Transcript

Encoding Generalized Quantifiers in

Dependency-based Compositional Semantics

Yubing Dong โ€“ University of Southern California

Ran Tian โ€“ Tohoku University

Yusuke Miyao โ€“ National Institute of Informatics, Japan

BackgroundGeneralized Quantifiers (GQ)

Generalized Quantifiers (GQ)

Most students like noodles.

Generalized Quantifier

Generalized Quantifiers (GQ)

Most students like noodles.

Property-denoting noun phrase

Generalized Quantifier

Generalized Quantifiers (GQ)

Most students like noodles.

Property-denoting noun phrase

PredicateGeneralized Quantifier

Generalized Quantifiers (GQ)

Most (Student) (LikeNoodles) โˆˆ {0,1}

DenotationsStudent โŠ† ๐‘Š

LikeNoodles โŠ† ๐‘ŠBinary Relation over ๐‘Š

Generalized Quantifiers (GQ)

Most (Student) (LikeNoodles)

iff

๐’๐ญ๐ฎ๐๐ž๐ง๐ญ โˆฉ ๐‹๐ข๐ค๐ž๐๐จ๐จ๐๐ฅ๐ž๐ฌ

๐’๐ญ๐ฎ๐๐ž๐ง๐ญ> 80%

The relation imposed by a GQ is usually based on the notion โ‹… of set cardinalities

Generalized Quantifiers (GQ)

Most (Student) (LikeNoodles)Many

ALotOf

Few

AFew

AtMost[n]

AtLeast[n]

BackgroundRecognizing Textual Entailment (RTE)

Recognizing Textual Entailment (RTE)

Example:โ€ข ๐‘‡1: Mary loves every dog.โ€ข ๐‘‡2: Tom has a dog.โ€ข ๐ป: Tom has an animal that Mary loves.โ€ข ๐‘‡1, ๐‘‡2 โ‡’ ๐ป i.e. ๐‘‡1 and ๐‘‡2 entails ๐ป

Definition: โ€œ๐‘‡ entails ๐ป" (๐‘‡ โ‡’ ๐ป) if, typically, a human

reading ๐‘‡ would infer that ๐ป is most likely trueโ€ข Relatively loose, compared to logical entailment

GQ in RTE

At most 5 students like noodles.

At most 5 Japanese students like udon noodles.

GQ in RTE

At least 5 students like noodles.

At least 5 Japanese students like udon noodles.

GQ in RTE

Most students like noodles.

Most Japanese students like udon noodles.

GQ in RTE

The FraCaS Corpus:โ€ข Built in mid-1990sโ€ข A set of hand-crafted entailment problems covering

wide range of semantic phenomena

Section 1 - Generalized Quantifiers:โ€ข 74 problems:

โ€ข 44 have single premise sentenceโ€ข 30 have multiple premise sentence

GQ in RTE

SystemAccuracy

Single Multi Overall

NatLogMacCartney07 84.1%

N/AMacCartney08 97.7%

CCG-DistParser Syntax 70.5% 50.0% 62.2%

Gold Syntax 88.6% 80.0% 85.1%

Accuracies of previous systems on Section 1 of FraCaS corpus

GQ in RTE

SystemAccuracy

Single Multi Overall

NatLogMacCartney07 84.1%

N/AMacCartney08 97.7%

CCG-DistParser Syntax 70.5% 50.0% 62.2%

Gold Syntax 88.6% 80.0% 85.1%

TIFMO

Baseline 79.5% 86.7% 82.4%

Selection 90.9% 93.3% 91.9%

Relation 88.6% 93.3% 90.5%

Selection+Relation 93.2% 96.7% 94.6%

Accuracies of previous systems on Section 1 of FraCaS corpus

But Iโ€™m getting ahead of myselfโ€ฆ

BackgroundProperties of GQs

Properties of GQsProblem with encoding the โ€œperfect semanticsโ€

Most (Student) (LikeNoodles)

iff

๐’๐ญ๐ฎ๐๐ž๐ง๐ญ โˆฉ ๐‹๐ข๐ค๐ž๐๐จ๐จ๐๐ฅ๐ž๐ฌ

๐’๐ญ๐ฎ๐๐ž๐ง๐ญ> 80%

Challenge: set cardinalities are difficult to perfectly encode

Properties of GQs

Compromise: only encode major GQ propertiesโ€ข Interaction with universal and existential quantificationsโ€ข Conservativityโ€ข Monotonicity

Properties of GQsInteraction with universal and existential quantifications

Case 1:๐ด โŠ† ๐ต โ‡’ ๐น ๐ด ๐ต โ‡’ ๐ด โˆฉ ๐ต โ‰  โˆ…

Example: โ€œmostโ€

All students like noodles.

Most students like noodles.

There are students who like noodles.

Properties of GQsInteraction with universal and existential quantifications

Case 2:๐ด โŠ† ๐ต โ‡’ ๐น ๐ด ๐ต โ‡’ ๐ด โˆฉ ๐ต โ‰  โˆ…

Example: โ€œa lot ofโ€

All students like noodles.

A lot of students like noodles.

There are students who like noodles.

Properties of GQsInteraction with universal and existential quantifications

Case 3:๐ด โŠ† ๐ต โ‡’ ๐น ๐ด ๐ต โ‡’ ๐ด โˆฉ ๐ต โ‰  โˆ…

Example: โ€œat most nโ€

All students like noodles.

At most 5 students like noodles.

There are students who like noodles.

Properties of GQsConservativity

The โ€œdomain restrainingโ€ role of the noun argumentโ€ข Eliminates objects that do not have the noun propertyโ€ข Only need to consider which of the rest has the predicate property

๐น ๐ด ๐ต โŸบ ๐น(๐ด)(๐ด โˆฉ ๐ต)

Example:โ€ข โ€œFew apples are toxic.โ€โŸบโ€œFew apples are toxic apples.โ€โ€ข We donโ€™t care non-apples toxicants, e.g. toxic oranges

Properties of GQsMonotonicity

A GQ ๐น โ‹… โ‹… is upward entailing in the noun argument if:๐น ๐ดโ€ฒ ๐ต โ‡’ ๐น ๐ด ๐ต โˆ€๐ดโ€ฒ โŠ† ๐ด

Similarly, a GQ can also beโ€ข downward entailing in the noun argument, and โ€ข upward/downward entailing in the predicate argument

Properties of GQsMonotonicity

At most 5 students like noodles.

At most 5 Japanese students like udon noodles.

Example: โ€œat most ๐‘›โ€ is downward entailing in each argument

Properties of GQsMonotonicity

Example: โ€œat least ๐‘›โ€ is upward entailing in each argument

At least 5 students like noodles.

At least 5 Japanese students like udon noodles.

Properties of GQsMonotonicity

Example: โ€œmostโ€ is neither upward nor downward entailing in the noun argument

Most students like noodles.

Most Japanese students like noodles.

Properties of GQsMonotonicity

Example: but is upward entailing in the predicate argument

Most students like noodles.

Most students like udon noodles.

BackgroundDependency-based Compositional Semantics (DCS) for RTE

โ€ข Proposed by Tian et al. (2014)

DCS for RTE

DCS tree for โ€œAll students like udon noodlesโ€

DCS for RTE

DCS tree for โ€œAll students like udon noodlesโ€

Abstract Denotations:

๐ง๐จ๐จ๐๐ฅ๐ž โŠ† ๐‘Š๐ฎ๐๐จ๐ง โŠ† ๐‘Š๐ฌ๐ญ๐ฎ๐๐ž๐ง๐ญ โŠ† ๐‘Š๐ฅ๐ข๐ค๐ž โŠ† ๐‘Š ร—๐‘Š

DCS for RTE

๐ท1 = ๐ง๐จ๐จ๐๐ฅ๐ž โˆฉ ๐ฎ๐๐จ๐ง

DCS tree for โ€œAll students like udon noodlesโ€

โ€œudon noodlesโ€

DCS for RTE

๐ท1 = ๐ง๐จ๐จ๐๐ฅ๐ž โˆฉ ๐ฎ๐๐จ๐ง

๐ท2 = ๐ฅ๐ข๐ค๐ž โˆฉ ๐‘Š๐‘†๐ต๐ฝ ร— ๐ท1 ๐‘‚๐ต๐ฝ

DCS tree for โ€œAll students like udon noodlesโ€

โ€œlike udon noodlesโ€

DCS for RTE

๐ท1 = ๐ง๐จ๐จ๐๐ฅ๐ž โˆฉ ๐ฎ๐๐จ๐ง

๐ท2 = ๐ฅ๐ข๐ค๐ž โˆฉ ๐‘Š๐‘†๐ต๐ฝ ร— ๐ท1 ๐‘‚๐ต๐ฝ

๐ท3 = ๐œ‹๐‘†๐ต๐ฝ ๐ท2

DCS tree for โ€œAll students like udon noodlesโ€

โ€œsubjects who like udon noodlesโ€

DCS for RTE

๐ท1 = ๐ง๐จ๐จ๐๐ฅ๐ž โˆฉ ๐ฎ๐๐จ๐ง

๐ท2 = ๐ฅ๐ข๐ค๐ž โˆฉ ๐‘Š๐‘†๐ต๐ฝ ร— ๐ท1 ๐‘‚๐ต๐ฝ

๐ท3 = ๐œ‹๐‘†๐ต๐ฝ ๐ท2๐ท4 = ๐‘žโŠ†

๐‘†๐ต๐ฝ ๐ท3, ๐ฌ๐ญ๐ฎ๐๐ž๐ง๐ญ

DCS tree for โ€œAll students like udon noodlesโ€

qโŠ†r R,C โ‰ก x โˆ…โ‰ Rโˆฉ x ร—Wr โŠ† x ร—Cr

If ๐‘… and ๐ถ have the same dimension,โ€ข ๐‘žโŠ†

๐‘Ÿ ๐‘…, ๐ถ = โˆ— (0-dimension point set) when ๐ถ โŠ† ๐‘…,โ€ข ๐‘žโŠ†

๐‘Ÿ ๐‘…, ๐ถ = โˆ… otherwise

wide reading of โ€œโŠ†โ€

DCS for RTE

๐ท1 = ๐ง๐จ๐จ๐๐ฅ๐ž โˆฉ ๐ฎ๐๐จ๐ง

๐ท2 = ๐ฅ๐ข๐ค๐ž โˆฉ ๐‘Š๐‘†๐ต๐ฝ ร— ๐ท1 ๐‘‚๐ต๐ฝ

๐ท3 = ๐œ‹๐‘†๐ต๐ฝ ๐ท2๐ท4 = ๐‘žโŠ†

๐‘†๐ต๐ฝ ๐ท3, ๐ฌ๐ญ๐ฎ๐๐ž๐ง๐ญ

๐ท5 = ๐‘žโŠ†๐‘†๐ต๐ฝ

๐ท2, ๐ฌ๐ญ๐ฎ๐๐ž๐ง๐ญ

DCS tree for โ€œAll students like udon noodlesโ€

qโŠ†r R,C โ‰ก x โˆ…โ‰ Rโˆฉ x ร—Wr โŠ† x ร—Cr

If ๐‘… and ๐ถ have the same dimension,โ€ข ๐‘žโŠ†

๐‘Ÿ ๐‘…, ๐ถ = โˆ— (0-dimension point set) when ๐ถ โŠ† ๐‘…,โ€ข ๐‘žโŠ†

๐‘Ÿ ๐‘…, ๐ถ = โˆ… otherwise

narrow reading of โ€œโŠ†โ€(โ€œthe set of udon noodles that all student likeโ€)

DCS for RTE

๐ท1 = ๐ง๐จ๐จ๐๐ฅ๐ž โˆฉ ๐ฎ๐๐จ๐ง

๐ท2 = ๐ฅ๐ข๐ค๐ž โˆฉ ๐‘Š๐‘†๐ต๐ฝ ร— ๐ท1 ๐‘‚๐ต๐ฝ

๐ท3 = ๐œ‹๐‘†๐ต๐ฝ ๐ท2๐ท4 = ๐‘žโŠ†

๐‘†๐ต๐ฝ ๐ท3, ๐ฌ๐ญ๐ฎ๐๐ž๐ง๐ญ

๐ท5 = ๐‘žโŠ†๐‘†๐ต๐ฝ

๐ท2, ๐ฌ๐ญ๐ฎ๐๐ž๐ง๐ญ

DCS tree for โ€œAll students like udon noodlesโ€

Prove statement

โ€ข ๐ท4 โ‰  โˆ… (wide reading) orโ€ข ๐ท5 โ‰  โˆ… (narrow reading)

using forward chaining

DCS for RTE

Basic operators / functions:โ€ข ร— - Cartesian product of setsโ€ข โˆฉ - Set intersectionโ€ข ๐œ‹๐‘Ÿ - Projection onto domain of semantic role ๐‘Ÿโ€ข ๐‘™๐‘Ÿ - Relabelingโ€ข ๐‘žโŠ†

๐‘Ÿ - Division

Basic types of statements:โ€ข Non-emptiness: ๐ด โ‰  โˆ…โ€ข Subsumption: ๐ด โŠ† ๐ต

BackgroundDCS for RTE: the selection operator

โ€ข Also introduced in Tian et al. (2014)

DCS for RTE: the selection operator

โ€ข Introduced as an extension to represent the generalized selection operation in relational algebra

โ€ข Marked on a DCS tree nodeโ€ข Wrap the abstract denotation ๐ท to form a new abstract

denotation ๐‘ ๐‘“ ๐ท

โ€ข The properties of ๐‘ ๐‘“ ๐ท can be user defined

Example:

the set of highest mountains: ๐‘ โ„Ž๐‘–๐‘”โ„Ž๐‘’๐‘ ๐‘ก(๐ฆ๐จ๐ฎ๐ง๐ญ๐š๐ข๐ง)

Encoding Generalized Quantifiersas selections

Encoding GQs as Selections

We encode a GQ ๐น using selection ๐‘ ๐น as:

๐น ๐ด ๐ต โ‰ก ๐‘ ๐น ๐ด โŠ† ๐ต

Basic requirement:โ€ข ๐น should be upward-entailing in the predicate

argument ๐ตโ€ข A major limitation

Encoding GQs as Selections

๐น ๐ด ๐ต โ‰ก ๐‘ ๐น ๐ด โŠ† ๐ต

โ€ข Entailment from universal quantification now written as:๐ด โŠ† ๐ต โ‡’ ๐‘ ๐น ๐ด โŠ† ๐ต

โ€ข Conservativity as:๐‘ ๐น ๐ด โŠ† ๐ด โˆฉ ๐ต โ‡” ๐‘ ๐น ๐ด โŠ† ๐ต

โ€ข Both hold if we add axiom:๐‘ ๐น ๐ด โŠ† ๐ด

Encoding GQs as Selections

๐น ๐ด ๐ต โ‰ก ๐‘ ๐น ๐ด โŠ† ๐ต

โ€ข Entailment to existence quantification now written as:๐‘ ๐น ๐ด โŠ† ๐ต โ‡’ ๐ด โˆฉ ๐ต โ‰  โˆ…

โ€ข Holds if we add axiom:๐‘ ๐น ๐ด โˆฉ ๐ด โ‰  โˆ…

Encoding GQs as Selections

๐น ๐ด ๐ต โ‰ก ๐‘ ๐น ๐ด โŠ† ๐ต

โ€ข Monotonicity in the noun argument ๐ด (e.g. upward) now written as:

A โŠ† Aโ€ฒ โˆง ๐‘ ๐น ๐ด โŠ† ๐ต โ‡’ ๐‘ ๐น ๐ดโ€ฒ โŠ† ๐ต

โ€ข Holds if we add axiom:A โŠ† Aโ€ฒ โ‡’ ๐‘ ๐น ๐ด โŠ‡ ๐‘ ๐น ๐ดโ€ฒ

DCS tree for โ€œAt least 5 students like udon noodles.โ€where the GQ โ€œat least 5โ€ is encoded as selection ๐‘ ๐ด๐‘ก๐ฟ๐‘’๐‘Ž๐‘ ๐‘ก 5

Encoding GQs as Selections

Example: at least ๐‘›

โ€ข Satisfied: upward-entailing in predicate argument

โ€ข Entails existential quantification:โˆ€๐ด ๐‘ ๐ด๐‘ก๐ฟ๐‘’๐‘Ž๐‘ ๐‘ก 5 ๐ด โˆฉ ๐ด โ‰  โˆ…

โ€ข Upward-entailing in noun argument:โˆ€๐ด, ๐ดโ€ฒ ๐‘ . t. A โŠ† Aโ€ฒ

๐‘ ๐ด๐‘ก๐ฟ๐‘’๐‘Ž๐‘ ๐‘ก 5 ๐ด โŠ‡ ๐‘ ๐ด๐‘ก๐ฟ๐‘’๐‘Ž๐‘ ๐‘ก 5 ๐ดโ€ฒ

Encoding GQs as SelectionsExample:

โ€œAt least 5 Japanese students like udon noodles.โ€

โ‡’ โ€œ At least 5 students like noodles.โ€

๐ท1 = ๐ง๐จ๐จ๐๐ฅ๐ž โˆฉ ๐ฎ๐๐จ๐ง

๐ท2 = ๐ฅ๐ข๐ค๐ž โˆฉ ๐‘Š๐‘†๐ต๐ฝ ร— ๐ท1 ๐‘‚๐ต๐ฝ

๐ท3 = ๐œ‹๐‘†๐ต๐ฝ ๐ท2

๐ท3โ€ฒ = ๐œ‹๐‘†๐ต๐ฝ ๐ฅ๐ข๐ค๐ž โˆฉ ๐‘Š๐‘†๐ต๐ฝ ร— ๐ง๐จ๐จ๐๐ฅ๐ž๐‘‚๐ต๐ฝ

Encoding Generalized Quantifiersas relations

Encoding GQs as RelationsIntro to Relations

โ€ข Review: GQ can be seen as binary relation over 2๐‘Š

โ€ข Therefore, we introduce a new extension: relationโ€ข A new type of statementโ€ข A relation ๐‘Ÿ๐น ๐ด, ๐ต can represent arbitrary custom

relation between abstract denotations ๐ด and ๐ต

Encoding GQs as RelationsIntro to Relations

Relation ๐‘Ÿ๐น ๐ด, ๐ต

โ€ข The inference engine keeps track of which term pairs are labeled with which relationsโ€ข Does ๐ด and ๐ต have relation ๐‘Ÿ๐น?โ€ข What terms have relation ๐‘Ÿ๐น to ๐ด?

โ€ข Supports custom axioms for a relationโ€ข What entails ๐‘Ÿ๐น ๐ด, ๐ต ?โ€ข What does ๐‘Ÿ๐น ๐ด, ๐ต entail?

Encoding GQs as Relations

We intuitively encode a GQ ๐น using relation ๐‘Ÿ๐น as:

๐น ๐ด ๐ต โ‰ก r๐น ๐ด, ๐ต

๐ท1 = ๐ง๐จ๐จ๐๐ฅ๐ž โˆฉ ๐ฎ๐๐จ๐ง

๐ท2 = ๐ฅ๐ข๐ค๐ž โˆฉ ๐‘Š๐‘†๐ต๐ฝ ร— ๐ท1 ๐‘‚๐ต๐ฝ

๐ท3 = ๐œ‹๐‘†๐ต๐ฝ ๐ท2

Statement:๐‘Ÿ๐ด๐‘ก๐‘€๐‘œ๐‘ ๐‘ก 5 ๐ฌ๐ญ๐ฎ๐๐ž๐ง๐ญ, ๐ท3

Encoding GQs as Relations

๐น ๐ด ๐ต โ‰ก r๐น ๐ด, ๐ต

โ€ข Entailment from universal quantification:๐ด โŠ† ๐ต โ‡’ ๐‘Ÿ๐น ๐ด, ๐ต

โ€ข Entailment to existential quantification:๐‘Ÿ๐น ๐ด, ๐ต โ‡’ ๐ด โˆฉ ๐ต โ‰  โˆ…

โ€ข Monotonicity (e.g. downward in both arguments):๐‘Ÿ๐น ๐ด, ๐ต โˆง ๐ด โŠ‡ ๐ดโ€ฒ โˆง ๐ต โŠ‡ ๐ตโ€ฒ โ‡’ ๐‘Ÿ๐น ๐ดโ€ฒ, ๐ตโ€ฒ

Encoding GQs as Relations

๐น ๐ด ๐ต โ‰ก r๐น ๐ด, ๐ต

โ€ข Conservativity:๐‘Ÿ๐น ๐ด, ๐ต โ‡’ ๐‘Ÿ๐น ๐ด, ๐ด โˆฉ ๐ต

โ€ข How about the other direction?๐‘Ÿ๐น ๐ด, ๐ด โˆฉ ๐ต โ‡’ ๐‘Ÿ๐น ๐ด, ๐ต

Encoding GQs as Relations

๐‘Ÿ๐น ๐ด, ๐ด โˆฉ ๐ต โ‡’ ๐‘Ÿ๐น ๐ด, ๐ต

Challenge:โ€ข The inference engine is based on forward chaining:

โ€ข Always try to deduce all possible implications from given premisesโ€ข Efficientโ€ข Opens the possibility of adapting DCS for entailment

generation

Encoding GQs as Relations

๐‘Ÿ๐น ๐ด, ๐ด โˆฉ ๐ต โ‡’ ๐‘Ÿ๐น ๐ด, ๐ต

Challenge:โ€ข The inference engine is based on forward chainingโ€ข Therefore itโ€™s infeasible to enumerate all forms ๐‘‹ = ๐ด โˆฉ ๐ต

when ๐‘Ÿ๐น ๐ด, ๐‘‹ is claimedโ€ข Number of possibilities explodes exponentially

โ€ข e.g. ๐‘‹ = ๐‘‹ โˆฉ ๐ถ โˆ€๐ถ, ๐‘‹ = ๐ด โˆฉ ๐ต โˆฉ ๐ถ = ๐ด โˆฉ ๐ต โˆฉ ๐ถ

Encoding GQs as Relations

๐‘Ÿ๐น ๐ด, ๐ด โˆฉ ๐ต โ‡’ ๐‘Ÿ๐น ๐ด, ๐ต

Implementation: limit search using conditions ๐‘‹ โŠ† ๐ด โˆง ๐‘‹ โŠ† ๐ต

If ๐‘Ÿ๐น ๐ด, ๐‘‹ and ๐‘‹ โŠ† ๐ด:โ€ข For each ๐ต โŠ‡ ๐‘‹:

โ€ข Check if ๐‘‹ = ๐ด โˆฉ ๐ต

We emphasize this detail because formal semantic researchers are often not aware of these difficulties.

Encoding GQs as RelationsLimitations

๐น ๐ด ๐ต โ‰ก r๐น ๐ด, ๐ต

Limitation:

Relations in DCS trees are always explained as having the widest scope, hence cannot deal with multiple relations in a sentence.

Encoding GQs as RelationsLimitations

Example:๐‘ƒ: At most 10 commissioners spend a lot of time at home.

We want to state๐‘Ÿ๐ด๐‘ก๐‘€๐‘œ๐‘ ๐‘ก 10 ๐œ๐จ๐ฆ๐ข๐ฌ๐ฌ๐ข๐จ๐ง๐ž๐ซ๐ฌ, ๐ท

where ๐ท = โ€œpeople who spend a lot of time at homeโ€

But this is impossible if โ€œa lot ofโ€ is also encoded as a relation

Encoding GQs as RelationsLimitations

Example:๐‘Ÿ๐ด๐‘ก๐‘€๐‘œ๐‘ ๐‘ก 10 ๐œ๐จ๐ฆ๐ข๐ฌ๐ฌ๐ข๐จ๐ง๐ž๐ซ๐ฌ, ๐ท

๐ท = "people who spend a lot of time at home"

Workaround:Since โ€œa lot ofโ€ is upward-entailing in predicate argument, we can encode it using selection ๐‘ ๐ด๐ฟ๐‘œ๐‘ก๐‘‚๐‘“, while still encode โ€œat

most 10โ€ using ๐‘Ÿ๐ด๐‘ก๐‘€๐‘œ๐‘ ๐‘ก 10

Encoding GQs as RelationsLimitations

Example:๐‘Ÿ๐ด๐‘ก๐‘€๐‘œ๐‘ ๐‘ก 10 ๐œ๐จ๐ฆ๐ข๐ฌ๐ฌ๐ข๐จ๐ง๐ž๐ซ๐ฌ, ๐ท

๐ท = ๐‘žโŠ†๐‘‚๐ต๐ฝ

๐ทโ€ฒ, ๐‘ ๐ด๐ฟ๐‘œ๐‘ก๐‘‚๐‘“ ๐ญ๐ข๐ฆ๐ž

where

๐ทโ€ฒ = ๐ฌ๐ฉ๐ž๐ง๐ โˆฉ ๐‘Š๐‘†๐ต๐ฝ ร—๐‘Š๐‘‚๐ต๐ฝ ร— ๐ก๐จ๐ฆ๐ž๐‘€๐‘‚๐ท

(โ€œspend at homeโ€)

Evaluation

EvaluationSet-up

The FraCaS Corpus:โ€ข Built in mid-1990sโ€ข A set of hand-crafted entailment problems covering

wide range of semantic phenomena

Section 1 - Generalized Quantifiers:โ€ข 74 problems:

โ€ข 44 have single premise sentenceโ€ข 30 have multiple premise sentence

EvaluationSet-up

Settings:โ€ข Baselineโ€ข Selectionโ€ข Relationโ€ข Selection+Relation

EvaluationSet-up

Settings:โ€ข Baseline

โ€ข Simply drop GQsโ€ข Same tree structure as follows

โ€ข Selectionโ€ข Relationโ€ข Selection+Relation

EvaluationSet-up

Settings:โ€ข Baselineโ€ข Selection

โ€ข Implement all GQs as selections, even for those that are downward-entailing in predicate argument

โ€ข Relationโ€ข Selection+Relation

EvaluationSet-up

Settings:โ€ข Baselineโ€ข Selectionโ€ข Relation

โ€ข Implement all GQs as relationsโ€ข Selection+Relation

EvaluationSet-up

Settings:โ€ข Baselineโ€ข Selectionโ€ข Relationโ€ข Selection+Relation

โ€ข Use relations to encode GQs that are downward-entailing in predicate argument

โ€ข Encode the rest with selections

Evaluation

SystemAccuracy

Single Multi Overall

NatLogMacCartney07 84.1%

N/AMacCartney08 97.7%

CCG-DistParser Syntax 70.5% 50.0% 62.2%

Gold Syntax 88.6% 80.0% 85.1%

TIFMO

Baseline 79.5% 86.7% 82.4%

Selection 90.9% 93.3% 91.9%

Relation 88.6% 93.3% 90.5%

Selection+Relation 93.2% 96.7% 94.6%

Accuracies of previous systems on Section 1 of FraCaS corpus

Conclusion

Conclusion

โ€ข Generalized Quantifiers are important (for RTE)

โ€ข We explored ways of encoding GQs in DCS for RTEโ€ข via selection extensionโ€ข via relation extension (newly proposed)

โ€ข Significant improvement in performance, but not perfectโ€ข which suggests towards more powerful logical systems