Does Wikidata Support Analogical Reasoning Filip Ilievski Jay Pujara and Kartik Shenoy Information Sciences Institute University of Southern California

2025-04-27 0 0 1.36MB 14 页 10玖币

侵权投诉

Does Wikidata Support Analogical Reasoning?

Filip Ilievski, Jay Pujara, and Kartik Shenoy

Information Sciences Institute, University of Southern California

{ilievski, jpujara, kshenoy}@isi.edu

Abstract. Analogical reasoning methods have been built over various

resources, including commonsense knowledge bases, lexical resources,

language models, or their combination. While the wide coverage of knowl-

edge about entities and events make Wikidata a promising resource for

analogical reasoning across situations and domains, Wikidata has not

been employed for this task yet. In this paper, we investigate whether

the knowledge in Wikidata supports analogical reasoning. Speciﬁcally,

we study whether relational knowledge is modeled consistently in Wiki-

data, observing that relevant relational information is typically missing

or modeled in an inconsistent way. Our further experiments show that

Wikidata can be used to create data for analogy classiﬁcation, but this

requires much manual eﬀort. To facilitate future work that can support

analogies, we discuss key desiderata, and devise a set of metrics to guide

an automatic method for extracting analogies from Wikidata.

Keywords: Wikidata ·Analogical Reasoning ·Ontologies ·User Expe-

rience

1 Introduction

Cognitive science research has provided rich evidence that humans use analogical

reasoning to understand, explain, or imagine novel situations within or across

domains [13]. Analogical thinking can connect the Great Depression and the ﬁ-

nancial crisis based on causal knowledge [21], or compare the Sun and the Earth

to the Earth and the Moon based on the revolves relation [11]. Correspond-

ing cognitive systems have been build to realize and test this skill algorithmi-

cally, such as the Structured Mapping Engine [9] and the Companion architec-

ture [10]. Natural Language Processing research on analogy has been popularized

through the proportional analogy task, illustrated through the famous example

of man:woman-king:queen by the word2vec system [19]. Recognizing the gap

between the large-scale word analogy systems and the expressive cognitive sys-

tems, recent AI research has focused on integrating neural (language) models

with cognitive systems to solve tasks like sketch object recognition [3], product

innovation [14], narrative understanding [21], and moral decision making [5].

As it can be expected, these analogical reasoning eﬀorts are often centered

around a knowledge base that enables models to understand implicit relations,

such as causes or revolves. Curiously, despite the large quality and richness of

arXiv:2210.00620v1 [cs.AI] 2 Oct 2022

2 Ilievski et al.

Wikidata, and its increasing adoption for many knowledge-intensive tasks [20,16,23],

prior work on analogical reasoning has not considered leveraging its ontology

or its factual knowledge to reason by analogy. Instead, existing systems have

typically leveraged publicly available parts of Cyc, semantic lexical resources,

language models [27], or their combination [10].

Considering the coverage of millions of ontological concepts and instances,

intuitively, Wikidata could serve as a valuable resource for analogical reasoning.

In this paper, we perform an initial study on whether the Wikidata knowledge

supports analogical reasoning. Speciﬁcally, we focus on three key questions:

1. Does Wikidata express relational information consistently? Analogical rea-

soning revolves around relational similarity, therefore, consistency in the

knowledge modeling of relational knowledge is crucial to enable reason-

ing systems to connect between two situations or domains. We investigate

whether relational information is consistently modeled in Wikidata.

2. Does Wikidata support extraction of analogy classiﬁcation data? Given its

wide coverage, Wikidata may have the potential to generate large-scale anal-

ogy detection tasks automatically. We investigate how much manual eﬀort

is required to create such benchmarks for subclass-of relations in Wikidata,

and we evaluate the performance of state-of-the-art NLP systems.

3. Which desiderata and metrics can guide automatic generation of analogies

using the Wikidata structure? Considering that the Wikidata ontology is not

uniform in terms of its granularity and expressivity, it is important to deﬁne

desired properties for analogical reasoning and design automated metrics

that can quantify this variation and enable the selection of potential analog-

ical correspondences automatically.

2 Does Wikidata Express Relational Information

Consistently?

2.1 Data and Setup

We sample 20 subclass-of (P279) relations from Wikidata, whose subject label

is a superstring of its object label. For example, we keep the subclass pair red

wine - wine, while we discard dog - pet. We prioritize Qnodes with low identiﬁers

as a simple proxy for well-known entities and concepts. Example pairs include

stellar atmosphere - atmosphere and computer keyboard - keyboard. We analyze

whether the subclass-of relation is complemented by additional information that

can help us categorize the nature of the inheritance relation, expressed either as

other relations of the subject or qualiﬁers on the subclass-of relation. Based on

prior work on categorization of semantic relations for noun compounds [12], we

deﬁne an initial set of seven inheritance categories: PURPOSE, PROPERTY,

LOCATION, OWNERSHIP, MATERIAL, INSTANCE, and TEMPORAL, and

annotate each pair with one category. Besides obtaining relations for the original

pair, we obtain siblings of the subject (other Qnodes that are direct children of

the same object) and investigate their structures seeking for regularities.

Does Wikidata Support Analogical Reasoning? 3

Table 1. Ten exemplar compound noun pairs.

Subject Object Category Qualiﬁers Statements Siblings

computer

keyboard

(Q250)

keyboard

(Q1921606)

PURPOSE follows:

mobile

phone

computer key-

board - part

of - computer

typewriter keyboard,

Braille keyboard, musical

keyboard

natural

science

(Q7991)

science

(Q336)

PROPERTY - - human science, informa-

tion science, modern sci-

ence, Ancient Egyptian sci-

ence, ...

beach

volleyball

(Q4543)

volleyball

(Q1734)

LOCATION - - snow volleyball, women’s

volleyball, men’s volleyball

fairy tale

(Q699)

tale

(Q17991521)

PROPERTY - - old-fashioned tale, cumula-

tive tale, urbain tale, Ger-

man folk tale

Shia Islam

(Q9585)

Islam

(Q432)

INSTANCE - - Sunni Islam, Islam in Den-

mark, Islamic eschaetology,

Gospel in Islam

stellar at-

mosphere

(Q6311)

atmosphere

(Q8104)

LOCATION of: star - extrasolar atmosphere, ex-

traterrestrial atmosphere,

Reducing atmosphere

electric

charge

(Q1111)

charge

(Q73792)

PROPERTY of: electro-

magnetic

ﬁeld

- magnetic charge, color

charge, weak hypercharge

red wine

(Q1827)

wine

(Q282)

PROPERTY - red wine -

color - red

white wine, Mexican wine,

Polish wine, straw wine, de-

alcoholised wine, ...

day sky

(Q4812)

sky (Q527) PROPERTY - - blue sky, morning sky, Ve-

lazquez sky, ...

animal

rights

(Q426)

right

(Q2386606)

PROPERTY of: non-

human

animal

- hunting rights, women’s

rights, right to property, ...

2.2 Findings

We show ten out of the twenty pairs in Table 1. Overall, we ﬁnd that Wikidata

describes relations sparsely, which does not help us identify the compound re-

lation category. Speciﬁcally, out of 20 pairs, we found 4 cases where Wikidata

provided a qualiﬁer to further specify the relation. Among these four qualiﬁers,

three were expressed with of (e.g., stellar atmosphere - atmosphere is further

speciﬁed by the qualiﬁer of - star ) and a single case used the follows qualiﬁer

(computer keyboard - keyboard is speciﬁed by follows - mobile phone). In addi-

tion to being sparse, we ﬁnd the qualiﬁer information to correlate weakly with

our semantic categories, as of corresponds to both LOCATION (in stellar atmo-

sphere - atmosphere - of: star) and to PROPERTY (in electric charge - charge

- of: electromagnetic ﬁeld).

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

DoesWikidataSupportAnalogicalReasoning?FilipIlievski,JayPujara,andKartikShenoyInformationSciencesInstitute,UniversityofSouthernCaliforniafilievski,jpujara,kshenoyg@isi.eduAbstract.Analogicalreasoningmethodshavebeenbuiltovervariousresources,includingcommonsenseknowledgebases,lexicalresources,language...

展开>> 收起<<

Does Wikidata Support Analogical Reasoning Filip Ilievski Jay Pujara and Kartik Shenoy Information Sciences Institute University of Southern California.pdf

共14页,预览3页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Does Wikidata Support Analogical Reasoning Filip Ilievski Jay Pujara and Kartik Shenoy Information Sciences Institute University of Southern California

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: