learning to construct and reason with a large kb of extracted information

Learning to Construct and Reason with a Large KB of

Extracted InformationWilliam W. Cohen

Machine Learning Dept and Language Technology Dept

joint work with:

Tom Mitchell, Ni Lao, William Wang, Kathryn Rivard Mazaitis,Richard Wang, Frank Lin, Ni Lao, Estevam Hruschka, Jr., Burr

Settles, Partha Talukdar, Derry Wijaya, Edith Law, Justin Betteridge, Jayant Krishnamurthy, Bryan Kisiel, Andrew

Carlson, Weam Abu Zaki , Bhavana Dalvi, Malcolm Greaves, Lise Getoor, Jay Pujara, Hui Miao, …

Outline• Background: information extraction and NELL• Key ideas in NELL

– Coupled learning– Multi-view, multi-strategy learning

• Inference in NELL– Inference as another learning strategy

• Learning in graphs • Path Ranking Algorithm• ProPPR

– Promotion as inference

• Conclusions & summary

But first….some backstory

..and an unrelated project…

..called SimStudent…

SimStudent will learn rules to solve a problem step-by-step and guide a student through

how solve problems step-by-step

Quinlan’s FOIL

Summary of SimStudent• Possible for a human author (eg middle school teacher) to

build an ITS system– by building a GUI, then demonstrating problem solving and

having the system learn how from examples• The rules learned by SimStudent can be used to construct

a “student model” – with parameter tuning this can predict how well individual

students will learn– better than state-of-the-art in some cases!

• AI problem solving with a cognitively predictive model … and ILP is a key component!

Information Extraction• Goal:

– Extract facts about the world automatically by reading text

– IE systems are usually based on learning how to recognize facts in text

• .. and then (sometimes) aggregating the results• Latest-generation IE systems need not require large

amounts of training• … and IE does not necessarily require subtle analysis of

any particular piece of text

Never Ending Language Learning (NELL)

• NELL is a broad-coverage IE system– Simultaneously learning 500-600 concepts and relations

(person, celebrity, emotion, aquiredBy, locatedIn, capitalCityOf, ..)

– Starting point: containment/disjointness relations between concepts, types for relations, and O(10) examples per concept/relation

– Uses 500M web page corpus + live queries– Running (almost) continuously for over three years– Has learned over 50M beliefs, over 1M high-confidence ones

• about 85% of high-confidence beliefs are correct

Demo

• http://rtw.ml.cmu.edu/rtw/

http://rtw.ml.cmu.edu/rtw/

NELL Screenshots

More examples of what NELL knows

Bootstrapped SSL learning of lexical patterns

ParisPittsburgh

SeattleCupertino

mayor of arg1live in arg1

San FranciscoAustindenial

arg1 is home oftraits such as arg1

it’s underconstrained!!

anxietyselfishness

Berlin

Extract cities:

Given: four seed examples of the class “city”

NP1 NP2

Krzyzewski coaches the Blue Devils.

athleteteam

coachesTeam(c,t)

person

coach

sport

playsForTeam(a,t)

NP

Krzyzewski coaches the Blue Devils.

coach(NP)

hard (underconstrained)semi-supervised learning

problem

much easier (more constrained)semi-supervised learning problem

teamPlaysSport(t,s)

playsSport(a,s)

One Key to Accurate Semi-Supervised Learning

1. Easier to learn many interrelated tasks than one isolated task2. Also easier to learn using many different types of information

Ontology and

populated KB

the Web

CBL

text extraction patterns

SEAL

HTML extraction patterns

evidence integration

PRA

learned inference

rules

Morph

Morphologybased

extractor

Another key idea: use multiple types of information




• Background: Learning in graphs • Path Ranking Algorithm• ProPPR



proposal

CMU

NSF

graph

William

6/18/07

6/17/07

Sent To

Term In Subject

[email protected]

Background: Personal Info Management as Similarity Queries on a Graph

[SIGIR 2006, EMNLP 2008, TOIS 2010]

Einat Minkov, Univ Haifa

Learning about graph similarity• Personalized PageRank aka Random Walk with Restart:

– Similarity measure for nodes in a graph, analogous to TFIDF for text in a WHIRL database

– natural extension to PageRank– amenable to learning parameters of the walk (gradient

search, w/ various optimization metrics):• Toutanova, Manning & NG, ICML2004; Nie et al,

WWW2005; Xi et al, SIGIR 2005– or: reranking, etc– queries:Given type t* and node x, find y:T(y)=t* and y~xGiven type t* and nodes X, find y:T(y)=t* and y~X

Many tasks can be reduced to similarity queries

Person namePerson namedisambiguationdisambiguation

ThreadingThreading

Alias findingAlias finding

[ term “andy” file msgId ]

“person”

[ file msgId ]

“file”

What are the adjacent messages in this thread? A proxy for finding “more messages like this one”

What are the email-addresses of Jason ?... [ term Jason ]

“email-address”

Meeting Meeting attendees finderattendees finder

Which email-addresses (persons) should I notify about this meeting? [ meeting mtgId ]

“email-address”

Learning about graph similarity:the next generation

• Personalized PageRank aka Random Walk with Restart:– Given type t* and nodes X, find y:T(y)=t* and y~X

• Ni Lao’s thesis (2012): New, better learning methods– richer parameterization– faster PPR inference– structure learning

• Other tasks:– relation-finding in parsed text– information management for biologists– inference in large noisy knowledge bases

Lao: A learned random walk strategy is a weighted set of random-walk “experts”, each of which is a walk constrained by a path (i.e., sequence of relations)

6) approx. standard IR retrieval

1) papers co-cited with on-topic papers

7,8) papers cited during the past two years

12-13) papers published during the past two years

Recommending papers to cite in a paper being prepared

Another study:learning inference rules for a noisy KB(Lao, Cohen, Mitchell 2011)(Lao et al, 2012)

Synonyms of the query team

American

IsA

PlaysIn

AthletePlaysInLeagueHinesWard

SteelersAthletePlaysForTeam

NFL

TeamPlaysInLeague

?

isa-1

Random walk interpretation is crucial

i.e. 10-15 extra points in MRR

Ontology and

populated KB

the Web

CBL


SEAL



PRA

learned inference

rules

Morph

Morphologybased

extractor

Another key idea: use multiple types of information

Outline

• Background: information extraction and NELL• Key ideas in NELL• Inference in NELL

– Inference as another learning strategy• Background: Learning in graphs • Path Ranking Algorithm• PRA + FOL: ProPPR and joint learning for inference



How can you extend PRA to

• Non-binary predicates?• Paths that include constants?• Recursive rules?• …. ?

• Current direction: using ideas from PRA in a general first-order logic: ProPPR

athletePlaySportViaRule(Athlete,Sport) onTeamViaKB(Athlete,Team), teamPlaysSportViaKB(Team,Sport)

teamPlaysSportViaRule(Team,Sport) memberOfViaKB(Team,Conference), hasMemberViaKB(Conference,Team2),playsViaKB(Team2,Sport).

teamPlaysSportViaRule(Team,Sport) onTeamViaKB(Athlete,Team), athletePlaysSportViaKB(Athlete,Sport)

A limitation• Paths are learned separately for each relation

type, and one learned rule can’t call another• PRA can learn this….

A limitation• Paths are learned separately for each relation

type, and one learned rule can’t call another• But PRA can’t learn this…..

athletePlaySport(Athlete,Sport) onTeam(Athlete,Team), teamPlaysSport(Team,Sport)

athletePlaySport(Athlete,Sport) athletePlaySportViaKB(Athlete,Sport)

teamPlaysSport(Team,Sport) memberOf(Team,Conference), hasMember(Conference,Team2),plays(Team2,Sport).

teamPlaysSport(Team,Sport) onTeam(Athlete,Team), athletePlaysSport(Athlete,Sport)

teamPlaysSport(Team,Sport) teamPlaysSportViaKB(Team,Sport)

Solution: a major extension from PRA to include large subset of Prolog






Sample ProPPR program….

Horn rules features of rules(vars from head ok)

.. and search space…

D’oh! This is a graph!

• Score for a query soln (e.g., “Z=sport” for “about(a,Z)”) depends on probability of reaching a ☐ node*• learn transition probabilities based on features of the rules• implicit “reset” transitions with (p≥α) back to query node

• Looking for answers supported by many short proofs

“Grounding” size is O(1/αε) … ie independent of DB size fast approx incremental inference (Reid,Lang,Chung, 08)

Learning: supervised variant of personalized PageRank (Backstrom & Leskovic, 2011)

*Exactly as in Stochastic Logic Programs[Cussens, 2001]

Sample Task: Citation Matching• Task:

• citation matching (Alchemy: Poon & Domingos).• Dataset:

• CORA dataset, 1295 citations of 132 distinct papers.• Training set: section 1-4.• Test set: section 5.• ProPPR program:

• translated from corresponding Markov logic network (dropping non-Horn clauses)

• # of rules: 21.

Task: Citation Matching

Time: Citation Matchingvs Alchemy

“Grounding” is independent of DB size

Accuracy: Citation Matching

AUC scores: 0.0=low, 1.0=hiw=1 is before learning

UW rules

Our rules

It gets better…..• Learning uses many example queries

• e.g: sameCitation(c120,X) with X=c123+, X=c124-, …

• Each query is grounded to a separate small graph (for its proof)

• Goal is to tune weights on these edge features to optimize RWR on the query-graphs.

• Can do SGD and run RWR separately on each query-graph

• Graphs do share edge features, so there’s some synchronization needed

Learning can be parallelized by splitting on the separate “groundings” of each query

Ontology and

populated KB

the Web

CBL


SEAL



PRA

learned inference

rules

Morph

Morphologybased

extractor

Back to NELL……

Experiment:•Take top K paths for each predicate learned by Lao’s PRA

• (I don’t know how to do structure learning for ProPPR yet)•Convert to a mutually recursive ProPPR program•Train weights on entire program






More details

• Train on NELL’s KB as of iteration 713• Test on new facts from later iterations• Try three “subdomains” of NELL

– pick a seed entity S– pick top M entities nodes in a (simple untyped

RWR) from S– project KB to just these M entities– look at three subdomains, six values of M





– Promotion as inference• Conclusions & summary

More detail on NELL

• For iteration i=1,….,715,…:– For each view (lexical patterns, …, PRA):

• Distantly-train for that view using KBi

• Propose new “candidate beliefs” based on the learned view-specific classifier

– Hueristically find the “best” candidate beliefs and “promote” them into KBi+1

Not obvious how to promote in a principled way …

Promotion: identifying new correct extractions from a pool of noisy extractions

• Many types of noise are possible:• co-referent entities• missing or spurious labels• missing or spurious relations• violations of ontology (e.g., an athlete that is not a person)

• Identifying true extractions requires joint reasoning, e.g.• Pooling information about co-referent entities• Enforcing mutual exclusion of labels and relations

Problem: How can we integrate extractions from multiple sources in the presence of ontological constraints at the scale of millions of extractions?

An example

Ontology:Dom(hasCapital, country)Mut(country, bird)

Sample Extractions:Lbl(Kyrgyzstan, bird)Lbl(Kyrgyzstan, country)Lbl(Kyrgyz Republic, country)Rel(Kyrgyz Republic, Bishkek,

hasCapital)Entity Resolution:SameEnt(Kyrgyz Republic,

Kyrgyzstan)

country

Kyrgyzstan Kyrgyz Republic

bird

Bishkek

SameEnt

Dom

Mut

Lbl

Rel(hasCapital)

Lbl Lbl

Kyrgyzstan

Kyrgyz RepublicBishkekcountry

Rel(hasCapital)Lbl

A knowledge graph view of NELL’s extractions

What you want

Knowledge graph

country

Kyrgyzstan Kyrgyz Republic

bird

Bishkek

SameEnt

Dom

Mut

Lbl

Rel(hasCapital)

Lbl Lbl

Representation as a noisy knowledge graph

Kyrgyzstan

Kyrgyz RepublicBishkekcountry

Rel(hasCapital)LblAfter Knowledge Graph Identification

graph identification

Lise Getoor, Jay Pujara, and Hui Miao @ UMD

Graph Identification as Joint Reasoning: Probabilistic Soft Logic (PSL)

• Templating language for hinge-loss MRFs, much more scalable!• Model specified as a collection of logical formulas

– Formulas are ground by substituting literal values– Truth values of atoms relaxed to [0,1] interval– Truth values of formulas derived from Lukasiewicz t-norm

• Each ground rule, r, has a weighted potential, ϕr corresponding to a distance to satisfaction

• PSL defines a probability distribution over atom truth value assignments, I:

• Most probable explanation (MPE) inference is convex• Running time scales linearly with grounded rules (|R|)

€

P(I) =1Zexp − wrr∈R∑ ?r(I)

p[ ]

PSL Representation of Heuristics for PromotionPromote any candidate

Promote “hints” (old promotion strategy)

Be consistent about labels for duplicate entities

PSL Representation of Ontological Rules

Adapted from Jiang et al., ICDM 2012

Be consistent with constraints from ontology

Too expressive for ProPPR

Datasets & Results• Evaluation on NELL dataset from iteration 165:

• 1.7M candidate facts • 70K ontological constraints

• Predictions on 25K facts from a 2-hop neighborhood around test data

• Beats other methods, runs in just 10 seconds!

F1 AUCBaseline .828 .873NELL .673 .765MLN (Jiang, 12) .836 .899KGI-PSL .853 .904

Summary• Background: information extraction and NELL• Key ideas in NELL






learning to construct and reason with a large kb of extracted information

Documents

learning strategylearning

language learning nellnell

cohen machine learning

nellcoupled learningmultiview

nellkey ideas

ni lao

graphs path

information extractiongoal