finalpresentation

Information about finalpresentation

Published on November 16, 2007

Author: Simo

Source: authorstream.com

Content

Collecting Labeled Linguistic Data through Online Games:  Collecting Labeled Linguistic Data through Online Games William Choi, Jason Turner-Maier, David Vickrey, Daphne Koller Motivation:  Motivation Commonly in Natural Language Processing, text is labeled with important features, such as parts of speech, sentence breaks, semantic roles, or related words. This data is used to train machine learning systems However, typically the amount of data that is labeled is small because it is expensive and time-consuming to obtain. Problem: Labeled linguistic data is useful for machine learning systems, but it is often expensive to get. Solution: Make the labeling task fun and easy by making online games. The Categorization Game:  The Categorization Game Examples of Category/Word Pairs Activities ----------------------------------------- Kayaking Things that are recovered from -------------- Grief Types of containers ---------------------------- Cups Things that are Turkish ------------------------ Baklava Things that are trusted ------------------------ Mentors Potential Uses Collect objects of a verb Collect attributes of a noun Enhance WordNet Identify noun-verb relationships The Categorization Game:  The Categorization Game Game Statistics:  Game Statistics 480 total games played 306 bot games, 174 human games Game Statistics:  Game Statistics Out of a sample of 100 randomly selected guesses, we have found that about 75~90% of them was of good quality. Examples of “bad” data: Misspelled guesses Things that are shaven ------ bears Things that listen ------------- hearing Enhancing WordNet Of the 1083 guess(noun)/category(noun) pairs, about 60.3% were not found in WordNet. Of those pairs that occurred two or more times, about 39.4% were not found in WordNet. Examples of collected pairs not in WordNet: Crime/Arson Diversion/Baking Protection/Guard Language/Mandarin, Language/Polish Music/Jam, Music Salsa Terrain/Plain, Terrain/Plateau, Terrain/Prairie Weather/Tornado The Free-Association Game:  The Free-Association Game Game Outline Given a word and a list of taboo words/phrases, two matched players type anything that is “associated” with the given word but does not appear on the taboo list. The goal is to match and move on to the next word as quickly as possible. This game collects a large web of related words and phrases. Possible Uses Useful for identifying implicit information from a sentence Can be extended to analyze multiple senses of the same word The Free-Association Game:  The Free-Association Game Game Statistics:  Game Statistics Around 95% of the word-association data was of good quality based on a sample. Slide10:  Type of – When one word is a “type of” the other, i.e. one is a superset of the other. Ex. Beef - Meat Synonyms – When the two words mean exactly or close to the same thing. Ex. Bounce - Jump Part of – When one word is a component or part of the other. Ex. Bicycle - Wheel Action – When one of the words is an action which involves the other. Ex. Hijack - Car Phrase – When the two words form a compound word or common phrase. Ex. Base - Ball Has – When one of the words has the other as a possession or an attribute, but that other word is not a “part of” the first word. Ex. Entry – Room Definition – When one word describes or defines the other word. Ex. Discard – Get rid of Word Part – When one word is related to the other by appearance but not necessarily by meaning. Ex. Work - Workforce Antonyms – Words that are opposites or are approximately so. Ex. Positive - Negative Misc –These don’t appear often enough to deserve their own category. Ex. Gravity - Apple This classification shows the kind of information that could be derived from the free-association data. A method for automatically classifying still needs to be developed. Game Statistics:  Game Statistics Definition – When one word describes or defines the other word. Ex. Discard – Get rid of Word Part – When one word is related to the other by appearance but not necessarily by meaning. Ex. Work - Workforce Antonyms – Words that are opposites or are approximately so. Ex. Positive - Negative Misc –These don’t appear often enough to deserve their own category. Ex. Gravity - Apple This classification shows the kind of information that could be derived from the free-association data. A method for automatically classifying still needs to be developed. The Sentence Game:  The Sentence Game Use the existing framework of the Free-Association Game to try to get facts that sentences entail. For example: Guide dogs are highly trained and very dependable, but occasionally make potentially dangerous mistakes. Guide dogs are highly trained. Guide dogs are dependable. Guide dogs can make mistakes. Guide dogs can be dangerous. Guide dogs are not always dependable. … Future Works:  Future Works Further develop the Sentence Game. Create other games for collecting different kinds of data: e.g. Telephone Advertise the games to get more users Improve the GUI, and come up with names for the games Enhanced Ranking/Titles System Differentiate between good and bad data automatically Put data in a more accessible form Classify the data to facilitate analysis

Related presentations


Other presentations created by Simo

MCC Presentation Kick Off v3
28. 09. 2007
0 views

MCC Presentation Kick Off v3

Ch 19
01. 10. 2007
0 views

Ch 19

eric tr03
05. 12. 2007
0 views

eric tr03

managing pain
12. 12. 2007
0 views

managing pain

LHC ATLAS Canada McPherson
27. 09. 2007
0 views

LHC ATLAS Canada McPherson

Cosplay
05. 11. 2007
0 views

Cosplay

Gerhard Haag TPA Pres
16. 11. 2007
0 views

Gerhard Haag TPA Pres

D1 10 Rad Olszewski
19. 11. 2007
0 views

D1 10 Rad Olszewski

Fire Safety Version 2
17. 12. 2007
0 views

Fire Safety Version 2

Stress Management Presentation
18. 12. 2007
0 views

Stress Management Presentation

virtualzoo 000
03. 01. 2008
0 views

virtualzoo 000

11 Grasses
04. 10. 2007
0 views

11 Grasses

Paquette MERLOT MAKES HISTORY
28. 12. 2007
0 views

Paquette MERLOT MAKES HISTORY

A THEORY OF HUMAN NEED lecture
07. 01. 2008
0 views

A THEORY OF HUMAN NEED lecture

NEOPOLIOS 2 final
15. 11. 2007
0 views

NEOPOLIOS 2 final

IROS2003
04. 01. 2008
0 views

IROS2003

copyright07
26. 02. 2008
0 views

copyright07

07 Paul Mullen
28. 02. 2008
0 views

07 Paul Mullen

vandergriffs future army
04. 03. 2008
0 views

vandergriffs future army

simile 3
11. 03. 2008
0 views

simile 3

munk economydemographyoil
18. 12. 2007
0 views

munk economydemographyoil

IS1
12. 03. 2008
0 views

IS1

globalspaces
14. 03. 2008
0 views

globalspaces

forum IRTG
18. 03. 2008
0 views

forum IRTG

D1L1
27. 03. 2008
0 views

D1L1

week 05
13. 04. 2008
0 views

week 05

haetc05
13. 12. 2007
0 views

haetc05

Affiliates
29. 10. 2007
0 views

Affiliates

psmp3 anna smrz
21. 11. 2007
0 views

psmp3 anna smrz

New Utility Marking Rule
04. 12. 2007
0 views

New Utility Marking Rule

Create a teaching webpage
05. 11. 2007
0 views

Create a teaching webpage

nomepoesia
21. 11. 2007
0 views

nomepoesia

Hedjazi presentation
19. 12. 2007
0 views

Hedjazi presentation

cre pp rcga rbca seminar
08. 11. 2007
0 views

cre pp rcga rbca seminar

erincook
03. 12. 2007
0 views

erincook

Lecture 110201
07. 11. 2007
0 views

Lecture 110201

ELI slideshow
23. 12. 2007
0 views

ELI slideshow

los meses
05. 11. 2007
0 views

los meses

2007 Laptop Handbook Slides
28. 11. 2007
0 views

2007 Laptop Handbook Slides