Parameter-Efﬁcient Tuning with Special Token Adaptation Xiaocong Yangy James Y. Huangz Wenxuan Zhouzand Muhao Chenz yTsinghua UniversityzUniversity of Southern California

2025-05-02 0 0 981.22KB 8 页 10玖币

侵权投诉

Parameter-Efﬁcient Tuning with Special Token Adaptation

Xiaocong Yang†∗

, James Y. Huang‡, Wenxuan Zhou‡and Muhao Chen‡

†Tsinghua University; ‡University of Southern California

yangxc.18@sem.tsinghua.edu.cn;{huangjam,zhouwenx,muhaoche}@usc.edu

Abstract

Parameter-efﬁcient tuning aims at updating

only a small subset of parameters when adapt-

ing a pretrained model to downstream tasks.

In this work, we introduce PASTA, in which

we only modify the special token representa-

tions (e.g., [SEP] and [CLS] in BERT) be-

fore the self-attention module at each layer in

Transformer-based models. PASTA achieves

comparable performance to full ﬁnetuning in

natural language understanding tasks includ-

ing text classiﬁcation and NER with up to only

0.029% of total parameters trained. Our work

not only provides a simple yet effective way of

parameter-efﬁcient tuning, which has a wide

range of practical applications when deploying

ﬁnetuned models for multiple tasks, but also

demonstrates the pivotal role of special tokens

in pretrained language models.1

1 Introduction

Built upon a pretrained language model (PLM; De-

vlin et al. 2019;Liu et al. 2019;Yang et al. 2019;

Chowdhery et al. 2022), many of the recent NLP

systems are developed based on task-speciﬁc ﬁne-

tuning. In this way, the PLM effectively leverages

the task-agnostic knowledge captured during self-

supervised pretraining and adapts itself to down-

stream tasks. However, full ﬁnetuning poses a

challenge to model deployment under multi-task,

memory-limited scenarios, where we need to train

and store a separate full-sized model for each sub-

stantially distinct task. As an alternative, parameter-

efﬁcient tuning (Ding et al.,2022) aims at only up-

dating a small number of parameters when adapting

PLMs to downstream tasks while making most of

the model parameters ﬁxed and shared among tasks,

thus reducing memory usage.

In this paper, we propose

rameter-efﬁcient

tuning with

pecial

oken

daptation (PASTA),

∗

Work done when visiting USC.

Our code is publicly available at:

https://github.

com/luka-group/PASTA/

L5-H3 L5-H5 L5-H6 L5-H7

L20-H3 L20-H5 L20-H8 L20-H15

Figure 1: Examples of vertical attention heads in the 5-

th and 20-th layer of BERT-large with a random sample

from CoLA (Warstadt et al.,2019) as input. Heads in

the ﬁrst row and second row assign most of maximal

attention weights to [CLS] and [SEP] respectively.

See Appx. §Cfor the full attention map.

where we only add trainable vectors to hidden rep-

resentations of special tokens

at each layer before

the multi-head attention module in Transformer-

based PLMs. Our work is motivated by the role

of special tokens in PLMs. First, special tokens

such as

[CLS]

collect information from the whole

input sequence and are typically regarded as the

global text representation (Devlin et al.,2019). For

sentence-level tasks such as GLUE (Wang et al.,

2018), a common practice is to add a new classiﬁer

head based on the

[CLS]

representation in the last

model layer. Thus, if trained properly, by updating

the

[CLS]

representations, we can approximate

the result of the information collection process in

PLMs. Second, many attention heads in PLMs fol-

low a vertical pattern

, where the attention scores

are mostly allocated to either the

[CLS]

[SEP]

token (Clark et al.,2019;Kovaleva et al.,2019), as

WLOG, we use the notation of special tokens

[CLS]

and

[SEP]

in BERT for the convenience of expression, while the

method applies in the same way to other paradigms such as

<S> and </S> in RoBERTa (Liu et al.,2019).

Following Voita et al. (2019) and Yao et al. (2021), an

attention head is regarded as vertical if at least 90% tokens

assign maximal attention scores to either [CLS] or [SEP].

arXiv:2210.04382v2 [cs.CL] 14 Feb 2023

illustrated in Fig. 1. Therefore, updates to special

tokens can also be disseminated to other tokens dur-

ing the forward pass through the vertical attention

heads (Elhage et al.,2021), enabling the PLMs to

adapt to both sentential and lexical tasks.

By tuning as few as up to 0.029% of the total

parameters, PASTA achieves competitive perfor-

mance on par with full ﬁnetuning and BitFit (Za-

ken et al.,2022) on GLUE (§4.2). It also outper-

forms P-tuning v2 (Liu et al.,2022) by 0.6% on

CoNLL2003 with 20

fewer additional parame-

ters (§4.3). The ablation study shows that we can

further reduce trainable parameters to 0.009% with

only a slight performance drop (§4.4), showing the

merit of adapting special token representations.

2 Related Work

A recent survey (Ding et al.,2022) categorizes three

types of parameter-efﬁcient tuning methods. Ad-

dition methods (Houlsby et al.,2019;Lester et al.,

2021;Liu et al.,2022) introduce a small number

of additional trainable parameters while keeping

those in the PLM unchanged. Speciﬁcation meth-

ods (Zaken et al.,2022;Guo et al.,2021;Zhao

et al.,2020) update a portion of parameters in the

PLM while keeping others frozen. Reparameteri-

zation methods (Aghajanyan et al.,2021;Hu et al.,

2021;Qin et al.,2021) modify PLMs’ structures to

parameter-efﬁcient forms. Our method belongs to

the addition-based methods and follows the basic

settings of P-tuning v2 (Liu et al.,2022), where

newly initialized hidden representations of tokens

are inserted into each Transformer layer. Different

from most prompt tuning methods that introduce

new tokens, we add the introduced vectors to the

hidden states of special tokens and keep the se-

quence length unchanged.

Previous works use probing tasks (Wu et al.,

2020) and pruning methods (Prasanna et al.,2020)

to study the roles of different modules inside BERT.

It has been shown that functional specialization

exists in BERT self-attention heads (Clark et al.,

2019), and vertical attention heads

take up a large

portion (Yao et al.,2021). Kovaleva et al. (2019)

ﬁnd that vertical attention heads are almost ex-

clusively associated with attention to

[SEP]

[CLS]

tokens, and Clark et al. (2019) conclude

that heads in early layers often attend to

[CLS]

while in middle layers attend to

[SEP]

. In this

work, we demonstrate that adapting hidden repre-

sentations of special tokens is sufﬁcient to bring the

Multi-Head Attention

h([CLS]) Sentence1

Frozen

Trainable

e(𝑣𝟏

𝒍)

h([SEP]) Sentence2 h([SEP])

e(𝑣𝟐

𝑙)e(𝑣𝟑

𝑙)

Add & Norm

Feed-Forward Network

Add & Norm

Figure 2: Architecture of PASTA layer in Transformer.

Skip-connections in Transformers are not shown for

brevity. At layer lwe add a trainable vector e(vl

p)∈

Rdto the hidden representation of the p-th special to-

ken in the input sequence, and freeze the weights of the

PLM.

performance of PLMs to the level of full ﬁnetuning.

3 PASTA

Given a large PLM, our goal is to develop a

parameter-efﬁcient tuning method that only up-

dates a small set of parameters when adapting to a

downstream task. To this end, we propose a simple

yet effective method called PASTA, in which we

train a hidden vector for every special token at each

Transformer layer, along with a task-speciﬁc clas-

siﬁer, while freezing the parameters of the PLM.

3.1 Special Token Adaptation

The special token adaption is illustrated in Fig. 2.

Although these adaptations are not directly applied

to non-special tokens, changes in special token hid-

den states can be effectively disseminated to other

tokens via self-attention during forward passes,

thanks to the prevalence of vertical attention heads

in PLMs.

Speciﬁcally, denote the inputs to the

-th Trans-

former layer as

Hl={hl

i}N

i=1,hl

i∈Rd

, where

is the number of input tokens,

is the hidden size,

PASTA modiﬁes the inputs as follows:

mod ={hl

i+ml

i}N

i=1,

Hl+1 =Trml(Hl

mod),

where

Trml

is the

-th Transformer layer,

i∈

is our special token adaptation deﬁned as fol-

lows:

i=(0if token iis not a special token

e(vl

p)if token iis the p-th special token

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：

Parameter-EfcientTuningwithSpecialTokenAdaptationXiaocongYangy,JamesY.Huangz,WenxuanZhouzandMuhaoChenzyTsinghuaUniversity;zUniversityofSouthernCaliforniayangxc.18@sem.tsinghua.edu.cn;{huangjam,zhouwenx,muhaoche}@usc.eduAbstractParameter-efcienttuningaimsatupdatingonlyasmallsubsetofparameterswhena...

展开>> 收起<<

Parameter-Efﬁcient Tuning with Special Token Adaptation Xiaocong Yangy James Y. Huangz Wenxuan Zhouzand Muhao Chenz yTsinghua UniversityzUniversity of Southern California.pdf

共8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！

Parameter-Efﬁcient Tuning with Special Token Adaptation Xiaocong Yangy James Y. Huangz Wenxuan Zhouzand Muhao Chenz yTsinghua UniversityzUniversity of Southern California

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: