Geolocating Blogs From Their Textual Content

1
Geolocating Blogs From Their Textual Content

• Clay Fink, Christine Piatko, James Mayfield,
  Danielle Chou (APL), Tim Finin, Justin Martineau
  (UMBC)

2
Motivation

• Where are people blogging?
• Why do we want to know?
• Support for analyzing the geographic dimension of
  sentiment
• To extract locations for geo-spatial reasoning
• Indentify sources of hyperlocal content
• Useful for mashups

3
Geolocation Issues For Blogs

• IP of blogger is not usually available for crawls
  and screen scraping
• Few blogs are self-hosted
• Most blogs are hosted by services such as
  Blogger, Wordpress, etc.
• From our crawl of 800,000 blogs (weblogs.com,
  2008)
• Only 3 with unique IPs
• 82 were hosted on IPs with at least 100 other
  crawled blogs

4
Metadata

• Metadata tags such as ICBM and geo.position are
  not used widely in blogs
• Found 900 blogs with such tags in 800,000 crawled
  blogs

ltmeta name'ICBM content'38.9906657,
-77.0260880' /gt
ltmeta name"geo.position" content"
'38.9906657-77.0260880 "gt ltmeta
name"geo.placename" content Silver Spring,
MDgt ltmeta name"geo.region" content US"gt
5
About Me And Profiles

• No consistent way across blogging platforms for
  expressing the bloggers location

6
Blogging Behavior and Location

• People can indicate their position from their
  linguistic behavior
• 54 of bloggers worldwide write about personal
  experience
• Can we leverage the presence of named location
  entities in text to determine a bloggers
  location?

7
Textual Clues In Blog Posts
8
Geolocating Blogs From Textual Clues

• Smith, et al (2001) and Amitay, et al (2004)
  describe how to extract the geographic focus of
  digital documents in just such a way
• Extract location entities
• Extract toponyms for entities from gazetteer
• Disambiguate location entities to a particular
  toponym
• Determine focus from clustering of disambiguated
  places
• Apply to blog posts
• Can the geographic focus of a blog can be derived
  the same way by accumulating disambiguated entity
  mentions across all posts?
• We tested how well the geographic focus extracted
  from a blog matches a bloggers location based on
  a collection of ground truth blogs with known
  locations

9
Toponyms From Post Location Mentions

• One US blog with known location
• 294 posts processed
• 211 unique entities
• 3429 toponyms

10
Disambiguated Post Location Mentions

• Disambiguated to 67 locations

11
Predicted/Actual Location of Blog
Predicted Washington County, OR
Actual Beaverton, OR

• Predicted location is Washington County, Oregon
• Subsumes true location of Beaverton, Oregon

12
Method

• We collected posts from blogs with know locations
  as our ground truth
• Our crawl looked for blogs with geolocating meta
  tags (ICBM, geo.position) in the home page
• Also used a list of blogs with reported locations
  from feedmap.net
• Used a named entity recognizer to extract
  location entities from posts
• Used an online gazetteer to get toponyms for the
  location entities
• Disambiguated locations
• Determined geographic focus by looking for
  clustering of disambiguated locations in a
  particular locale

13
Method

• NER
• Used the Stanford Named Entity Recognizer
• Used the Geonames online gazetteer
• Disambiguation
• Assume continent, country, or first-level
  administrative area for matching location names
• Look for qualified location names
• Vancouver, BC
• Disambiguate remaining mentions to most populous
  toponym
• Assume one sense per discourse across a given
  blogs posts

14
Method

• Determining Geographic Focus
• Populate a tree of the disambiguated places using
  the geographic inclusion hierarchy from the
  gazetteer
• Implemented tree using simple OWL ontology that
  describes a location hierarchy
• Set initial scores for hierarchy nodes
• Apply scoring function (Amitay et al, 2004) to
  accumulate scores up hierarchy
• sn In ? s2i DL-1
• Nodes lower in hierarchy with more sub-nodes will
  be higher scoring
• Success if highest scoring node subsumes known
  location, or is within 100 miles, or is within
  the same state or province

15
Results

• Processed posts from 844 English language, US
  blogs published between January 1, 2005, and
  April 24, 2009
• 526 hits out of 829 acceptable results
• 63 accuracy based on acceptance criteria
• 97 correctly identified as in the US
• Used an initial node score or 0.5 and a decay
  constant of 0.8
• Success correlated with number of disambiguated
  locations near known location (0.5)
• 95 accuracy if number of nearby locations is gt
  10

16
Results Geolocated US Blogs
17
Discussion

• Issues
• NER model we used was not trained on blog text
• Disambiguation techniques were very simple
• No confidence measure for assigned geographic
  focus
• No comparison to human judgments
• Worth Further Study
• Integrate into sentiment analysis
• Investigate use of locative language for finding
  references to proximate locations
• Use to populate an ontology instance implementing
  more complex geospatial relationships
• Apply to other social media, such as Twitter