Discourse Analysis via Questions and Answers Parsing Dependency Structures of Questions Under Discussion Wei-Jen Ko1Yating Wu2Cutter Dalton3Dananjay Srinivas3

2025-04-26 0 0 864.97KB 13 页 10玖币

侵权投诉

Discourse Analysis via Questions and Answers:

Parsing Dependency Structures of Questions Under Discussion

Wei-Jen Ko1Yating Wu2Cutter Dalton3Dananjay Srinivas3

Greg Durrett1Junyi Jessy Li4

1Computer Science, 2Electrical and Computer Engineering, 4Linguistics,

The University of Texas at Austin

3Linguistics, University of Colorado Boulder

wjko@outlook.com, yating.wu@utexas.edu, cutter.dalton@colorado.edu, dananjay.srinivas@gmail.com,

gdurrett@cs.utexas.edu, jessy@utexas.edu

Abstract

Automatic discourse processing is bottle-

necked by data: current discourse formalisms

pose highly demanding annotation tasks in-

volving large taxonomies of discourse rela-

tions, making them inaccessible to lay anno-

tators. This work instead adopts the linguis-

tic framework of Questions Under Discussion

(QUD) for discourse analysis and seeks to

derive QUD structures automatically. QUD

views each sentence as an answer to a ques-

tion triggered in prior context; thus, we charac-

terize relationships between sentences as free-

form questions, in contrast to exhaustive ﬁne-

grained taxonomies. We develop the ﬁrst-

of-its-kind QUD parser that derives a depen-

dency structure of questions over full docu-

ments, trained using a large, crowdsourced

question-answering dataset DCQA (Ko et al.,

2022). Human evaluation results show that

QUD dependency parsing is possible for lan-

guage models trained with this crowdsourced,

generalizable annotation scheme. We illustrate

how our QUD structure is distinct from RST

trees, and demonstrate the utility of QUD anal-

ysis in the context of document simpliﬁcation.

Our ﬁndings show that QUD parsing is an ap-

pealing alternative for automatic discourse pro-

cessing.

1 Introduction

Discourse structure characterizes how each sen-

tence in a text relates to others to reﬂect the au-

thor’s high level reasoning and communicative in-

tent. Understanding discourse can be widely use-

ful in applications such as text summarization (Hi-

rao et al.,2013;Gerani et al.,2014;Durrett et al.,

2016;Xu et al.,2020), classiﬁcation (Bhatia et al.,

2015;Ji and Smith,2017), narrative understand-

ing (Lee and Goldwasser,2019), machine compre-

hension (Narasimhan and Barzilay,2015), etc.

However, automatically inferring discourse

structure is challenging which hinders wider appli-

cation (Atwell et al.,2021). At its root lies the issue

of data annotation: popular coherence formalisms

like the Rhetorical Structure Theory (RST, Mann

and Thompson (1988), Segmented Discourse Rep-

resentation Theory (SDRT, Asher et al. (2003), and

the Penn Discourse Treebank (PDTB, Prasad et al.

(2008) require experts—typically linguists trained

for the task—to reason through long documents

over large relation taxonomies. These features,

coupled with the difﬁculties of annotating full struc-

tures in the case of RST and SDRT, make the task

inaccessible to lay annotators. The taxonomies dif-

fer across formalisms (Demberg et al.,2019), and

their coverage and deﬁnitions are being actively re-

searched and reﬁned (Sanders et al.,1992;Taboada

and Mann,2006;Prasad et al.,2014).

In contrast, this work aims to derive discourse

structures that ﬁt into the linguistic framework

of Questions Under Discussion (QUD) (Von Stut-

terheim and Klein,1989;Van Kuppevelt,1995),

which neatly avoids reliance on a strict taxonomy.

In QUD, “each sentence in discourse addresses a

(often implicit) QUD either by answering it, or

by bringing up another question that can help an-

swering that QUD. The linguistic form and the

interpretation of a sentence, in turn, may depend

on the QUD it addresses” (Benz and Jasinskaja,

2017). Thus relationships between sentences can

be characterized by free-form questions instead of

pre-deﬁned taxonomies. For instance, consider the

following two sentences:

(S3)

: A route out of Sarajevo was expected to open

later today — but only for international humanitarian

agencies that already can use another route.

(S6)

: A four-month cease-ﬁre agreement signed Dec.

31 made possible the medical evacuation and opening

of the route into Sarajevo today.

Sentence 6 is the answer to a question from sen-

tence 3: “Why can they open a route?”. The

question-answer view is in line with recent work

arXiv:2210.05905v2 [cs.CL] 12 May 2023

[1] California legislators, searching for ways to pay for the $4 billion to $6 billion in damages from last week's earthquake, are laying

the groundwork for a temporary increase in the state's sales tax. [2] The talk of a sales tax rise follows a rebuff from Congress on the

question of how much the federal government is willing to spend to aid in California's earthquake relief efforts. [3] The state had sought

as much as $4.1 billion in relief, but yesterday the House approved a more general scaled-back measure calling for $2.85 billion in aid,

the bulk of which would go to California, with an unspecified amount going to regions affected by Hurricane Hugo. [4] That leaves the

state roughly $2 billion to $4 billion short. [5] A sales tax increase appears to be the fastest and easiest to raise funds in a hurry.

[6] According to the state department of finance, a one-penny increase in the state's six-cent per dollar sales tax could raise $3 billion.

1-2 What is prompting

the California legislators

to seek alternative forms

of relief?

1-3 What type of relief

is being considered?

3-4 What would the

shortfall be?

2-5 What would be the

advantage of the proposed tax?

2-6 What would be the impact

of the proposed increase?

Summary

background

elaboration-

additional

result

hypothetical

(a) (b)

Figure 1: A snippet of a WSJ article from the intersecting subset of DCQA (Ko et al.,2022) and the RST Discourse

Treebank (Carlson et al.,2001). (a) shows a QUD dependency structure derived from DCQA. Edges are deﬁned

by questions, connecting where the question arose from (the “anchor” sentence) and the sentence that answers the

question. (b) shows the annotated RST tree above the sentence level.

reformulating linguistic annotation as question an-

swering (He et al.,2015;Pyatkin et al.,2020;Klein

et al.,2020), which reduces the bar for data collec-

tion and allows advancements in QA systems to be

recruited (Aralikatte et al.,2021). Furthermore,

QUD’s reliance on natural language annotation

aligns with large language models (e.g., GPT-3)

using language as a universal “interface” across

various tasks.

Despite the richness in theoretical research re-

lated to QUD, data-driven efforts are scarce; re-

cent work has started corpora development under

QUD (De Kuthy et al.,2018;Westera et al.,2020;

Hesse et al.,2020), but these dedicated datasets are

small and no computational models have yet been

built to automatically derive QUD structures.

This work seeks to ﬁll this gap, and presents the

ﬁrst-of-its-kind QUD parser. This parser takes a

document as input and returns a question-labeled

dependency structure over the sentences in the

document, as depicted in Figure 1(a). For train-

ing, we use the intra-document question answering

dataset DCQA (Ko et al.,2022); DCQA’s anno-

tation scheme is both compatible with QUD and

easily crowdsourced, making QUD parsing a much

less costly option than existing frameworks.

Each question in DCQA is considered to arise

from an “anchor” sentence, and answered by an-

other sentence later in the same article. In line with

QUD, we consider each sentence as the answer to

an implicit question from prior context (Hunter and

Abrusán,2015), in particular the anchor sentence.

We view the anchor sentence as the parent node

of the answer sentence, with the question describ-

ing the relation between the two; this results in a

dependency tree structure.

Conveniently, a subset of DCQA overlaps with

the RST Discourse Treebank (Carlson et al.,2001),

allowing us to directly compare the two types of

structures (Figure 1(b)). We show that the QUD

trees are structurally distinct from RST trees. A

close inspection of relation-question correspon-

dence reveals that QUD’s free-form questions are

more ﬁne-grained, and that their presence reduces

annotator disagreement in selecting RST relations.

Trained on DCQA, our QUD parser consists of

two models used in a pipeline. The ﬁrst model

predicts the anchor sentence for each (answer) sen-

tence in the article; the second model performs

question generation given the answer sentence and

the predicted anchor sentence. Our comprehen-

sive human evaluation shows that readers approve

of 71.5% of the questions generated by our best

model; among those, the answer sentence answers

the generated question 78.8% of the time. Finally,

we demonstrate the analytical value of QUD analy-

sis in the context of news document simpliﬁcation:

the questions reveal how content is elaborated and

reorganized in simpliﬁed texts.

In sum, this work marks the ﬁrst step in QUD

parsing; our largely positive human evaluation re-

sults show that this is a promising data-driven ap-

proach to discourse analysis with open, crowd-

sourced annotation that is so far infeasible to

do at scale with other discourse frameworks.

We release our models at

https://github.com/

lingchensanwen/DCQA-QUD-parsing.

2 Background and related work

Discourse frameworks

Questions Under Dis-

cussion is a general framework with vast theo-

retical research especially in pragmatics, e.g., in-

formation structure (Roberts,2012;Büring,2003;

Velleman and Beaver,2016), presuppositions (Si-

mons et al.,2010), and implicature (Hirschberg,

1985;Van Kuppevelt,1996;Jasinskaja et al.,2017).

Ginzburg et al. (1996) extended Stalnaker (1978)’s

dynamic view of context to dialogue by integrating

QUD with dialogue semantics, where the speakers

are viewed as interactively posing and resolving

queries. In QUD analysis of monologue, each sen-

tence aims to answer a (mostly implicit) question

triggered in prior context. Sometimes the questions

form hierarchical relationships (stacks where larger

questions have sub-questions, starting from the root

question “What is the way things are?”) (Büring,

2003;Roberts,2004;De Kuthy et al.,2018;Ri-

ester,2019). However, because of the inherent

subjectivity among naturally elicited QUD ques-

tions (Westera et al.,2020;Ko et al.,2020), we

leave question relationships for future work.

QUD and coherence structures are closely re-

lated. Prior theoretical work looked into the map-

ping of QUDs to discourse relations (Jasinskaja

et al.,2008;Onea,2016) or the integration of the

two (Kuppevelt,1996). Hunter and Abrusán (2015)

and Riester (2019) studied structural correspon-

dances between QUD stacks and SDRT specif-

ically. Westera et al. (2020) showed that QUD

could be a useful tool to quantitatively study the

predictability of discourse relations (Garvey and

Caramazza,1974;Kehler et al.,2008;Bott and

Solstad,2014). In Pyatkin et al. (2020), discourse

relation taxonomies were also converted to tem-

platic questions, though not in the QUD context.

Traditionally, discourse “dependency parsing”

refers to parsing the RST structure (Hirao et al.,

2013;Bhatia et al.,2015;Morey et al.,2018). Since

QUD structures are marked by free-form questions,

the key aspect of “parsing” a QUD structure is

thus question generation, yielding a very different

task and type of structure than RST parsing. As

we show in the paper, the two are complementary

to each other and not comparable. This work fo-

cuses on automating and evaluating a QUD parser;

we leave for future work to explore what types of

structure is helpful in different downstream tasks.

The DCQA dataset

Corpora speciﬁc for QUD

are scarce. Existing work includes a handful of in-

terviews and 40 German driving reports annotated

with question stacks (De Kuthy et al.,2018;Hesse

et al.,2020), as well as Westera et al. (2020)’s 6

TED talks annotated following Kehler and Rohde

(2017)’s expectation-driven model (eliciting ques-

tions without seeing upcoming context). Ko et al.

(2020)’s larger INQUISITIVE question dataset is an-

notated in a similar manner, but INQUISITIVE only

provides questions for the ﬁrst 5 sentences of an

article, and they did not annotate answers.

This work in contrast repurposes the much larger

DCQA

dataset (Ko et al.,2022), consisting of more

than 22K questions crowdsourced across 606 news

articles. DCQA was proposed as a way to more

reliably and efﬁciently collect data to train QA sys-

tems to answer high-level questions, speciﬁcally

QUD questions in INQUISITIVE. Though not orig-

inally designed for QUD parsing, DCQA is suit-

able for our work because its annotation procedure

follows the reactive model of processing that is

standard in QUD analysis (Benz and Jasinskaja,

2017), where the questions are elicited after ob-

serving the upcoming context. Concretely, for each

sentence in the article, the annotator writes a QUD

such that the sentence is its answer, and identiﬁes

the “anchor” sentence in preceding context that the

question arose from. Figure 1(a) shows questions

asked when each of the sentences 2-6 are consid-

ered as answers, and their corresponding anchor

sentences. As with other discourse parsers, ours

is inevitably bound by its training data. However,

DCQA’s crowdsourcable paradigm makes future

training much easier to scale up and generalize.

3 Questions vs. coherence relations

We ﬁrst illustrate how questions capture inter-

sentential relationships, compared with those in

coherence structures. We utilize the relation taxon-

omy in RST for convenience, as in Section 5.3 we

also compare the structure of our QUD dependency

trees with that of RST.

Given each existing anchor-answer sentence pair

across 7 DCQA documents, we asked two gradu-

ate students in Linguistics to select the most ap-

propriate discourse relation between them (from

the RST relation taxonomy (Carlson and Marcu,

2001)). Both students were ﬁrst trained on the

taxonomy using the RST annotation manual.

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：

DiscourseAnalysisviaQuestionsandAnswers:ParsingDependencyStructuresofQuestionsUnderDiscussionWei-JenKo1YatingWu2CutterDalton3DananjaySrinivas3GregDurrett1JunyiJessyLi41ComputerScience,2ElectricalandComputerEngineering,4Linguistics,TheUniversityofTexasatAustin3Linguistics,UniversityofColoradoBoulderw...

展开>> 收起<<

Discourse Analysis via Questions and Answers Parsing Dependency Structures of Questions Under Discussion Wei-Jen Ko1Yating Wu2Cutter Dalton3Dananjay Srinivas3.pdf

共13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！

Discourse Analysis via Questions and Answers Parsing Dependency Structures of Questions Under Discussion Wei-Jen Ko1Yating Wu2Cutter Dalton3Dananjay Srinivas3

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: