707.009 Foundations of Knowledge Management g g Participative - - PowerPoint PPT Presentation
Knowledge Management Institute 707.009 Foundations of Knowledge Management g g Participative Knowledge Acquisition Methods Markus Strohmaier Univ. Ass. / Assistant Professor Knowledge Management Institute Graz University of
Knowledge Management Institute
Markus Strohmaier
Knowledge Management Institute Graz University of Technology, Austria e-mail: markus.strohmaier@tugraz.at web: http://www.kmi.tugraz.at/staff/markus
1
Markus Strohmaier 2011
Knowledge Management Institute
P bl f C t i ti – Problems of Categorization
– WordNet CyC ConceptNet and others WordNet, CyC, ConceptNet and others
– Knowledge and Ontology Engineering g gy g g – Collaborative Knowledge Acquisition – Game-Based Knowledge Acquisition Systems Perspective
2
Markus Strohmaier 2011
Knowledge Management Institute
Homonyme: Mehrdeutige Benennungen (z B Bank) Homonyme: Mehrdeutige Benennungen (z.B. Bank) Homophone: Gleichlautende Benennungen (z.B. Mohr, Moor) Homographen: Gleiche Schreibweisen (z.B. Wach(-)s(-)tube) S M h B i h t h fü d lb B iff Synonyme: Mehrere Bezeichnungen stehen für denselben Begriff (Auto, PKW) Antonyme: Gegensätze (z.B. hart - weich) H /H Ab t kt / S ifi h B iff ( B F h / Hyper/Hyponyme: Abstraktere / Spezifischere Begriffe (z.B. Fahrzeug / PKW) Formale Begriffssysteme zielen oft darauf ab wenig Raum für Interpretation zu lassen! Interpretation zu lassen!
– Homonymzusätze (Qualifikatoren) (z.B. „Ring <Schmuckstück>, Ring <Mathematik>) – Korrekte Zuordnung von Begriffen und Benennungen oft erst aus dem Kontext g g g heraus interpretierbar!
3
Markus Strohmaier 2011
Knowledge Management Institute
4
Markus Strohmaier 2011
Knowledge Management Institute
Liu, H. & Singh, P. (2004) ConceptNet: A Practical Commonsense Reasoning Toolkit. BT Technology Journal,. Volume 22, Kluwer Academic Publishers.
5
Markus Strohmaier 2011
Knowledge Management Institute
„Reale Welt“
Objekt W ort
Sem iotisches Dreieck
Ausdruck Sym bol Begriff Konzept
Sprache Wissen
Dolphin
Sprache Wissen
8
Markus Strohmaier 2011
Knowledge Management Institute
Example: Example:
Granulität Klarheit Vollständigkeit Wiederverwendbarkeit Modularität Konsistenz Redundanz Redundanz Verbreitung Zugänglichkeit Zugangsform
9
Markus Strohmaier 2011
Knowledge Management Institute
Anforderungsspezifikation Anforderungsspezifikation Wissenssammlung, Aufbau von Domänen-Wissen K tid tifik ti Konzeptidentifikation Informelle Repräsentation Entwickung einer ersten Taxonomie Konzeptualisierung und Formalisierung Integration evtl. vorhandener Integration evtl. vorhandener Ontologien Evaluation und Dokumentation Instandhaltung Instandhaltung Weiterentwicklung, Iteration
Grafik in Anlehnung an Studer
10
Markus Strohmaier 2011
Knowledge Management Institute
Automatic Acquisition of Hyponyms from Large Text Corpora 1992
cereals: rice* wheat* countries: Cuba Vietnam France* hydrocarbon: ethylene hydrocarbon: ethylene substances: bromine* hydrogen* protozoa: paramecium liqueurs: anisette* absinthe* q rocks: granite* substances: phosphorus* nitrogen* species: steatornis oilbirds bivalves: scallop* bivalves: scallop* fungi: smuts* rusts* fabrics: acrylics* nylon* silk* antibiotics: ampicillin erythromycin* p y y institutions: temples king seabirds: penguins albatross* flatworms: tapeworms planaria amphibians: frogs*
11
Markus Strohmaier 2011
amphibians: frogs*
Knowledge Management Institute
Automatic Acquisition of Hyponyms from Large Text Corpora 1992
(S1) The bow lute, such as the Bambara ndang, is plucked and has an individual curved neck for each string curved neck for each string.
12
Markus Strohmaier 2011
Knowledge Management Institute
Automatic Acquisition of Hyponyms from Large Text Corpora 1992
Process for identifying further patterns: 1 Decide on a lexical relation R that is of interest e g \group/member" (in our
formulation this is a subset of the hyponymy relation).
country". This list can be found automatically using the method described here, y y g , bootstrapping from patterns found by hand, or by bootstrapping from an existing lexicon
another and record the environment another and record the environment.
p p y , g the target relation and go to Step 2.
13
Markus Strohmaier 2011
Knowledge Management Institute
http://commonsense.media.mit.edu/ C id t Cyc uses paid experts to enter facts in CycL – a proprietary language to represent knowledge ConceptNet leverages ConceptNet leverages
Two types of input:
i iti acquisition
(restricted in length)
14
Markus Strohmaier 2011
( g )
Knowledge Management Institute
http://commonsense.media.mit.edu/
Types of relations:
15
Markus Strohmaier 2011
Knowledge Management Institute
http://commonsense.media.mit.edu/
Liu, H. & Singh, P. (2004) ConceptNet: A Practical Commonsense Reasoning Toolkit. BT Technology Journal,. Volume 22, Kluwer Academic Publishers.
Building ConceptNet Building ConceptNet 1. Extraction phase
50 extraction rules in regular expression form expression form Syntactic and semantic constraints are enforced
2 Normalisation phase
Spelling correction, lemmatization (replacing terms with their base form), removal of determiners ( the“ a“) Inferring assertions: removal of determiners („the , „a )
3. Relaxation phase
Improving the connectivity of the
e
adding frequency metadata, heuristics, Utilization of WordNet‘s and FrameNet‘s synsets and class- hi hi
16
Markus Strohmaier 2011
hierarchies
Knowledge Management Institute
http://commonsense.media.mit.edu/ DEMO: Openmind Common Sense htt // di it d / http://commonsense.media.mit.edu/ Example: Example: A car „is a kind of“ animal
17
Markus Strohmaier 2011
Knowledge Management Institute
http://commonsense.media.mit.edu/
Liu, H. & Singh, P. (2004) ConceptNet: A Practical Commonsense Reasoning Toolkit. BT Technology Journal,. Volume 22, Kluwer Academic Publishers.
Extraction Phase Each node is an english fragment composed of 4 syntactic constructions:
N h ( d l t t )
Adjectival phrases (very sour, red) Verbs must precede noun phrases and adj. Phrases, which in turn must precede prepositional phrases
If you want to own an expensive car then you should have lots of money or rich parents“ Illustration:
18
Markus Strohmaier 2011
„If you want to own an expensive car then you should have lots of money or rich parents
Knowledge Management Institute
http://commonsense.media.mit.edu/
Liu, H. & Singh, P. (2004) ConceptNet: A Practical Commonsense Reasoning Toolkit. BT Technology Journal,. Volume 22, Kluwer Academic Publishers.
Normalization Phase
Lemmatization:
Words are stripped of tense (is/are/were > be)
„If you want to own an expensive car then you should have earned lots of money or have rich parents“
Illustration:
19
Markus Strohmaier 2011
Knowledge Management Institute
http://commonsense.media.mit.edu/
Liu, H. & Singh, P. (2004) ConceptNet: A Practical Commonsense Reasoning Toolkit. BT Technology Journal,. Volume 22, Kluwer Academic Publishers.
Relaxation Phase Goal: Improve the connectivity of the
Apple IsA fruit
„Lifting“ knowledge by leveraging the IsA relationship
network Merge duplicate assertions Merge duplicate assertions Add additional metadata field „frequency“
OMCS corp s OMCS corpus.
during the relaxation phase
Produce „intermediate“ knowledge such as semantic and lexical generalisations Helps bridge other knowledge and to improve the
„Lifting“ knowledge by leveraging adjectival modifiers
20
Markus Strohmaier 2011
connectivity of the knowledgebase
Knowledge Management Institute
http://commonsense.media.mit.edu/
Liu, H. & Singh, P. (2004) ConceptNet: A Practical Commonsense Reasoning Toolkit. BT Technology Journal,. Volume 22, Kluwer Academic Publishers.
Relaxation Phase Goal: Improve the connectivity of the network Resolve vocabulary discrepancies and morphological variations (bike / bicycle) bicycle) Adding SuperThematicKline to reconcile action/state differences (relax/relaxation) or adjective/nominal differences (sad/sadness) Utilizing WordNet and FrameNet‘s verb synonym-sets and class- hierarchies hierarchies
21
Markus Strohmaier 2011
Knowledge Management Institute
http://commonsense.media.mit.edu/
Liu, H. & Singh, P. (2004) ConceptNet: A Practical Commonsense Reasoning Toolkit. BT Technology Journal,. Volume 22, Kluwer Academic Publishers.
f t th b f ti f t i
uttered in the OMCS corpus.
assertion was inferred during the relaxation phase
22
Markus Strohmaier 2011
Knowledge Management Institute Also see: Wired Magazin Article: “Inside the High Tech Hunt for a Missing Silicon Valley Legend” http://www.wired.com/techbiz/people/magazine/15-08/ff_jimgray?currentPage=all
Wi d N St F tt i i Wired News on Steve Fossett gone missing
http://www.wired.com/science/planetearth/news/2007/09/f
23
Markus Strohmaier 2011
Knowledge Management Institute
24
Markus Strohmaier 2011
Knowledge Management Institute
Demo: http://www.mturk.com
Examples: Powerset, Steve Fossett Criticism: minimum wage, exploitation of workers and exploitation of workers, and more
25
Markus Strohmaier 2011
Knowledge Management Institute
Chaplin's Modern Times, mins 04:00-05:30) (1936)
26
Markus Strohmaier 2011
Knowledge Management Institute
Now, Luis van Ahn, Ass. Prof. at CMU, asks:
This work would be fun? And people would love to do it? And people would love to do it? Even if they do not get paid?
Recommended lecture: http://video.google.com/videoplay?docid=-8246463980976635143
27
Markus Strohmaier 2011
p g g p y
Knowledge Management Institute
EXAMPLE EXAMPLE The Problem:
Observation: humans can easily identify complex concepts that consist of multiple parts in images, such as trees, bicycles, … Idea: human computation
28
Markus Strohmaier 2011
Knowledge Management Institute
Slide: Luis von Ahn
29
Markus Strohmaier 2011
Slide: Luis von Ahn
Knowledge Management Institute
Purse Addresses two bl t th Bag Handbag problems at the same time: 1) metadata ti d brown generation and 2) metadata validation
30
Markus Strohmaier 2011
Knowledge Management Institute
Slide: Luis von Ahn
31
Markus Strohmaier 2011
Slide: Luis von Ahn
Knowledge Management Institute
32
Markus Strohmaier 2011
Knowledge Management Institute
DEMO: Let‘s play the Google Image Labeling Game: Let s play the Google Image Labeling Game:
http://images.google.com/imagelabeler/
33
Markus Strohmaier 2011
Knowledge Management Institute
Luis von Ahn, Mihir Kedia and Manuel Blum, Verbosity: a game for collecting common-sense facts, CHI '06: Proceedings of the SIGCHI conference on Human Factors in computing systems, : 75--78, 2006.
Instead of asking users to enter true or false statements, or to rate statements, Verbosity leverages the fact that a game exists that requires users to state common-sense facts:TabooTM Example: Players have to describe the word „apple“ without saying „apple“ and without saying „red, pie, fruit, macintosh etc. without saying „red, pie, fruit, macintosh etc. The player has to give a good enough description of the word to get his teamates guess the right concept („you can make juice out of it“) The game requires players to say a list of common-sense facts about each word in order to get their teammates to guess it
34
Markus Strohmaier 2011
Knowledge Management Institute
Luis von Ahn, Mihir Kedia and Manuel Blum, Verbosity: a game for collecting common-sense facts, CHI '06: Proceedings of the SIGCHI conference on Human Factors in computing systems, : 75--78, 2006.
One of the players is chosen as the “Narrator” Fact: Laptop contains a keyboard One of the players is chosen as the Narrator while the other is the “Guesser.” The Narrator gets a secret word and must get the Guesser to type that word by sending hints to the Fact: Laptop contains a keyboard Guesser. The hints take the form of sentence templates with blanks to be filled in. The Narrator can fill in the blanks with any word they wish except the secret word (or any string containing the secret word) secret word). For example, if the word is LAPTOP, the Narrator might say: “it has a KEYBOARD.” Narrator might say: it has a KEYBOARD. The Guesser guesses. The Narrator can see all of these guesses, and can tell the Guesser
35
Markus Strohmaier 2011
g , whether each is „hot“ or „cold“.
Knowledge Management Institute
Luis von Ahn, Mihir Kedia and Manuel Blum, Verbosity: a game for collecting common-sense facts, CHI '06: Proceedings of the SIGCHI conference on Human Factors in computing systems, : 75--78, 2006.
Player can describe the secret by using sentence templates only. Reasons:
Template Examples: is a kind of
__ is a kind of__ __ is used for __ __ is typically near/in/on __ __ is the opposite of __
__ is related to __ __ (wildcard)
36
Markus Strohmaier 2011
Knowledge Management Institute
Google Techtalks, Luis von Ahn, 2006
http://video.google.com/videoplay?docid=-8246463980976635143
NARRATOR GUESSER
near CEREAL MILK
Is typically near CEREAL MILK! CEREAL
37
Markus Strohmaier 2011
Knowledge Management Institute
Google Techtalks, Luis von Ahn, 2006
http://video.google.com/videoplay?docid=-8246463980976635143
NARRATOR GUESSER
sense facts b t ilk MILK about milk MILK MILK
38
Markus Strohmaier 2011
Knowledge Management Institute
Luis von Ahn, Mihir Kedia and Manuel Blum, Verbosity: a game for collecting common-sense facts, CHI '06: Proceedings of the SIGCHI conference on Human Factors in computing systems, : 75--78, 2006.
Validation and Strategies for assuring accuracy of facts:
enters the correct word
indicator for the quality of the Narrator‘s statementss
39
Markus Strohmaier 2011
Knowledge Management Institute
Google Techtalks, Luis von Ahn, 2006
http://video.google.com/videoplay?docid=-8246463980976635143
NARRATOR GUESSER INPUT (SECRET) OUTPUT INPUT OUTPUT Asymmetric verification game
40
Markus Strohmaier 2011
Properties: Often fun, verified output
Knowledge Management Institute
Google Techtalks, Luis von Ahn, 2006
http://video.google.com/videoplay?docid=-8246463980976635143
Player 1 Player 2 INPUT INPUT OUTPUT2 OUTPUT1 Symmetric verification game
41
Markus Strohmaier 2011
If Output1 = Output2, both player get points
Knowledge Management Institute
Luis von Ahn, Mihir Kedia and Manuel Blum, Verbosity: a game for collecting common-sense facts, CHI '06: Proceedings of the SIGCHI conference on Human Factors in computing systems, : 75--78, 2006.
Symmetric vs. Asymmetric games Symmetric games: Symmetric games: Constraint is number of outputs per input
The ESP game If a given input has too many outputs a symmetric game
Asymmetric games:
Verbosity is not going to work, both players are never going to agree on the same output.
Asymmetric games: Constraint is number of inputs that yield the same output
near CEREAL If there are too many inputs that yield the same output, then given only the output you‘ll never guess what the
then given only the output you‘ll never guess what the input was.
42
Markus Strohmaier 2011
Knowledge Management Institute
43
Markus Strohmaier 2011
Knowledge Management Institute
44
Markus Strohmaier 2011
Knowledge Management Institute
45
Markus Strohmaier 2011
Knowledge Management Institute
Whoever has a CCTV camera, be it the police, local authorities or business or home
very soon.” Players are awarded one point for spotting a t d i d th i t if th suspected crime and three points if they see someone committing an actual crime They also lose points if the camera operator rules the alert was not a crime. Players who help catch the most criminals each month will win prizes up to £1,000.
Source: http://www.telegraph.co.uk/news/uknews/crime/6263882/ Snoopers-could-win-1000-prizes-for-monitoring-CCTV- cameras-on-the-internet.html
47
Markus Strohmaier 2011
Knowledge Management Institute
Researchers at Microsoft have invented a system for tagging images by reading brain scans from an electroencephalograph (EEG). Tagging images is an important task because Tagging images is an important task because many images on the web are unlabeled and have no semantic information. This new method allows an appropriate tag to be generated by an AI algorithm interpreting be generated by an AI algorithm interpreting the EEG scan of a person's brain while they view an image. The person need only view the image for as little as 500 ms. Other current methods for generating tags include current methods for generating tags include flat out paying people to do it manually, putting the task on Amazon Mechanical Turk,
Source: http://scitedaily.wordpress.com/2009/11/25/tag-images- with-your-mind/
48
Markus Strohmaier 2011
Knowledge Management Institute
49
Markus Strohmaier 2011