Modeling Document-level Temporal Structures for Building Temporal Dependency Graphs Prafulla Kumar Choubey

2025-05-06 0 0 458.56KB 9 页 10玖币

侵权投诉

Modeling Document-level Temporal Structures for Building Temporal

Dependency Graphs

Prafulla Kumar Choubey∗

Salesforce Research

pchoubey@salesforce.com

Ruihong Huang

Texas A&M University

huangrh@tamu.edu

Abstract

We propose to leverage news discourse proﬁl-

ing to model document-level temporal struc-

tures for building temporal dependency graphs.

Our key observation is that the functional roles

of sentences used for proﬁling news discourse

signify different time frames relevant to a news

story and can, therefore, help to recover the

global temporal structure of a document. Our

analyses and experiments with the widely used

knowledge distillation technique show that dis-

course proﬁling effectively identiﬁes distant

inter-sentence event and (or) time expression

pairs that are temporally related and otherwise

difﬁcult to locate1.

1 Introduction

Grounding all events and time expressions to a ref-

erence timeline is fundamental to text understand-

ing. Recently, Yao et al. (2020) proposed a new

task and dataset for building temporal dependency

graph (TDG)

. TDG is based on the notion of nar-

rative time and temporal anaphora, and references

each timex to a timex or a meta node and each

event to a timex and maybe an event. The reference

timex of an event is either the smallest time (when

identiﬁable) that encloses the event or the docu-

ment creation time (DCT). Similarly, the reference

event is selected such that it gives the most precise

temporal interpretation for a child event.

Because each event and timex is referenced to

only one timex (or additionally an event), identi-

ﬁed temporal relations represent the most salient

relations that can potentially be used to infer ad-

ditional temporal relations through transitivity or

commonsense reasoning (Yao et al.,2020). This

makes identifying reference timex and reference

∗Work done while at Texas A&M University

Code is available at

https://github.com/

prafulla77/Discourse_TDG_AACL2022

The dataset was obtained from

https://github.

com/Jryao/temporal_dependency_graphs_

crowdsourcing

Figure 1: Temporal structures induced by different con-

tent types from the News Discourse Proﬁling.

event more challenging, especially when they are

mentioned across sentences. Human evaluations by

(Yao et al.,2020) also found that identifying the ap-

propriate reference timex and reference event was

the most challenging aspect of their annotation.

In this work, we focus on improving cross-

sentence reference timex and event mentions identi-

ﬁcation by exploring discourse-level temporal cues.

We choose the news discourse proﬁling structure

(DP) (Choubey et al.,2020). DP classiﬁes sen-

tences in a news document into one of eight content

types, deﬁned based on the functional role of a sen-

tence in describing the main news story (Teun A,

1986;Van Dijk,1988a,b;Choubey et al.,2020),

and provides an event-based functional interpreta-

tion of sentences. The eight content types include

main, consequence, previous event, current context,

historical, anecdotal, evaluation and expectation.

As shown in Figure 1, different content types in-

duce different time frames relevant to a news story

that can be beneﬁcial for the global interpretation of

temporal orders among event and timex mentions.

For instance, mentions in historical sentences have

a temporal adjacency with other mentions in histor-

ical sentences but are likely to be distant from men-

tions in other content types. Similarly, mentions

in previous event sentences may have a temporal

adjacency with mentions from one of the previous

event, main event or current-context sentences but

are likely to be separated from mentions in any

arXiv:2210.11787v1 [cs.CL] 21 Oct 2022

of the historical, expectation or consequence sen-

tences.

We ﬁrst summarize the distributional association

between the position of reference mentions and dis-

course content types in

2.3. Then, we propose

a knowledge distillation-based method to incorpo-

rate discourse knowledge into the TDG system.

We experiment with the BERT Devlin et al. (2019)

and RoBERTa Liu et al. (2019) pre-trained lan-

guages models and ﬁnd that the proposed knowl-

edge distillation-based TDG system is effective in

using discourse-level cues and achieves improved

performance on identifying cross-sentence refer-

ence mentions while retaining performance on the

intra-sentence mention pairs.

2 Background and Analysis

2.1 News Discourse Proﬁling (DP)

Following the news content schemata proposed by

Van Dijk (Teun A,1986;Van Dijk,1988a,b), DP

(Choubey et al.,2020) deﬁnes eight content types.

Each content type describes the functional role of a

sentence in describing the main news event. Main

event (M1) sentence describes the major events and

subjects of the news article. Consequence (M2) de-

scribes events that are triggered by the main event.

Previous Event (C1) describes recent events that are

a possible cause of the main event. Current Context

(C2) describes remaining contextual information.

Historical Event (D1) describes past events that

precede the main events in months and years, Anec-

dotal Event (D2) describes unveriﬁable facts, Eval-

uation (D3) describes opinionated contents from

immediate participants, experts or journalists, and

Expectation (D4) describes speculations or possible

consequences of the main or context events.

2.2 Temporal Dependency Graph (TDG)

TDG (Yao et al.,2020) is a directed edge-labeled

graph in which each node is either an event, a

timex, or a meta node (e.g. document creation

time). The reference for each timex/event node is

another timex node or a meta node. Optionally,

the temporal position of some events can be more

precisely determined by referencing them to an-

other event, and thus they can also have a reference

event node. For instance, in Figure 2, the event

incident can only be temporally positioned with

respect to the timex August 23 while the tempo-

ral order of event broke can be determined with

respect to both the timex later and the event oc-

Figure 2: An example TDG.

curred. The edges between event/ timex node pairs

are labeled with one of the overlap,after,before

and included temporal relations while the edges

between a timex node and a meta node is assigned

a generic depend-on label. In this work, we focus

exclusively on identifying the reference timex (and

event) for each timex (event) without predicting the

temporal relations between them.

2.3 Analysis of TDG Structures w.r.t. DP

Sentence Types

As illustrated in Figure 1, discourse roles have tem-

poral interpretations that are useful to locate event

and timex relations in a document. Therefore, we

use the recently proposed discourse proﬁling sys-

tem by Choubey and Huang (2021)

to assign con-

tent type labels to all sentences in the training data

and analyze the distribution of reference timex and

event mentions across different content types. Note

that our analyses are based on a neural network

model-predicted discourse content types which are

noisy. Additionally, a sentence often contains more

than one event and timex mentions and its content

type can only provide a broad temporal ordering

for constituent mentions.

First, we observe that reference timex for both

timex (66% to 100%) and event (54% to 80%) men-

tions from all content types, except the historical, is

majorly the DCT. Further, among the events from

non-historical sentences that are not referenced to

DCT, we observe that majority (71% to 89%) of

them are referenced to a time expression from main,

The discourse proﬁling system was obtained from

https://github.com/prafulla77/Discoure_

Profiling_RL_EMNLP21Findings.

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：

ModelingDocument-levelTemporalStructuresforBuildingTemporalDependencyGraphsPrafullaKumarChoubeySalesforceResearchpchoubey@salesforce.comRuihongHuangTexasA&MUniversityhuangrh@tamu.eduAbstractWeproposetoleveragenewsdiscourseprol-ingtomodeldocument-leveltemporalstruc-turesforbuildingtemporaldependenc...

展开>> 收起<<

Modeling Document-level Temporal Structures for Building Temporal Dependency Graphs Prafulla Kumar Choubey.pdf

共9页,预览2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！

Modeling Document-level Temporal Structures for Building Temporal Dependency Graphs Prafulla Kumar Choubey

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: