identity resolution without product feature
extraction
– Used later as a baseline for identity resolution
• We developed the Free Text Preprocessor
– Makes the data more structured by extracting new property-
value pairs from free-text properties
– https://www.assembla.com/spaces/silk/wiki/Silk_Free_Text_Preprocessor

Free Text Preprocessor by Example
<http://wdc.org/resource/2> <http://schema.org/Product/title> "Apple iPod nano (8 GB, 6th generation, Graphite)" .
<http://wdc.org/resource/2> <http://schema.org/Product/description>
"Memory size: 8GB. Colour: Graphite Generation: 6th generation. Memory type: Integrated. Weight: 21.1g.
Radio: With Radio. Audio/Video formats: AAC, AIFF, Audible, MP3, WAV, VBR Display: 1.5-inch" .

Free Text Preprocessor by Example
<http://wdc.org/resource/2> <http://schema.org/Product/Brand> "Apple" .
<http://wdc.org/resource/2> <http://schema.org/Product/Model> "iPod nano" .
<http://wdc.org/resource/2> <http://schema.org/Product/Storage> "8GB" .
<http://wdc.org/resource/2> <http://schema.org/Product/Display> "1.5-inch" .

Silk Free Text Preprocessor by Example
<http://wdc.org/resource/2> <http://schema.org/Product/Brand> "Apple" .
<http://wdc.org/resource/2> <http://schema.org/Product/Model> "iPod nano" .
<http://wdc.org/resource/2> <http://schema.org/Product/Storage> "8GB" .
<http://wdc.org/resource/2> <http://schema.org/Product/Display> "1.5-inch" .
Free Text Preprocessor
Specification

Extractors – Bag-of-words
• Learning
• Creating a list of words for every feature in the training set
• Extraction
• Matching tokens against the learned lists
• Pros
• Good for extracting nominal and numerical (with units of measurement) attributes
• Cons
• Bad for extracting multi-token values
• Inconclusive for values that refer to more than one feature
Brand
Storage
Display
Samsung Benq Apple Cannon …
64 GB megabytes 512GB …
42-inch 3.5-inches Inches 15.24cm …

Extractors – Feature-Value Pairs
Learns feature-value pairs from the structured data
Extraction
• Tagging – taking n-grams up to 4 and matching against the values from the training set
• Parsing – taking the combination of feature-value pairs that best describes an object
from the training dataset
• Pros
• Extracting multi-token values
Cons
• Inconclusive for values that refer to more than one feature
<Model, Asus EEE 10.1 Inch>
<Processor, 1.66 GHz Intel Atom N445>
<Display, 10.1-inches>
..
<Model, Panasonic Viera>
<Display, 42-Inch>

Extractors – Manual Configuration
Manually configure features and extraction methods
1. Regular expressions
• E.g. Processor - d*.?d+GHz
2. Dictionary search
• E.g. Dictionary of brands (Samsung, Panasonic, Lenovo, Apple)
• Pros
• Extraction process can be fine-tuned according to the data
• Good solution when no training (structured) data are available
• Cons
• Needs domain knowledge
• Non-trivial to efficiently pick extraction methods manually

Extraction Experiments
• Dataset for extraction 5,000 electronic
products from WDC
• Training dataset (structured data)
– 20 electronics products Amazon dataset

Extraction Accuracy
Brand Model Storage Display Processor Dimension
iPod Nano .92 .98 .86 .49 .12 .78
Galaxy SII .72 .87 .89 .81 .40 .91
GalaxyTab 7.7 .80 .92 .89 .85 .72 .93
Ixus 120IS 1 .96 N/A .89 N/A .56
Vaio VPC .99 .65 .81 .77 .73 .32
Viera 42 .95 .72 N/A .82 N/A .64
Sandisk 1 1 .85 N/A N/A .31
• Extraction using Combination configuration
(bag-of-words for Brand, Storage and Display;
feature-value pairs for Model and Dimension;
custom regular expression for the Processor)

Outline
2. Classification
5. Data Fusion
3. Conclusions

Identity Resolution
• We used Silk – a tool for discovering relationships
between data items within different linked data
sources
Provides a expressive language for defining linkage rules
Uses genetic programming to learn linkage rules
Has shown high performance on various datasets
https://www.assembla.com/spaces/silk/wiki/Home

Identity Resolution Experiments
• Gold standard: 5,000 links manually annotated
• 2,500 positive/2,500 negative
• 20 electronics products Amazon dataset (reference set)
• Experiment on 5 configurations
– Baseline (no feature extraction step)
– Bag-of-words
– Feature-value pairs
– Manual configuration
– Combinations

Silk Output: Learned Linkage Rule
:Property
wdc:Model
:Transform
lowerCase
:Comparison
func = Levensthein
threhold = 1.134
:Property
wdc:Display
:Aggregation
func= max
:Aggregation
func= average
:Transform
lowerCase
:Property
amazon:Model
:Transform
tokenize
:Transform
tokenize
:Property
amazon:Display
:Comparison
func = Jaccard
threhold = 0.23
:Comparison
func = Jaccard
threhold = 0.02
:Property
amazon:Storage
:Property
wdc:Storage

Identity Resolution Results
Precision % Recall % F-Measure %
Baseline 69 90 78.1
Bag-of-words 75 82 77.9
Feature-value pairs 80 77 78.4
Custom 82 80 80.9
Combination 85 80 82.4

Outline
2. Classification
5. Data Fusion
3. Conclusions

Data Fusion
• Input: clusters of products after identity resolution
• Properties worth fusing/combining
– AggregateRating and Review

Fusion Results
Product Offers Reviews Ratings
iPod Nano 8GB 829 84 0
iPhone 4 16GB 624 35 52
Sony Ericsson Xperia Mini 450 31 12
iPad 16GB 423 40 48
Motorola XOOM 32GB 270 12 0
Samsun Galaxy SII 142 8 0

Conclusions
• By using Microdata, thousands of websites help us to
understand their content
• We have implemented the 5-step data integration pipeline
– From Microdata markup to an integrated dataset
• A newly introduced feature extraction step is crucial for the
precision of data integration
– Identity resolution precision increases from 69% to 85%
• Future work
– Automatically learning regular expressions
– Automatically discovering combinations of extractors

Questions?

�ݺ�ߣ

Integrating Product Data from Websites offering Microdata Markup

More Related Content

Integrating Product Data from Websites offering Microdata Markup