Bayesian Logic Programs for Plan Recognition and Machine Reading

Bayesian Logic Programsfor

Plan Recognition and Machine Reading

Sindhu RaghavanAdvisor: Raymond Mooney

PhD Oral DefenseNov 29th, 2012

1

Outline• Motivation• Background

– Bayesian Logic Programs (BLPs)• Plan Recognition• Machine Reading

– BLPs for inferring implicit facts– Online Rule Learning– Scoring Rules using WordNet

• Future Work• Conclusions

2





3

Machine ReadingMachine reading involves the automatic extraction of knowledge from natural language text

4

Example“Barack Obama is the current President of the USA……. Obama was born on August 4, 1961, in Hawaii, USA…….”

Extracted factsnationState(usa)person(barackobama)isLedBy(usa,barackobama)hasBirthPlace(barackobama,usa)employs(usa, barackobama)

Data is relational in nature - several entities and several relations between them

Characteristics of Real World Data• Relational or structured data

– Several entities in the domain– Several relations between entities– Not always independent and identically

distributed (i.i.d)• Presence of noise or uncertainty

– Uncertainty in the types of entities– Uncertainty in the relations

Traditional approaches like first-order logic or probabilistic models can handle either structured data or uncertainty, but not both. 5

Statistical Relational Learning (SRL)

• Integrates first-order logic and probabilistic graphical models [Getoor and Taskar, 2007]

– Overcome limitations of traditional approaches

• SRL formalisms– Stochastic Logic Programs (SLPs) [Muggleton, 1996]

– Probabilistic Relational Models (PRMs) [Friedman et al., 1999]

– Bayesian Logic Programs (BLPs) [Kersting and De Raedt, 2001] – Markov Logic Networks (MLNs) [Richardson and Domingos, 2006]

6

Statistical Relational Learning (SRL)

• Integrates first-order logic and probabilistic graphical models [Getoor and Taskar, 2007]

– Overcome limitations of traditional approaches

• SRL formalisms– Stochastic Logic Programs (SLPs) [Muggleton, 1996]

– Probabilistic Relational Models (PRMs) [Friedman et al., 1999]

– Bayesian Logic Programs (BLPs) [Kersting and De Raedt, 2001]

– Markov Logic Networks (MLNs) [Richardson and Domingos, 2006]

7

Bayesian Logic Programs (BLPs)[Kersting and De Raedt, 2001]

• Integrate first-order logic and Bayesian networks

• Why BLPs?– Efficient grounding mechanism that includes only

those variables that are relevant to the query– Easy to extend by incorporating any type of logical

inference to construct networks– Well suited for capturing causal relations in data

8

Objectives

Plan Recognition

Machine Reading

9

Objectives

10

Plan recognition involves predicting the top-level plan of an agent based on its observed actions

Machine Reading

Objectives

11

Plan Recognition

Machine Reading involves automatic extraction of knowledge from natural language text

Common characteristics

• Inference and learning from partially observed or incomplete data

• Plan recognition– Top-level plan is not observed– Some of the executed actions can be unobserved

• Machine Reading– Information that is implicit is rarely observed in

data– Common sense knowledge is not always explicitly

stated12

Thesis Contributions

• Plan Recognition– Bayesian Abductive Logic Programs (BALPs) [ECML 2011]

• Machine Reading– BLPs for learning to infer implicit facts from natural

language text [ACL 2012]

– Online rule learner for learning common sense knowledge from natural language extractions [In Submission]

– Approach to scoring first-order rules (common sense knowledge) using WordNet [In Submission]

13

Thesis Contributions

• Plan Recognition– Bayesian Abductive Logic Programs (BALPs) [ECML 2011]

• Machine Reading– BLPs for learning to infer implicit facts from natural

language text [ACL 2012]

– Online rule learner for learning common sense knowledge from natural language extractions [In Submission]

– Approach to scoring first-order rules (common sense knowledge) using WordNet [In Submission]

14





15

Bayesian Logic Programs (BLPs)[Kersting and De Raedt, 2001]

• Set of Bayesian clauses a|a1,a2,....,an– Definite clauses that are universally quantified– Range-restricted, i.e variables{head} variables{body}– Associated conditional probability table (CPT)

• P(head|body)

• Bayesian predicates a, a1, a2, …, an have finite domains– Combining rule like noisy-or for mapping multiple CPTs into

a single CPT

• Given a set of Bayesian clauses and a query, SLD resolution is used to construct ground Bayesian networks for probabilistic inference

16

Probabilistic Inference and Learning

• Probabilistic inference– Marginal probability

• Exact Inference• Sample Search [Gogate and Dechter, 2007]

• Learning [Kersting and De Raedt, 2008]

– Parameters• Expectation Maximization• Gradient-ascent based learning

17





18

Plan Recognition

• Predict an agent’s top-level plan based on its observed actions• Abductive reasoning involving inference of

cause from effect

• Since SLD resolution used in BLPs is deductive in nature, BLPs cannot be used as is plan recognition

19

Extending BLPs for Plan Recognition

20

BLPs Logical Abduction

BALPs

BALPs – Bayesian Abductive Logic Programs

Extending BLPs for Plan Recognition

21

BLPsStickel’s

Abduction Algorithm

BALPs

BALPs – Bayesian Abductive Logic Programs

Experimental Evaluation

• Data• Monroe [Blaylock and Allen, 2005]• Linux [Blaylock and Allen, 2005]• Story Understanding [Ng and Mooney, 1992]

• Systems compared– BALPs– MLN-HCAM [Singla and Mooney, 2011]

– Blaylock and Allen’s system [Blaylock and Allen, 2005]– ACCEL-Simplicity [Ng and Mooney, 1992]– ACCEL-Coherence [Ng and Mooney, 1992]

22

Summary of Results• Monroe and Linux

– BALPs outperform both MLN-HCAM and the system by Blaylock and Allen

• Story Understanding– BALPS outperform both MLN-HCAM and ACCEL-

Simplicity– ACCEL-Coherence outperforms BALPs and other

systems• Specifically developed for text interpretation

• Automatic learning of model parameters using EM

23





24

Machine Reading

• Natural language text is typically “incomplete”– Some information is always implicit– Common sense information is not always explicitly

stated– Grice’s maxim of quantity [1975]

• Information extraction (IE) systems extract information that is explicitly stated [Cowie and

Lenhert, 1996; Sarawagi, 2008] – Cannot extract information that is implicit

25

ExampleNatural language text“Barack Obama is the President of the United States of America.”

Query“Barack Obama is a citizen of what country?”

IE systems cannot answer this query since citizenship information is not explicitly stated.

26

Objective• Infer implicit facts from explicitly stated

information– Extract explicitly stated facts using an off-the-shelf

IE system– Learn common sense knowledge in the form of

first-order rules to deduce additional facts– Use BLPs for inference of additional facts

27

Related Work• Logical deduction based approaches

– Learning propositional rules [Nahm and Mooney, 2000]– Purely logical deduction is brittle since it cannot assign

probabilities to inferences– Learning probabilistic first-order rules using FOIL and

FARMER [Carlson et al., 2010; Doppa et al., 2010]– Probabilities are not computed using well-founded

probabilistic graphical models • Use MLN based approaches for inferring additional

facts [Schoenmackers et al., 2010; Sorower et al., 2011]

– “Brute force” inference could result in intractably large networks for large domains

– Scaling of MLNs to large domains [Schoenmackers et al., 2010; Niu et al., 2012]

28

Objectives

• BLPs for learning to infer implicit facts from natural language text

• Online rule learner for learning common sense knowledge from natural language extractions

• Approach to scoring first-order common sense knowledge using WordNet

29





30

System ArchitectureTraining

DocumentsInformation Extractor

(IBM SIRE)Extracted

Facts

Rule learnerFirst-OrderLogical Rules

BLP Weight Learner

Bayesian LogicProgram (BLP)

BLP InferenceEngine

TestDocument

Extractions

Inferences withprobabilities 31

.

.

.

.

.

.

Barack Obama is the current President of USA……. Obama was born on August 4, 1961, in Hawaii, USA.

.

.

.

.

.

.

nationState(USA)Person(BarackObama)isLedBy(USA,BarackObama)hasBirthPlace(BarackObama,USA)hasCitizenship(BarackObama,USA)

nationState(B) ∧ isLedBy(B,A) hasCitizenship(A,B)nationState(B) ∧ employs(B,A) hasCitizenship(A,B)

hasCitizenship(A,B) | nationState(B) , isLedBy(B,A) .9hasCitizenship(A,B) | nationState(B) , employs(B,A) .6

nationState(malaysia)Person(mahathir-mohamad)isLedBy(malaysia,mahathir-mohamad)employs(malaysia,mahatir-mohamad)

hasCitizenship(mahathir-mohamad, malaysia) 0.75

System ArchitectureTraining

DocumentsInformation Extractor

(IBM SIRE)Extracted

Facts

Inductive LogicProgramming

(LIME)

First-OrderLogical Rules

BLP Weight Learner

Bayesian LogicProgram (BLP)

BLP InferenceEngine

TestDocument

Extractions

Inferences withprobabilities 32

Inductive Logic Programming (ILP) for learning first-order rules

ILP Rule Learner

Target relationhasCitizenship(X,Y)

Positive instanceshasCitizenship(BarackObama, USA)

hasCitizenship(GeorgeBush, USA)

hasCitizenship(IndiraGandhi,India)

.

.

Negative instanceshasCitizenship(BarackObama, India)

hasCitizenship(GeorgeBush, India)

hasCitizenship(IndiraGandhi,USA)

.

.

KBhasBirthPlace(BarackObama,USA)person(BarackObama)nationState(USA)nationState(India)

.

.

RulesnationState(Y) ∧person(X)∧ isLedBy(Y,X) hasCitizenship(X,Y)

..

Generated using clo

sed-

world assu

mption

33

Inference using BLPsTest document“Barack Obama is the current President of the USA……. Obama was born on August 4, 1961, in Hawaii, USA…….”

Extracted factsnationState(usa)person(barackobama)isLedBy(usa,barackobama)hasBirthPlace(barackobama,usa)employs(usa, barackobama)

Learned rulesnationState(B) ∧ person(A) ∧ isLedBy(B,A) hasCitizenship(A,B)nationState(B) person(A) ∧ ∧ employs(B,A) hasCitizenship(A,B)

34

Logical Inference - Proof 1

hasCitizenship(barackobama,usa)

nationState(usa) person(barackobama) isLedBy(usa,barackobama)

nationState(B) ∧ person(A) ∧ isLedBy(B,A) hasCitizenship(A,B)

35

Logical Inference - Proof 2

hasCitizenship(barackobama,usa)

nationState(usa) person(barackobama) employs(usa,barackobama)

nationState(B) ∧ person(A) ∧ employs(B,A) hasCitizenship(A,B)

36

Bayesian Network ConstructionnationState

(usa)

isLedBy(usa,

barackobama)

employs(usa,

barackobama)

hasCitizenship(barackobama, usa)

37

person(barackobama)


(usa)

isLedBy(usa,

barackobama)

employs(usa,

barackobama)


38

person(barackobama)


(usa)

isLedBy(usa,

barackobama)

employs(usa,

barackobama)


39

person(barackobama)


(usa)

isLedBy(usa,

barackobama)

- - -

- - -

- - -

- - -

Logical

And

employs(usa,

barackobama)

dummy1 dummy2


- - -- - -

- - -

- - -

Logical

And- - -- - -

- - -

- - -

Noisy

Or

40

person(barackobama)

Marginal Probability ??


• Data– DARPA’s intelligence community (IC) data set from

the Machine Reading Project (MRP)– Consists of news articles on politics, terrorism,

and other international events– 10,000 documents in total

• Perform 10-fold cross validation

41


• Learning first-order rules using LIME [McCreath and Sharma, 1998]

– Learn rules for 13 target relations– Learn rules using both positive and negative instances

and using only positive instances– Include all unique rules learned from different models

• Learning BLP parameters– Learn noisy-or parameters using Expectation

Maximization (EM)– Set priors to maximum likelihood estimates

42


• Performance evaluation– Lack of ground truth for evaluation– Manually evaluated inferred facts from 40

documents, randomly selected from each test set– Compute precision

– Fraction of inferences that are correct– Compute two precision scores

• Unadjusted (UA) – does not account for extractor’s mistakes• Adjusted (AD) – account for extractor’s mistakes

– Rank inferences using marginal probabilities and evaluate top-n

43


• Systems compared– BLP Learned Weights

• Noisy-or parameters learned using EM– BLP Manual Weights

• Noisy-or parameters set to 0.9– Logical Deduction– MLN Learned Weights

• Learn weights using generative online weight learner– MLN Manual Weights

• Assign a weight of 10 to all rules and MLE priors to all predicates

44

Unadjusted Precision

45

Inferior performance of EM

• Insufficient training data• Lack of ground truth information for relations

that can be inferred– Implicit relations seen less frequently in training

data– EM learns lower weights for rules corresponding

to implicit relations

46

Performance of MLNs• Inferior performance of MLNs

– Insufficient training data for learning– Use of closed world assumption for inference and learning– Lack of strictly typed ontology

• GeopoliticalEntity could be an Agent as well as Location

• Improvements to MLNs– Integrity constraints to avoid inference of spurious facts

like employs(a,a)– Incorporate techniques proposed by Sorower et al. [2011]

47





48

Limitations of LIME• Assumes data is accurate

– Negative instances artificially generated are usually noisy and inaccurate

– Extraction errors result in noisy data• Does not scale to large corpora

49

Develop an approach that can learn first-order rules from noisy and incomplete IE extractions

Online Rule Learning

• Incorporates the incomplete nature of natural language text– Body consists of relations that are explicitly stated– Head is a relation that can be inferred

• Relations that are implicit occur less frequently than those that are explicitly stated– Use frequency of occurrence as a heuristic to

distinguish different types of relations• Process examples in an online manner to scale to

large corpora

50

Approach

• For each example, construct a directed graph of relation extractions

• Add directed edges between nodes that share one or more constants– Relations connected by edges are related and

participate in the same rule• Traverse the graph to learn first-order rules

51

Learning from positive instances only

Example

“Barack Obama is the current President of the USA……. Obama, citizen of the USA was born on August 4, 1961, in Hawaii, USA…….”

Extracted facts

nationState(USA)person(BarackObama)isLedBy(USA,BarackObama)hasBirthPlace(BarackObama,USA)hasCitizenship(BarackObama,USA)

52

Example

Extracted facts


Entities


53

Example

Extracted facts


Relations


54

Directed graph constructionisLedBy(USA,

Barack Obama)

hasBirthPlace(Barack Obama,

USA)

hasCitizenship(Barack Obama,

USA)

isLedBy 33hasBirthPlace 25hasCitizenship 17

?

55

Graph Traversal

isLedBy(USA, Barack Obama) hasBirthPlace(Barack Obama, USA)

isLedBy(USA,

Barack Obama)


USA)

56

Graph Traversal

isLedBy(USA, Barack Obama) ∧person(Barack Obama) ∧ nationState(USA) hasBirthPlace(Barack Obama, USA)

isLedBy(USA,

Barack Obama)


USA)

57

Graph Traversal

isLedBy(X, Y) ∧person(Y) ∧ nationState(X) hasBirthPlace(Y, X)

isLedBy(USA,

Barack Obama)


USA)

58

Rules learned

isLedBy(X, Y) ∧ person(Y) ∧ nationState(X) hasBirthPlace(Y, X)

isLedBy(X, Y) ∧ person(Y) ∧ nationState(X) hasCitizenship(Y, X)

hasBirthPlace(X, Y) ∧ person(X) ∧ nationState(Y) hasCitizenship(X, Y)

59

Sample rules

employs(X, Y) commercialOrganization(X)∧ hasMemberPerson(X, Y)

isLedBy(X, Y) nationState(X)∧ hasCitizenship(Y, X)

isLedBy(X, Y) nationState(X) person(Y) ∧ ∧ hasBirthPlace(Y, X)

60

Experimental Evaluation• Learn first-order rules for 14 target relations

– Full-set– Subset

• 10 target relations

• Manually set noisy-or parameters to 0.9• Systems compared

– Online Rule Learner (ORL)– LIME [McCreath and Sharma, 1998]

– Combined

61

Full-set

62

Inferior performance of ORL on Full-set

• Several incorrect inferences with high marginal probabilities– Instances of thingPhysicallyDamaged and

eventLocationGPE– High probabilities due to multiple rules inferring

these instances– Rules not very accurate resulting in inaccurate

inferences

63

Subset

64

Running Time

• LIME– Learns rules for one target relation at a time– Includes time taken to learn from positive only and

from positive and negative examples• ORL

– Learns rules for all target relations at once

65

ORL LIME

3.8 mins 11.23 hrs





66

Scoring first-order rules• Predicate names employ English words• Confident rules typically have predicates whose

words are semantically related• Use word similarity or relatedness to calculate

weights– Word similarity computed using WordNet

• Compute weights between 0 and 1, which are then used as noisy-or parameters– Higher weights indicate more confident rules

67

WordNet[Fellbaum, 1998]

• Lexical knowledge base consisting of 130,000 English words

• Nouns, verbs, adjectives, and adverbs organized into “synsets” (synonym sets)

• wup [Wu and Palmer, 1994] similarity measure to compute word similarity– Computes scaled similarity scores between 0 and 1– Computes the depth of the least common subsumer

of the given words and scales it by the sum of the depths of the given words

68

Scoring rules using WUP

• Compute word similarity using wup for every pair of words (wi,wj)– wi refers to words in the body

– wj refers to words in the head

• Compute average similarity for all pairs of words• Predicate names like hasCitizenship and

hasMember are segmented into has, citizenship, and member– Stop words are removed

69

Example

employs(X,Y) governmentOrganization(X) ∧ hasMember(X,Y)

70

Example

employs(X,Y) governmentOrganization(X) ∧ hasMember(X,Y)(employs, government, organization) (member)

71

Example

employs(X,Y) governmentOrganization(X) ∧ hasMember(X,Y)(employs, government, organization) (member)

72

Word pair wup scoreemploys, member .50government, member .75organization, member .85

Average .70

Example

employs(X,Y) governmentOrganization(X) ∧ hasMember(X,Y) (.70)(employs, government, organization) (member)

employs(X,Y) person(Y) nationState(X)∧ ∧ hasBirthPlace(Y,X) (.67)(employs, person, nation, state) (birth, place)

73

Scoring rules using WUP• WUP-AVG

– Use words from both entities and relations– Use the average similarity between all pairs of words as

the weight• WUP-MAX

– Use words from both entities and relations– Use maximum similarity among all pairs of words as the

weight• WUP-MAX-REL

– Use words from relations only– Use maximum similarity among all pairs of words as the

weight

74

Experimental Evaluation• Target relations

– Full-set– Subset

• Models– COMBINED

• Rule scoring approaches compared– WUP-AVG– WUP-MAX– WUP-MAX-REL– Default (Manual weights set to 0.9)– EM (Weights learned from EM)

75

Full-set

76

Subset

77

Summary

• BLP approach for inferring implicit facts with high precision

• Superior performance of BLPs over purely logical deduction and MLNs

• Efficient learning of probabilistic first-order rules using online rule learning

• Efficacy of WUP-AVG for scoring first-order rules

78





79

Future Work

• Plan recognition– Structure learning of abductive knowledge bases

for BALPs– Comparison of BALPs to other SRL models

• ProbLog [Kimmig et al., 2008]

• PRISM [Sato, 1995]

• Poole’s Horn Abduction [Poole, 1993]

• Abductive Stochastic Logic Programs [Tamaddoni-Nezhad, Chaleil, Kakas, & Muggleton, 2006]

80

Future Work

• Machine Reading– Large scale evaluation using crowdsourcing– Comparison of BLPs to existing approaches on

machine reading [Schoenmackers et al., 2010; Carlson et al., 2010; Doppa et al., 2010; Sorower et al., 2011]

– Alternate approaches to scoring rules• Use models from distributional semantics [Garrette et

al., 2011]

81

Long-term Directions

• Parameter learning – Using approximate inference techniques– Discriminative learning of parameters

• Lifted inference for BLPs and BALPs

82

Conclusions

• Demonstrated the efficacy of BLPs on two diverse tasks– Plan recognition

• BALPs

– Machine reading• Infer implicit facts from natural language text• Online rule learner for efficient learning of first-order

rules from noisy IE extractions• Scoring first-order rules using WordNet

83

Conclusions

• Demonstrated superior performance of BLPs over MLNs on both tasks

• Contributions could have direct impact on the advancement of applications that use plan recognition and machine reading– SIRI– IBM’s Watson system

84

Questions

85