Chenliang Li, Aixin Sun @ SIGIR‘14

1. +
Fine-Grained Location Extraction from
Tweets with Temporal Awareness
2014/3/13 (Fri.)
Chang Wei-Yuan @ MakeLab Group Meeting
Chenliang Li, Aixin Sun
SIGIR‘14

2. +
Outline
n Introduction
n Problem Definition
n Method：PETAR
n POI inventory
n Data analysis and observations
n Time-aware POI tagger
n Experiment
n Conclusion
n Thought
2

3. +
Outline
n Introduction
n Problem Definition
n Method：PETAR
n POI inventory
n Data analysis and observations
n Time-aware POI tagger
n Experiment
n Conclusion
n Thought
3

4. +
Introduction
n Twitter has Large volume of timely data
n Social and business value.
n Users casually or implicitly reveal their
locations and short term visiting plans.
4

5. +
Outline
n Introduction
n Problem Definition
n Method：PETAR
n POI inventory
n Data analysis and observations
n Time-aware POI tagger
n Experiment
n Conclusion
n Thought
5

6. +
Problem Definition
n Given：
n a tweet t published by a user from a
predefined geographical region
n a set of temporal awareness label
n c ∈ { past, present, future, non-temporal }
n Goal：
n extract all POI and temporal awareness label
pairs from t.
n ( POI name, temporal awareness label )
6

7. +
Challenges
n Predominate usage of short names or
informal abbreviations
n Many POI names are ambiguous
7

8. +
Outline
n Introduction
n Problem Definition
n Method：PETAR
n POI inventory
n Data analysis and observations
n Time-aware POI tagger
n Experiment
n Conclusion
n Thought
8

9. +
Overview
9

10. +
POI inventory
n The POI inventory is constructed by
extracting the POI names mentioned
n Foursquare check-in from tweets.
10

11. +
POI inventory
n 239,499 Foursquare check-in tweets
made by Singaporean users
n POI names are extracted by applying
handcrafted rules, Regexp
11

12. +
POI inventory
n Partial POI names are extracted by
taking all the sub-sequences of the
names.
12

13. +
Data analysis and observations
n Data set
n 4.33M tweets from 19,256 unique Singapore-
based users during June 2010
n 222,201 tweets mentions at least one
candidate POI name from by 13,758 unique
users.
13

14. +
Data analysis and observations
n Observation 1
Many users reveal their fine-grained locations
in their tweets.
n 71.4% of all users in the dataset
n 91.3% of the users who had published at
least 20 tweets
14

15. +
Data analysis and observations
n Observation 2
The candidate POI mentions are mostly very
short with one or two words. Many of the
mentions are partial location names.
n 46.7% of the candidate POI names are
unigrams
n 41.6%+ of the candidate POI names are
partial POI names
15

16. +
Data analysis and observations
n Observation 3
About half of the candidate POI mentions
indeed refer to locations and their associated
temporal awareness can be determined.
16

17. +
Data analysis and observations
n Observation 3
About half of the candidate POI mentions
indeed refer to locations and their associated
temporal awareness can be determined.
17

18. +
Data analysis and observations
n Observation 4
Among all POIs that were visited, or to be
visited, about 90% of the visits to these POIs
happen within a day.
18

19. +
Time-aware POI tagger
n Simultaneously disambiguate the
candidate POI mentions and resolve the
temporal awareness
19

20. +
Time-aware POI tagger
n Lexical features
n Basic: word itself, lowercase form, all-cap …
n Contextual: context window, preceding,
following
n Off to jp now ! Hope it DOESN’t rain
n Off to jp now ! Hope it DOESN’t rain
n Off to jp now ! Hope it DOESN’t rain
20

21. +
Time-aware POI tagger
n Grammatical Features
n Part-of-speech (POS)
n The time-trend score T(t) of the tweet
n The closest verb/time-trend word
n its time-trend score, distance, left/right-hand side
indicator
21

22. +
Time-aware POI tagger
n a dictionary D of time-trend words,
containing commonly used words with
manually assigned time-trend scores
n D = { “future”:1, “present”:0, “past”:-1, “is”:0,
“was”:-1 … }
n a tweet t: “Soccer fever at mac now!”
n Is w ∈ t matching an entry in D?
n Considering tags of Verbs.
22

23. +
Time-aware POI tagger
n Geographical Features
n Spatial randomness.
n Location name confidence.
n Multiple candidate POI mention.
23

24. +
Time-aware POI tagger
n Geographical Features
n Spatial randomness ： POI name entorpy
n Location name confidence.
n Multiple candidate POI mention.
24

25. +
Time-aware POI tagger
n Geographical Features
n Spatial randomness.
n Location name confidence.
n Multiple candidate POI mention.
25

26. +
Time-aware POI tagger
n Geographical Features
n Spatial randomness. – entorpy/identifiability
n Location name confidence. – count
n Multiple candidate POI mention.
26

27. +
Time-aware POI tagger
n BILOU Schema Features
n Identify Beginning, Inside and Last word of a
multi-word POI name, and Unit-length POI
name, and the words Outside of any POI
names
27

28. +
Outline
n Introduction
n Problem Definition
n Method：PETAR
n POI inventory
n Data analysis and observations
n Time-aware POI tagger
n Experiment
n Conclusion
n Thought
28

29. +
Experiment
n Experiment Setup
n manually annotated 4,000 tweets as ground truth
n Comparative Methods
n Random Annotation (RA)
n K-Nearest Neighbor (KNN)
n StanfordNER (CRF-Classifier)
29

30. +
Experiment: results
n POI extraction with temporal awareness
n Disambiguating POIs (ignoring temporal
awareness)
30

31. +
Experiment: feature analysis
n Lex features are better for POI mention
disambiguation
n Gra features are better for resolving
temporal awareness
31

32. +
Experiment
n Lex + Gra features are better in most
cases
32

33. +
Outline
n Introduction
n Problem Definition
n Method：PETAR
n POI inventory
n Data analysis and observations
n Time-aware POI tagger
n Experiment
n Conclusion
n Thought
33

34. +
Conclusion
n PETAR facilitate the fine-grained location-
based services/marketing and
personalization.
n a POI inventory exploits the crowd wisdom
of Foursquare community to enable fine-
grained location extraction.
n a time-aware POI tagger conducts the
location extraction and temporal
awareness resolution in an effective and
efficient way.
34

35. +
Outline
n Introduction
n Problem Definition
n Method：PETAR
n POI inventory
n Data analysis and observations
n Time-aware POI tagger
n Experiment
n Conclusion
n Thought
35

36. +
Thought 1
n Location Extraction with Temporal
Awareness
n Activities Extraction with Temporal
Awareness
36

37. +
Thought 2
37

38. +
Thanks for listening.
2014 / 3 / 13 (Fri.) @ MakeLab Group Meeting
v123582@gmail.com