Towards Domain-Independent Supervised Discourse Parsing Through Gradient Boosting Patrick Huber and Giuseppe Carenini

2025-05-06 0 0 399.19KB 5 页 10玖币

侵权投诉

Towards Domain-Independent Supervised Discourse Parsing

Through Gradient Boosting

Patrick Huber and Giuseppe Carenini

Department of Computer Science

University of British Columbia

Vancouver, BC, Canada, V6T 1Z4

{huberpat, carenini}@cs.ubc.ca

1 Introduction

Discourse analysis and discourse parsing have

shown great impact on many important prob-

lems in the ﬁeld of Natural Language Process-

ing (NLP) (e.g., Ji and Smith (2017); Bhatia

et al. (2015); Nejat et al. (2017); Gerani et al.

(2014)). Given the direct impact of discourse

annotations on model performance and inter-

pretability, robustly extracting discourse struc-

tures from arbitrary documents is a key task

to further improve computational models in

NLP. To this end, a variety of complementary

discourse theories have been proposed in the

past, such as the lexicalized discourse frame-

work (Webber et al.,2003), the Segmented Dis-

course Representation Theory (SDRT) (Asher,

1993;Asher et al.,2003), and the Rhetorical

Structure Theory (RST) (Mann and Thomp-

son,1988), with RST focusing on the semantic

and pragmatic structure of complete monologue

documents, as used in this work.

Despite the importance of discourse analysis

and discourse parsing for the ﬁeld of NLP and

the obvious value of the RST discourse theory

for many downstream applications, one major

limitation for a wider application of discourse

information is the severe data sparsity issue (for

instance, the popular RST-DT (Carlson et al.,

2002) and GUM (Zeldes,2017) treebanks do not

exceed a minuscule number of 400 documents).

Furthermore, while the data sparsity issue has

been a long-standing problem, modern, data-

intensive machine learning approaches further

reinforce its severity.

In general, three modelling alternatives have

been established in the current landscape: (i)

Supervised approaches (e.g., Ji and Eisenstein

(2014); Feng and Hirst (2014); Joty et al. (2015);

Li et al. (2016); Wang et al. (2017); Guz

et al. (2020)), performing well in the domain

in which they are trained, however, obtain

severely reduced performance if a domain shift

is present, as shown in Huber and Carenini

(2019,2020). (ii) Distantly supervised models

(e.g., Huber and Carenini (2019,2020); Nishida

and Nakayama (2020); Karimi and Tang (2019);

Liu et al. (2019); Huber et al. (2021); Xiao

et al. (2020)), aiming to overcome the domain

adaptation problem by exploiting large-scale

supervised datasets from context-sensitive aux-

iliary tasks (e.g., sentiment analysis). (iii) Self-

supervised/unsupervised methods (e.g., Zhu

et al. (2020); Koto et al. (2021); Wu et al.

(2020); Kobayashi et al. (2019); Huber and

Carenini (2021,2022)), predicting discourse

from either pre-trained language models, auto-

encoder style frameworks, or by recursively

computing dissimilarity scores.

In this landscape of models aiming to over-

come the data sparsity and domain dependency

of current discourse parsers, we present a new,

supervised paradigm directly tackling the do-

main adaptation issue. Speciﬁcally, we intro-

duce the ﬁrst fully supervised discourse parser

designed to alleviate the domain dependency

through a staged model of weak classiﬁers by

introducing the gradient boosting framework

(Schaal and Atkeson,1995;Drucker et al.,1994;

Schwenk and Bengio,1997;Badirli et al.,2020)

into the process of discourse parsing. Using

the underlying assumption that any discourse

treebank contains a mix of frequently appear-

ing, general discourse features (applicable to

any domain) as well as a number of dataset-

related nuances (which are domain-speciﬁc), we

postulate that a set of weak classiﬁers is likely

to learn increasingly speciﬁc and rare features

of the training data. Using this assumption,

we can reasonably assume that there exists a

threshold of weak classiﬁers, which eﬀectively

separates the general features of discourse from

arXiv:2210.09565v1 [cs.CL] 18 Oct 2022

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：

TowardsDomain-IndependentSupervisedDiscourseParsingThroughGradientBoostingPatrickHuberandGiuseppeCareniniDepartmentofComputerScienceUniversityofBritishColumbiaVancouver,BC,Canada,V6T1Z4{huberpat,carenini}@cs.ubc.ca1IntroductionDiscourseanalysisanddiscourseparsinghaveshowngreatimpactonmanyimportantpr...

展开>> 收起<<

Towards Domain-Independent Supervised Discourse Parsing Through Gradient Boosting Patrick Huber and Giuseppe Carenini.pdf

共5页,预览1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！

Towards Domain-Independent Supervised Discourse Parsing Through Gradient Boosting Patrick Huber and Giuseppe Carenini

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: