Semantic Framework based Query Generation for Temporal Question Answering over Knowledge Graphs

2025-04-15 0 0 417.7KB 11 页 10玖币

侵权投诉

Semantic Framework based Query Generation for Temporal Question

Answering over Knowledge Graphs

Wentao Ding, Hao Chen, Huayu Li, Yuzhong Qu

State Key Laboratory for Novel Software Technology, Nanjing University, China

{wtding, haochen97, lihuayuu}@smail.nju.edu.cn, yzqu@nju.edu.cn

Abstract

Answering factual questions with temporal in-

tent over knowledge graphs (temporal KGQA)

attracts rising attention in recent years. In

the generation of temporal queries, existing

KGQA methods ignore the fact that some in-

trinsic connections between events can make

them temporally related, which may limit their

capability. We systematically analyze the pos-

sible interpretation of temporal constraints and

conclude the interpretation structures as the Se-

mantic Framework of Temporal Constraints,

SF-TCons. Based on the semantic frame-

work, we propose a temporal question answer-

ing method, SF-TQA, which generates query

graphs by exploring the relevant facts of men-

tioned entities, where the exploring process is

restricted by SF-TCons. Our evaluations show

that SF-TQA signiﬁcantly outperforms exist-

ing methods on two benchmarks over different

knowledge graphs.

1 Introduction

With the rapid growth of knowledge graphs, tem-

poral question answering over knowledge graphs

(temporal KGQA) is attracting rising attention in

recent years (Jia et al.,2018b,2021). In tempo-

ral KGQA, a common phenomenon is that ques-

tions express temporal relations between events

or time expressions, while knowledge graphs de-

scribe the facts resulting from each event. Existing

methods handle the heterogeneity between natural

language and knowledge graph representation in

two ways. Some systems express temporal intents

by constructing executable queries, some apply

time-sensitive neural models to rank candidate an-

swers. Considering that neural models are difﬁcult

to characterize the clear boundaries of concepts

(e.g., exactly ﬁlter all events that occur “before

2022”), this paper focuses on generating queries

that correspond to the meaning of questions.

From the logic perspective, formulated queries

are actually logical restrictions about KG facts. The

answers to a question is a set of KG objects, each

of which satisﬁes the corresponding logical restric-

tions. In previous studies (e.g., Jia et al.,2018b),

temporal intents are converted into restrictions over

KG facts with quantitative time values. Example 1

illustrates a typical conversion from a temporal

question to such restriction.

Example 1.

“Who was the president of the U.S.

when John Lennon was shot?”

The corresponding query on Wikidata can be

formulated as the following logical restriction:

T1=time(position_held(ANS, U.S._president))

∧T2=time(Murder_of_John_Lennon)

∧OVERLAPS(T1, T2).

However, the idea of constructing queries with

quantitative restrictions can not exhaust all possible

scenarios. As illustrated in Example 2, facts with

time values are not a necessary premise to introduce

a temporal relation.

Example 2.

“Where was John Lennon standing

when he was shot?”

To construct a comparison restriction, we need

to enumerate the “standing” of J.L. (i.e. all the

experiences of his life). The enumeration is hard

to implement and might introduce errors.

In fact,

the temporal intent does not rely on any time value.

The two events occur simultaneously just because

they are different aspects of the same entity (

wd:

Q2341090), the murder of John Lennon.

The above example reveals that intrinsic connec-

tions can also make events temporally related. We

argue that the neglect of such cases may limit the

capability of existing methods. Therefore, the pos-

sible temporal constraints, especially those that do

not rely on explicit time values, need to be speciﬁ-

cally studied. The main challenges in concluding

For example, Wikidata says that J.L.’s “residence”(

wd:

P551

) includes Liverpool and New York, but does not provide

the corresponding time duration.

arXiv:2210.04490v3 [cs.CL] 11 May 2023

such constraints come from the complexity of nat-

ural language and the lack of supervision signals.

Practical KGQA tasks often provide only question-

answer pairs. i.e., the constraints on the relevant

facts are unknown. Manually enumerating all possi-

ble constraint structures in a huge search space will

be cumbersome or even infeasible. Thus, there is a

need for a lightweight method to model the various

constraints that correspond to possible temporal

intents.

Inspired by the basic idea of frame semantics

that “one cannot understand the meaning of a word

without access to all the encyclopedic knowledge

that relates to that word.” (Fillmore et al.,2006),

we assume that temporal intents are expressed as

certain constraints about corresponding knowledge

and could be interpreted by some structures over

KG facts. Speciﬁcally, the events involved in a tem-

poral constraint should provide certain KG facts,

which support a possible interpretation of it. We

conclude the temporal constraints and their corre-

sponding interpretation structures as the

emantic

ramework of

emporal

Cons

traints, SF-TCons.

SF-TCons describes what kinds of knowledge are

needed and how they are composed in the poten-

tial interpretations. It consists of 6 interpretation

structures, which will be presented in Section 2. To

the best of our knowledge, SF-TCons is the ﬁrst

work to systematically summarize the interpreta-

tion structures for temporal KGQA tasks.

Based on SF-TCons, we propose a semantic-

framework-based question answering method, SF-

TQA, to convert SF-TCons into executable queries.

SF-TQA generates query graphs by exploring

the relevant facts of mentioned entities, where

the query graph is a graph representation of ex-

ecutable logical queries that resembles subgraphs

of KG (Yih et al.,2015). SF-TQA improves the ac-

curacy of query generation by regarding SF-TCons

as restrictions in the exploration. SF-TQA ﬁrstly

evokes possible interpretations of temporal intents

according to TimeML (Pustejovsky et al.,2010) an-

notations. It then grounds the temporal elements in

corresponding interpretation structures by the rele-

vant KG facts. The grounding phase will generate

multiple candidate queries, the best candidate will

be distinguished by ranking the pairs of questions

and serialized queries with a BERT model.

The rest of this paper is organized as follows:

Section 2discusses the SF-TCons in detail. Sec-

tion 3presents SF-TQA. Section 4evaluates the SF-

TQA with two benchmarks over different knowl-

edge graphs. Section 5summarizes the related

work. The last section concludes this paper.

2 Semantic Framework of Temporal

Constraints

As previously introduced, temporal intents reﬂect

constraints on events and time expressions. We

argue that what really supports the constraints is

the essential knowledge underlying the involved

elements. For example, in a comparison like “be-

fore WWI”, what is needed is its start time “1914”

rather than the named entity

wd:Q361

in KG. There-

fore, temporal constraints can be interpreted by de-

scribing what kind of knowledge is needed and how

they are composed. The interpretation structures

of the constraints are presented as SF-TCons, the

Semantic Framework of Temporal Constraints.

2.1 Temporal Constraints in Questions

Depending on whether the constraints concern

quantitative attributes of a single event or the re-

lations between events, we classify the temporal

constraints as follows.

Value Constraints.

The intentions about quan-

titative values are often expressed with time values

or ordinals (e.g., “

ﬁrst

president”). They require

certain events to have corresponding temporal or

ordinal attributes. Thus, they could be denoted as

follow.

HASVALUE(E1, T1),(VC-1)

HASVALUE(E1, O1),(VC-2)

where

E, T, O

denotes events, time expressions

and ordinals respectively. As an example, the

intent “ﬁrst president” could be denoted as

HASVALUE(“president”,“ﬁrst”)

. Speciﬁcally,

temporal interrogatives (e.g., “

when

did sth. hap-

pen?”) are denoted as

HASVALUE(E1, T ?)

, which

declare the existence of the temporal attributes but

has no restrict on the speciﬁc value.

Relation Constraints.

The possible relations

between time and events have been well studied in

the AI area. We follow TimeML, the most com-

monly used annotation speciﬁcation, to model the

relation constraints.

Example 3.

TimeML-style annotations for the

question in Example 2:

Where was John Lennon [

Event1

standing] [

Signal1

when] he was [Event2shot]?

TLINK reltype=SIMULTANEOUS target=EVENT

relatedTo=EVENT2signal=SIGNAL1/i

As illustrated in Example 3, temporal relations

are triggered by certain signals (e.g., “when”) and

classiﬁed into pre-deﬁned reltypes. For the practi-

cal demand of QA tasks, we formalized the relation

constraints as

RELATION(TR, E1, T1),(RC-1)

RELATION(TR, E1, E2),(RC-2)

where

denotes the 13 temporal reltypes in

TimeML (Pustejovsky et al.,2003),

and

de-

notes events and time expressions respectively. The

TimeML-style annotation in the example question

corresponds the following RC-2 constraint:

RELATION(SIMULTANEOUS,“standing”,“shot”)

2.2 Interpretation Structure for Temporal

Constraints

As previously mentioned, one temporal constraint

could be supported by various interpretations. We

summarize 6 interpretation structures (IS) accord-

ing to whether the involved event expressions are in-

trinsically connected and what connector between

them can correspond to the expected meanings. In

order to enhance the generality of the IS as much

as possible, we do not restrict the speciﬁc semantic

representations of involved events, but only focus

on the key knowledge that they can provide. The 6

IS are presented as follows.

IS-1 Comparison structure

HASVALUE(E1, T1)|RELATION(Tr, E1, T1|E2)

⇒COMPAREh◦,time(E1),time(T1|E2)i

This structure interprets

VC-1

and

, where

◦

denotes algebraic predicate for time values (Allen,

1983;Jia et al.,2018b). Speciﬁcally, the predicate

◦

is required to be EQUAL in

VC-1

and is determined

according to the identiﬁed type

. This

structure supposes that the involved events provides

certain time values.

For example, the question: “Which movie did

Alfred Hitchcock [

Event1

direct] [

Signal1

in] [

Time1

1960]?” corresponds to the following constraint

and KG facts, where the “direct” event provides

the value “1960-10-7”.

COMPAREhINCLUDES,time(“direct”),“1960”i

Alfred_Hitchcock Psycho (ANS)

director (E1)

“1960-10-7”

in_time

“1960” (T1)−INCLUDES→

IS-2 Ordering Structure

HASVALUE(E1, O1)⇒ORDERhattr(E1), O1i

This structure interprets

VC-2

by ordering entities

(or facts) that are described by

. It supposes that

describes a common attribute of certain objects

to be ordered. For example, the question: “When

did Henry the VIII [

Event1

marry]his [

Ordinal1

ﬁrst]

wife?” corresponds to

ORDERhattr(“marry”),“ﬁrst”i

Henry_VIII_of_England

. . .

Cathe. . . Aragon

(T=1506, 1st (O1))

spouse (E1)

Cathe. . . Parr

(T=1543, 6th)

spouse(E1)

IS-3 Direct Query Structure

HASVALUE(E1, X)⇒FINDhent(E1),attr(X)i

In some cases, the expected values are directly

represented in KG facts. This structure interprets

by directly ﬁnding the expected value

certain attributes of some related entity. It supposes

that the entity is related to the mentioned event

For example, the description: “. . . did the

[

Ordinal1

7th] [

Event1

Harry Potter book]...” cor-

responds to the following representation and KG

facts, where the entity “Harry Potter and the

Deathly Hallows” has some attribute with the value

“7”.

FINDhent(“. . . book”),attr(“7th”)i

Harry. . . Hallows

“7” (O1)

. . . .series_ordinal

H.P. (E1)

part_of. . .

IS-4 Same Entity Structure

RELATION(Tr, E1, E2)

⇒SAMEENTITYhe, attr(E1),attr(E2)i

This structure interprets simultaneous cases of

RC-

. It supposes that the events should be attributes

of a certain entity e.

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：

SemanticFrameworkbasedQueryGenerationforTemporalQuestionAnsweringoverKnowledgeGraphsWentaoDing,HaoChen,HuayuLi,YuzhongQuStateKeyLaboratoryforNovelSoftwareTechnology,NanjingUniversity,China{wtding,haochen97,lihuayuu}@smail.nju.edu.cn,yzqu@nju.edu.cnAbstractAnsweringfactualquestionswithtemporalin-tent...

展开>> 收起<<

Semantic Framework based Query Generation for Temporal Question Answering over Knowledge Graphs.pdf

共11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！

Semantic Framework based Query Generation for Temporal Question Answering over Knowledge Graphs

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: