Generalizing semantic relations - PowerPoint Presentation

Slide1

Generalizing semantic relations

１２月７日

研究会

祭都援炉

(

マットエンロ

)Slide2

Up until now: Getting to know NLP

“Speech and Language Processing” (

Jurafsky

& Martin)

論文：

On-Demand Information Extract

(

Sekine

)

Learning First-Order Horn Clauses From Web Text

[Sherlock] (

Schoenmackers

2010)

Coupled Semi-Supervised Learning for Information Extraction

[NELL] (Carlson)

Identifying Relations for Open Information Extraction

[

ReVerb

] (Fader)

Relation Acquisition using Word Classes and Partial Patterns

(

Saeger

)

Interpretation as Abduction

(Hobbs)

An ILP Formulation of

Abudctive

Inference for Discourse Interpretation

(Inoue

氏

)

Learning Dependency-Based Compositional Semantics

(Liang)Slide3

Motivation

Ultimate Goal: Inference

Inference requires: knowledge

Large scale database of semantic relations have been created from web textSlide4

ReVerb (Fader et al., 2011)

relation

(

arg1

,

arg2

) tuples acquired from large-scale Web data

Over 14.5 million semantic relations released to publicSlide5

5

ReVerb

- Predicate

Detection

POS tag patterns: V W* P

Milk

is a rich source of

calcium which

is critical for

building strong bones

.

ARG1

RELATION ARG2

(

is

a rich source

of

)

(

is

critical

for

)

Slide stolen from Eric Nichol’s

kenkyuukai

2011-11-02Slide6

6

ReVerb

- Argument

Detection

Noun Phrase chunking

Milk

is a rich source of

calcium

which

is critical for

building

strong bones

.

ARG1

RELATION

ARG2

(

Milk

is

a rich source of

calcium

)

(

calcium

is

critical for

strong bones

)

Slide

modified

from Eric Nichol’s

kenkyuukai

2011-11-02Slide7

Examples (from Fader 2011)Slide8

Problem

Many different ways to express equivalent meaning

Consider

resides

relation

Table shows counts of reverb relations containing

live

or

reside

Relations in red

should be generalized to:

reside(<PERSON>,<PLACE>)

We aim to generalize through

semantic clustering

Frequency | Relation

27,383

lives in

10,315

live in

8,653 lived in

5,185

currently resides in

4,002

currently lives in

3,310

now lives in

1,933

resides in

1,548

is a resident of

1,468 live on

1,308

now resides in

1,191 has lived in

1,055 resided in

876

lives on

590 lived on

531

live at

515

still lives in

461 can live up to

456

is a lifelong resident of

444 was a resident of

413 live for

382 must be residents of

332 lives with

332 lived forSlide9

Semantic Clustering

Store

generalizations

in the form

<

Rel

-type

>

(

<

Arg1 Type

>

,

<

Arg2 Type

>

)

Within a

semantic relation generalization dictionarySlide10

Semantic Clustering Goals

Semantic Relations Dictionary

Mapping from

ReVerb’s

specific instances to generalized semantic-placeholder that looks like

<Generalized-

Rel

>

(

<Arg1 Type>,<Arg2 Type>)

Method of mapping real-world relation instances to generalized semantic form

C

an

be accomplished with a semantic similarity function

Clustering and generalize relations

Looking up new relations from textSlide11

Semantic Similarity

Ontological: Similarity based on arguments’ hierarchy of semantic types

Lexical: Similarity based on lexical features of relation

Contextual: Similarity based on surrounding textSlide12

Ontological Similarity

Determine type of arg1 and arg2

WordNet

synset

Sherlock semantic class (

Schoenmackers

2010)Use WordNet similarity functions on argument typeSlide13

Ontological Similarity (Clustering)

Matt

resides in

Sendai

Eric

lives in

Japan

Should these be clustered together? (Yes!)

Matching arg1 type

<Person>

Matching arg2 type

<Place>

High ontological similarity means good chance of clusteringSlide14

Ontological Similarity (Lookup)

Matt

lives on

a

farm

=> ??

?

Eric

lives on

donuts

=> ??

?

Are these the same semantic relation? (NO!)

Multiple entries in dictionary for

lives_on

:

r

esides

(<Living Thing>,<Place>)

nourished_by

(<

Living

Thing>,<

Nourishment

>)

Use argument type similarity testing to differentiate between senses of

lives_onSlide15

Ontological Similarity (Lookup)

Matt

lives on

a

farm

=>

resides

(

Matt,a_farm

)

Eric

lives on

donuts

=> ???

Are these the same semantic relation? (NO!)

Multiple entries in dictionary for

lives_on

:

r

esides

(<Living Thing>,<Place>)

nourished_by

(<

Living

Thing>,<

Nourishment

>)

Use argument type similarity testing to differentiate between senses of

lives_onSlide16

Ontological Similarity (Lookup)

Matt

lives on

a

farm =>

resides

(

Matt,a_farm

)

Eric

lives on

donuts =>

nourished_by

(

Eric,donuts

)

Are these the same semantic relation? (NO!)

Multiple entries in dictionary for

lives_on

:

r

esides

(<Living Thing>,<Place>)

nourished_by

(<

Living

Thing>,<

Nourishment

>)

Use argument type similarity testing to differentiate between senses of

lives_onSlide17

Ontological Similarity (Lookup cont.)

Which version will ontological similarity suggest we return for each example?

Matt

lives on

a

farm

<

Person

>

lives on<?>

<

Place

>

resides

(

Matt,a_farm

)

Eric

lives on

donuts

<Person>

lives on<?>

<Food>

nourished_by

(

Eric,donuts

)

o

nto_sim

(

<Food>

,

<Nourishment>

) is greater than

onto_sim

(

<Food

>

,

<Place>

)

so we know knows

Eric

is

nourished_by

donutsSlide18

Lexical Similarity

Use relationship features to score similarity

N-gram overlap, bag-of-words, …

Weighting content/functional words differently

etcSlide19

Lexical Similarity

Correctly groups together

Lives at

Live in

But erroneously clusters

Lives for

Lives

with

And doesn’t cluster

resides in

(relying on ontological

sim

.

f

or that)

27,383 lives in

10,315 live in

8,653 lived in

5,185 currently resides in

4,002 currently lives in

3,310 now lives in

1,933 resides in

1,548 is a resident of

1,468 live on

1,308 now resides in

1,191 has lived in

1,055 resided in

876 lives on

590 lived on

531 live at

515 still lives in

461 can live up to

456 is a lifelong resident of

444 was a resident of

413 live for

382 must be residents of

332 lives with

332 lived forSlide20

Contextual Similarity

How similar is the surrounding text?

To answer this, we need original text

Will have to hunt down sentences on the web

Time consuming

Feasible?Slide21

Issues - Clustering

Huge data-set

O(n^2) clustering algorithms are infeasible

Investigating efficient methods:

Hierarchical Clustering

Co-Clustering (

Dhillon

et al., 2003)Probablistic Latent Semantic Indexing

Location Sensitive

HashingSlide22

Other Issues

Word tense

Does

lived in

belong with

lives in

?

Detection of conflicting polarity 

(

Acesulfame_Potassium

does_not_promote

tooth_decay

)

(Conservatives 

should_not_promote

 democracy) 

(Website 

must_not_promote

 hate) 

?

 (Environmentalists 

are_not_alone_in_promoting

renewable_energy

)

Semantic type coverage problems

Use lexical similarity-based lookup for semantic type too?Slide23

progress

Up to now

Looking aheadSlide24

進捗報告

Set up

git

repository

Implemented:

Wrapper for reverb

(data lookup)

WordNet type-lookupSherlock type-lookup

Ontological similarity

Made a slideshow for

研究会

ただ今ご覧になっていただいている物Slide25

計画！！！！！

Finish similarity score

Selecting a

wordnet

ontological similarity function

(Over 5 different evaluations already exist)

Implement lexical similarity(Should already be in NLTK somewhere)Implementing contextual similarity

(Prepare for the hunt!)

Selecting & implementing a clustering method

Test on

ReVerb

data

First on

wikipedia

…

Then on

clueweb