Zero-shot stance detection based on cross-domain feature enhancement by contrastive learning Xuechen ZhaoJiaying ZouZhong ZhangFeng XieBin Zhou Lei Tian

2025-04-29 0 0 1.67MB 9 页 10玖币

侵权投诉

Zero-shot stance detection based on cross-domain

feature enhancement by contrastive learning

Xuechen Zhao∗Jiaying Zou∗Zhong Zhang∗Feng Xie∗Bin Zhou∗ † Lei Tian∗

Abstract

Zero-shot stance detection is challenging because it requires

detecting the stance of previously unseen targets in the infer-

ence phase. The ability to learn transferable target-invariant

features is critical for zero-shot stance detection. In this

work, we propose a stance detection approach that can eﬃ-

ciently adapt to unseen targets, the core of which is to cap-

ture target-invariant syntactic expression patterns as trans-

ferable knowledge. Speciﬁcally, we ﬁrst augment the data by

masking the topic words of sentences, and then feed the aug-

mented data to an unsupervised contrastive learning mod-

ule to capture transferable features. Then, to ﬁt a speciﬁc

target, we encode the raw texts as target-speciﬁc features.

Finally, we adopt an attention mechanism, which combines

syntactic expression patterns with target-speciﬁc features to

obtain enhanced features for predicting previously unseen

targets. Experiments demonstrate that our model outper-

forms competitive baselines on four benchmark datasets.

1 Introduction

The goal of stance detection is to automatically identify

the attitude or stance (e.g., Favor, Against, or Neutral)

expressed in a text towards a speciﬁc target or topic1

[3, 23, 16]. The traditional target-speciﬁc stance detec-

tion assumes that the training and testing data belonged

to the same target [18]. However, due to the continuous

emergence of unseen targets, collecting data on all tar-

gets for training is infeasible in practice. Moreover, it is

expensive to obtain high-quality labels for a new target

[23]. Therefore, the study of zero-shot stance detection

for unseen targets goes beyond the target-speciﬁc task

and can help to predict stance more ﬂexibly.

For the zero-shot stance detection task, some ex-

isting approaches try to improve the model’s predic-

tive ability for unseen targets by employing attention

mechanisms [1, 33] or fusing external knowledge [22].

∗School of Computer, National University of Defense Technol-

ogy, Changsha, China. {zhaoxuechen, zoujiaying20, zhangzhong,

xiefeng, binzhou, leitian129}@nudt.edu.cn

†Corresponding Author, Key Lab. of Software Engineering for

Complex Systems, Changsha, China.

1In the paper, we will use the terms: target and topic

interchangeably.

However, transferring knowledge directly from a spe-

ciﬁc target to an unseen target is often limited in its

predictive eﬀectiveness due to the coupling of target-

speciﬁc features. [2, 32] use adversarial learning to

guide the model to learn target-invariant features via

discriminators, which may lead to degraded prediction

performance in the unbalanced distribution of targets.

[19] capture the target-invariant features by identifying

the stance feature categories and supervised contrastive

learning, so their model achieves a better generalization

capability. However, the data needs to be tagged with

soft labels by pretext tasks, which increases the com-

plexity of the model and brings some noise to the data.

Target-speciﬁc features are directly related to a speciﬁc

target, while target-invariant features are generic and

transferable, regardless of their targets. Consequently,

it is crucial to distinguish these two features when pre-

dicting the stance of a text on unseen targets.

Both linguistic and psychological ﬁelds divide lan-

guage into two aspects of representation: 1) syntactic

representation and 2) semantic representation. The for-

mer reﬂects the form of languages, such as word mor-

phology and sentence structure, and is the external rep-

resentation of language; the latter is the concept and

proposition, denoting the meaning referred to by the

form of language, which is abstract and is the inter-

nal representation of language [15]. Meanwhile, Event-

related Potentials (ERP) interaction theory [6] suggests

that text semantics is a fusion of syntactic and seman-

tic representations, where syntax and semantics inter-

act to complete the process of sentence comprehen-

sion and expression jointly. As shown in Table 1, it

is possible to use the same or similar syntactic expres-

sion patterns even for sentences with diﬀerent targets,

i.e., although the targets of Example 1 and Example 2

are distinct, they both use rhetorical question expres-

sion patterns, so these syntactic expression patterns are

target-invariant. The target-invariant syntactic expres-

sion patterns and the target-speciﬁc features jointly de-

termine the sentence’s meaning. Inspired by this, we

acquire syntactic expression patterns, which are natu-

rally target-invariant and have an important impact on

semantics. Furthermore, these syntactic expression pat-

arXiv:2210.03380v1 [cs.CL] 7 Oct 2022

Example 1: Target: Climate Change is a Real Concern Stance: Favor

Sentence: Today Europe is breaking heat records, while Asia is breaking the lowest temperature records!

Should we not be concerned?

Masked: Today Europe is [MASK], while Asia is [MASK]! Should we not be [MASK]?

Example 2: Target: Feminist Movement Stance: Favor

Sentence: When they say men look at women like a piece of meat, what do they even mean, they want

to cook & eat her?

Masked: When they say men look at [MASK] like [MASK] what do they even mean, they want to

[MASK]?

Table 1: Examples of syntactic expression patterns.

terns can combine with target-speciﬁc features to eﬀec-

tively predict the text stances.

More concretely, we propose a Feature Enhanced

Zero-shot Stance Detection Model via Contrastive

Learning (FECL). First, we capture syntactic expres-

sion patterns via contrastive learning as transferable fea-

tures. Second, based on supervised learning, we fully

use the labeled data to learn semantic information as

target-speciﬁc features. At last, a feature fusion mod-

ule fuses the target-invariant and the target-speciﬁc fea-

tures to achieve the cross-target prediction capability.

The main contributions of this paper can be sum-

marized as follows:

•We explore a novel self-supervised feature learning

scheme. The scheme augments the raw texts

by masking their topic words and then adopts

contrastive learning to capture syntactic expression

patterns as a bridge for knowledge transferring.

•We model the text stance expression into two parts:

syntactic and semantic expression, and adopt an

attention-based feature fusion mechanism that con-

siders both target-invariant and target-speciﬁc fea-

tures of texts. This mechanism improves the qual-

ity of feature representation and allows the model

to handle prediction tasks such as zero-shot, cross-

target, etc.

•Extensive experiments on four benchmark datasets

show that the proposed model performs well on the

zero-shot stance detection task. We also extend the

model to few-shot and cross-target stance detection

tasks, demonstrating the superiority and general-

ization of our approach.

2 Related Work

Stance Detection. Stance detection aims to ex-

amine the attitude of a text towards a given target

[13]. It is widely applied in argument mining [27], fake

news/rumor detection [17], fact-checking [28], and epi-

demic trend prediction [11]. Recently, with the rapid de-

velopment of social media, various unseen targets have

emerged, which also brings new challenges to stance de-

tection. Therefore, cross-target and zero-shot stance de-

tection have received extensive attention. Cross-target

stance detection is training on source targets and adapt-

ing to unseen but correlated targets [20]. While zero-

shot stance detection aims to train on multiple tar-

gets with labels and then automatically identify an un-

seen target [19]. Their core ideas are learning target-

invariant features and transferring them to unseen tar-

gets. [3] modeled the features of unseen targets via Bi-

Condition LSTM. [33] extracted shared features based

on a self-attention neural network. [2] applied a target-

speciﬁc stance dataset [23] to zero-shot stance detec-

tion and used adversarial learning to capture target-

invariant features. [22] proposed a Bert-based [9] com-

monsense knowledge-enhanced graph model to zero-

shot stance detection. These works try to extract trans-

ferable features from source targets and apply them to

unseen targets but ignore the most basic syntactic ex-

pression patterns, which is a natural target-invariant

and an essential factor that aﬀects semantics and can

eﬀectively detect stance with the features of speciﬁc tar-

gets.

Contrastive Learning. Contrastive learning is a

self-supervised representation learning method initially

proposed in computer vision. It aims to learn distinct

representations by pulling semantically close neighbors

together and pushing non-neighbors away [12]. Recent

studies have attempted to aid or enhance the perfor-

mance of natural language processing tasks by incor-

porating contrastive learning [24, 10, 5]. Contrastive

learning has proven comparable to supervised learning

in many domains and eﬀectively improves the quality

of feature learning. However, most traditional stance

detection algorithms only use supervised stance label

information and inadequately utilize the richer informa-

tion in the unlabeled text. Some researches [19, 21] be-

gan to apply contrastive learning to stance detection to

improve the feature representation ability of the model.

Therefore, our proposed model uses both the existing

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：

Zero-shotstancedetectionbasedoncross-domainfeatureenhancementbycontrastivelearningXuechenZhao*JiayingZou*ZhongZhang*FengXie*BinZhou*LeiTian*AbstractZero-shotstancedetectionischallengingbecauseitrequiresdetectingthestanceofpreviouslyunseentargetsintheinfer-encephase.Theabilitytolearntransferabletarg...

展开>> 收起<<

Zero-shot stance detection based on cross-domain feature enhancement by contrastive learning Xuechen ZhaoJiaying ZouZhong ZhangFeng XieBin Zhou Lei Tian.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！

Zero-shot stance detection based on cross-domain feature enhancement by contrastive learning Xuechen ZhaoJiaying ZouZhong ZhangFeng XieBin Zhou Lei Tian

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: