from textual entailment to knowledgeable machines peter clark allen institute for artificial...

38
From Textual Entailment to Knowledgeable Machines Peter Clark Allen Institute for Artificial Intelligence (AI2)

Upload: lydia-reynolds

Post on 15-Jan-2016

220 views

Category:

Documents


1 download

TRANSCRIPT

Page 1: From Textual Entailment to Knowledgeable Machines Peter Clark Allen Institute for Artificial Intelligence (AI2)

From Textual Entailment to Knowledgeable Machines

Peter Clark

Allen Institute for Artificial Intelligence (AI2)

Page 2: From Textual Entailment to Knowledgeable Machines Peter Clark Allen Institute for Artificial Intelligence (AI2)

Mission: achieve scientific breakthroughs by constructing AI systems with reasoning, learning, and reading capabilities.

2

Page 3: From Textual Entailment to Knowledgeable Machines Peter Clark Allen Institute for Artificial Intelligence (AI2)

Overall Goals

Vision: The Digital Aristotle Large volumes of general and scientific

knowledge, stored in a "computable“ form that supports reasoning and explanation.

Intermediate Focus: Pass science exams as written Heavy emphasis on semi-automated knowledge acquisition Textual entailment at every step

≈ the “modus ponens” of reasoning

“Explainable Reasoning”

Page 4: From Textual Entailment to Knowledgeable Machines Peter Clark Allen Institute for Artificial Intelligence (AI2)
Page 5: From Textual Entailment to Knowledgeable Machines Peter Clark Allen Institute for Artificial Intelligence (AI2)

The Task

Current focus: 4th grade, multiple choice

science questions Wide variety of question

types Requires general, lexical,

and scientific knowledge

Page 6: From Textual Entailment to Knowledgeable Machines Peter Clark Allen Institute for Artificial Intelligence (AI2)

The 4th Grade NY Regents’ Science Exam What types of questions are there? What would it take to answer them?

Page 7: From Textual Entailment to Knowledgeable Machines Peter Clark Allen Institute for Artificial Intelligence (AI2)

The 4th Grade NY Regents’ Science Exam What types of questions are there? What would it take to answer them?

Page 8: From Textual Entailment to Knowledgeable Machines Peter Clark Allen Institute for Artificial Intelligence (AI2)

Multiple Choice to Textual Entailment

QA: A potato is a fruit?QB: An onion is a fruit?QC: A carrot is a fruit?QD: A pumpkin is a fruit?

1. Convert to 4 true/false questions

2. Convert each true/false question to an entailment problem

T: potato H: fruit?

H: potato is a fruit?

entails?

T: entails?

OR: (for X isa Y questions, and questions with a setup)

Page 9: From Textual Entailment to Knowledgeable Machines Peter Clark Allen Institute for Artificial Intelligence (AI2)

Multiple Choice to Textual Entailment

QA: A potato is a fruit?QB: An onion is a fruit?QC: A carrot is a fruit?QD: A pumpkin is a fruit?

Confidence?

0.040.120.210.64

1. Convert to 4 true/false questions

2. Convert each true/false question to an entailment problem

T: potato H: fruit?

H: potato is a fruit?

entails?

T: entails?

OR: (for X isa Y questions, and questions with a setup)

Answer is D

3. Pick highest confidence answer

Page 10: From Textual Entailment to Knowledgeable Machines Peter Clark Allen Institute for Artificial Intelligence (AI2)

The 4th Grade NY Regents’ Science Exam What types of questions are there? What would it take to answer them? “Basic”

Page 11: From Textual Entailment to Knowledgeable Machines Peter Clark Allen Institute for Artificial Intelligence (AI2)

1. Taxonomic

Simple lexical entailment e.g., T:“sleet” H:“precipitation”

Several good sources of simple “isa” knowledge WordNet, Cyc, Wikipedia

Is a basic operation for more complex entailment tasks

entails?

Page 12: From Textual Entailment to Knowledgeable Machines Peter Clark Allen Institute for Artificial Intelligence (AI2)

2. Definitions

Search for best entailing definition

erosion: The process of being eroded by wind, water, or other natural agents.erosion: The wearing away of rocks and other deposits on the earth's surface …erosion: The gradual wearing away of land surface materials, especially rocks, …

Dictionary Resources

Page 13: From Textual Entailment to Knowledgeable Machines Peter Clark Allen Institute for Artificial Intelligence (AI2)

2. Definitions

Search for best entailing definition

erosion: The process of being eroded by wind, water, or other natural agents.erosion: The wearing away of rocks and other deposits on the earth's surface …erosion: The gradual wearing away of land surface materials, especially rocks, …

Dictionary Resources

T: The gradual wearing away of land surface materials, especially rocks, sediments, and soils, by the action of water, wind, or a glacier

H: The movement of soil by wind or water

entails?

Page 14: From Textual Entailment to Knowledgeable Machines Peter Clark Allen Institute for Artificial Intelligence (AI2)

T: The gradual wearing away of land surface materials, especially rocks, sediments, and soils, by the action of water, wind, or a glacier

H: The movement of soil by wind or water

entails?

# wordsin common

weighted∑ #words

# withhypernyms

∑ bi-grams

para-phrases … p(H|T)

3.0 4.31 3.45 2.0 1.20 ?

2.0 1.23 2.12 5.0 1.98

TRAINING DATAH TRUE

6.0 4.31 3.45 0.0 0.20 H FALSE1.0 5.43 1.11 1.0 0.24 H FALSE

3.0 1.12 3.45 2.0 1.76 H TRUE

p(H|T)

= 0.76

Page 15: From Textual Entailment to Knowledgeable Machines Peter Clark Allen Institute for Artificial Intelligence (AI2)

2. Definitions

Search for best entailing definition

erosion: The process of being eroded by wind, water, or otheragents.erosion: The wearing away of rocks and other deposits on the earth …erosion: The gradual wearing away of land surface materials, …

friction: The rubbing of surfaces against each otherfriction: a resistance encountered when one body moves relative to

another body with which it is in contactfriction: surface resistance to relative motion, as of a body sliding

Dictionary Resources0.540.430.76

0.210.11

0.13

Answer is erosion (C)

Page 16: From Textual Entailment to Knowledgeable Machines Peter Clark Allen Institute for Artificial Intelligence (AI2)

3. within-question entailments

T: A girl eating an apple

QB: A girl eating an apple is an example of an organism taking in nutrients?

H: an organism taking in nutrients

entails?

Page 17: From Textual Entailment to Knowledgeable Machines Peter Clark Allen Institute for Artificial Intelligence (AI2)

3. within-question entailments

T: A girl eating an apple

QB: A girl eating an apple is an example of an organism taking in nutrients?

H: an organism taking in nutrients

p(H|T) = 0.91

Page 18: From Textual Entailment to Knowledgeable Machines Peter Clark Allen Institute for Artificial Intelligence (AI2)

The 4th Grade NY Regents’ Science Exam What types of questions are there? What would it take to answer them?

“Entailmentfrom

Corpus”

Page 19: From Textual Entailment to Knowledgeable Machines Peter Clark Allen Institute for Artificial Intelligence (AI2)

4. Entailment from a corpus

HA: The flower is the part of a plant that produces the seedsHB: The leaves are the part of a plant that produces the seedsHC: The stem is the part of a plant that produces the seedsHD: The roots are the part of a plant that produces the seeds

T: HA? HB? HC? HD?

Page 20: From Textual Entailment to Knowledgeable Machines Peter Clark Allen Institute for Artificial Intelligence (AI2)

4. Entailment from a corpus

HA: The flower is the part of a plant that produces the seeds HB: The leaves are the part of a plant that produces the seedsHC: The stem is the part of a plant that produces the seedsHD: The roots are the part of a plant that produces the seeds

T:

…Plants can grow from a seed into a flower, tree, or bush.Plants reproduce by producing flowers and fruits that have seeds.The seeds then grow into plants.…

EntailmentConfidence?

0.840.120.210.04

Page 21: From Textual Entailment to Knowledgeable Machines Peter Clark Allen Institute for Artificial Intelligence (AI2)

The 4th Grade NY Regents’ Science Exam What types of questions are there? What would it take to answer them?

“Models”

“Diagrams”

Page 22: From Textual Entailment to Knowledgeable Machines Peter Clark Allen Institute for Artificial Intelligence (AI2)

5. Computational Models

Requires a specific computation over a representation

baby shake rattle rattle make noise

movement

mechanical energy

sound

sound energy

(C) Mechanical Energy

Page 23: From Textual Entailment to Knowledgeable Machines Peter Clark Allen Institute for Artificial Intelligence (AI2)

Performance (excluding diagrams)

Works okay… ~55% score (vs. 25% random guessing)

Page 24: From Textual Entailment to Knowledgeable Machines Peter Clark Allen Institute for Artificial Intelligence (AI2)

But…

If horses are kept inside in a barn, they require regular daily exercise for their physical health and mental well-being.

?

Carrots can also be used alone or with fruits in jam and preserves.

System answer: (B)

System answer: (C)

Page 25: From Textual Entailment to Knowledgeable Machines Peter Clark Allen Institute for Artificial Intelligence (AI2)

System answer: (A)

Graders are commonly used in the construction and maintenance of dirt roads and gravel roads

In experiments in which statoliths were replaced with metal shavings, researchers "tricked" crayfish into swimming upside down by using magnets to pull the shavings to the upper end of the statocysts located at the base of their antennae.

?System answer: (D)

Page 26: From Textual Entailment to Knowledgeable Machines Peter Clark Allen Institute for Artificial Intelligence (AI2)

What is going on?

Largely “smart guessing” based on word associations BUT:

doesn’t give us meaningful explanations doesn’t get us closer to machine reading performance max’es out at ~60%

What is missing? Richer representation of meaning

in the question in the corpus

Page 27: From Textual Entailment to Knowledgeable Machines Peter Clark Allen Institute for Artificial Intelligence (AI2)

Entailment

? ?- has-part(ribosome,?x).

Text Logic

Query

logicalentailment

“textual”entailment

Page 28: From Textual Entailment to Knowledgeable Machines Peter Clark Allen Institute for Artificial Intelligence (AI2)

Entailment

? ?- has-part(ribosome,?x).

Text Logic

Query

logicalentailment

“textual”entailment

Page 29: From Textual Entailment to Knowledgeable Machines Peter Clark Allen Institute for Artificial Intelligence (AI2)

Textual Entailment

? ?- has-part(ribosome,?x).

Text Logic

Query

Semi-

Formal

logicalentailment

“textual”entailment

Bag of wordsN-gramsParse treesDependency trees…

?Textual Entailment =The Science of Semi-Formal

Representations

Page 30: From Textual Entailment to Knowledgeable Machines Peter Clark Allen Institute for Artificial Intelligence (AI2)

Semi-formal representations

Lexical – good baseline Dependency trees – precise but complex to manipulate Full logic – very hard to translate into SVO units (“depth 1 parses”, “tuples”)?

top-level: syntactic structure lower level: phrasal

These simple propositions = a basic “unit of meaning” Inference is a mixture of structural and phrasal matching

The vibrations from sound move tiny bones in our ears.

(the vibrations from sound, move, tiny bones in our ears)

subject verb object [pps]

Text

Repn.

Page 31: From Textual Entailment to Knowledgeable Machines Peter Clark Allen Institute for Artificial Intelligence (AI2)

Sentences may contain one or more related units

Some animals grow thick fur in winter to stay warm.

P1 P2(some animals, grow, thick fur, in winter) (some animals, stay, , warm)

Types of relations between units:

P1 AND P2P1 IMPLIES P2P1 EFFECT P2P1 CAUSES P2P1 PURPOSE P2

EFFECT

Page 32: From Textual Entailment to Knowledgeable Machines Peter Clark Allen Institute for Artificial Intelligence (AI2)

Can define extraction patterns for these units

S V O “to help” V O P1 EFFECT P2

Some animals grow thick fur in winter to help maintain body heat

(some animals, grow, thick fur, in winter) EFFECT (some animals, maintain, body heat)

Page 33: From Textual Entailment to Knowledgeable Machines Peter Clark Allen Institute for Artificial Intelligence (AI2)

Pattern-based Extraction

Can transduce text into this form

P1P2 AND P3P4P5 CAUSES P6P7 ENABLES P8 AND P9P10P11P12 IMPLIES P13P14 CAUSES P15P16

S V O “to help” S V OS V O “in order to” V OS V “cause” S V O… … …

Page 34: From Textual Entailment to Knowledgeable Machines Peter Clark Allen Institute for Artificial Intelligence (AI2)

Pattern-based Extraction

S V O “to help” S V OS V O “in order to” V OS V “cause” S V O… … …

(gravity, pull) CAUSE (objects, fall, , towards Earth)(scientists, using, a model) EFFECT(scientists, understand, , better)(, dividing, a single cell) EFFECT(, form, two daughter cells)(Animals, use, saturated fatty acids) EFFECT(Animals, store, energy)(humans, get, regular rest) EFFECT (humans, be, healthy)(Fish, have, fins)EFFECT(fins, move, )(animals, move, to warmer climate) EFFECT (animals, avoid, change in seasons)(cactus, hold, water) EFFECT(cactus, survive, in the desert)

… … … …

Can transduce text into this form

Page 35: From Textual Entailment to Knowledgeable Machines Peter Clark Allen Institute for Artificial Intelligence (AI2)

The child exhibits the facial features characteristic of this disorder.

HA A facial scar is a characteristic that a human offspring can inherit?

T

System Answer: (A)

Page 36: From Textual Entailment to Knowledgeable Machines Peter Clark Allen Institute for Artificial Intelligence (AI2)

Some traits that can be inherited are color of hair, color of skin, color of eyes, and height.

HB Blue eyes is a characteristic that a human offspring can inherit?

T

System Answer: (B)

(human offspring, can inherit, the characteristic of blue eyes)

(, can inherit, color of hair)(, can inherit, color of skin)(, can inherit, color of eyes)(, can inherit, height)

Page 37: From Textual Entailment to Knowledgeable Machines Peter Clark Allen Institute for Artificial Intelligence (AI2)

Where does this break down?

1. Quality of extractions is low

2. Sentence-level units are too small, context independent

(, create, a diagram) EFFECT (, show, the cells of multicellular organisms may be organized at different levels)

(, Say, thanks to the authors) EFFECT (, access, a customizable version of this book)(All cells, are, small, very) EFFECT (one or more cells, need, )

Need larger-sized structures spanning several sentences

Page 38: From Textual Entailment to Knowledgeable Machines Peter Clark Allen Institute for Artificial Intelligence (AI2)

The Main Points Textual entailment is

less about “matching text” more about the science of

semi-formal representations Those representations

explicate the “world knowledge” the text encodes

can be sharable “knowledge resources” in their own right

take us a step closer to “knowledgeable machines”

?Query

Semi-

FormalText