Instruction Tuning for Few-Shot Aspect-Based Sentiment Analysis Siddharth Varia1Shuai Wang1Kishaloy Halder1Robert Vacareanu2 Miguel Ballesteros1Yassine Benajiba1Neha Anna John1

2025-04-24 0 0 404.72KB 9 页 10玖币

侵权投诉

Instruction Tuning for Few-Shot Aspect-Based Sentiment Analysis

Siddharth Varia1∗Shuai Wang1∗Kishaloy Halder1∗Robert Vacareanu2†∗

Miguel Ballesteros1Yassine Benajiba1Neha Anna John1

Rishita Anubhai1Smaranda Muresan1Dan Roth1

1AWS AI Labs

2University of Arizona, Tucson, AZ, USA

{siddhvar, wshui, kishaloh, ballemig, benajiy, ranubhai, smaranm, drot}@amazon.com

rvacareanu@arizona.edu

Abstract

Aspect-based Sentiment Analysis (ABSA) is a

ﬁne-grained sentiment analysis task which in-

volves four elements from user-generated texts:

aspect term, aspect category, opinion term, and

sentiment polarity. Most computational ap-

proaches focus on some of the ABSA sub-tasks

such as tuple (aspect term, sentiment polar-

ity) or triplet (aspect term, opinion term, senti-

ment polarity) extraction using either pipeline

or joint modeling approaches. Recently, gener-

ative approaches have been proposed to extract

all four elements as (one or more) quadruplets

from text as a single task. In this work, we take

a step further and propose a uniﬁed framework

for solving ABSA, and the associated sub-tasks

to improve the performance in few-shot scenar-

ios. To this end, we ﬁne-tune a

model with

instructional prompts in a multi-task learning

fashion covering all the sub-tasks, as well as

the entire quadruple prediction task. In experi-

ments with multiple benchmark datasets, we

show that the proposed multi-task prompting

approach brings performance boost (by abso-

lute

8.29

F1) in the few-shot learning setting.

1 Introduction

Aspect-Based Sentiment Analysis (ABSA) is a

ﬁne-grained sentiment analysis task where the goal

is to extract the sentiment associated with an entity

and all its aspects (Liu,2012;Pontiki et al.,2014,

2015,2016;Schouten and Frasincar,2015;Zhang

et al.,2018;Nazir et al.,2020;Zhang et al.,2022).

For example, in the context of Restaurant reviews

the relevant aspects could be food, ambience, lo-

cation, service with general used to represent the

subject itself (i.e., restaurant). ABSA can provide

valuable ﬁne-grained information for businesses

to analyze the aspects they care about. Annotated

∗Indicates equal contribution.

†Work done during internship at AWS.

Figure 1: Illustrative orientation of four ABSA elements

i.e., Aspect Term, Aspect Category, Opinion Term, and

Sentiment. The related tasks often involve predicting

either everything together or a subset of them.

datasets have been released to foster research in

this area (Pontiki et al.,2014,2015,2016).

A full ABSA task aims to extract four elements

from a user-generated text: aspect term, aspect

category, opinion term and the sentiment polar-

ity (see Figure 1for an example). Most existing

approaches have the focus on extracting some of

these elements such as a single element (e.g., as-

pect term), tuple (e.g., aspect term, sentiment po-

larity), or triplet (e.g., aspect term, aspect cate-

gory, sentiment polarity) (Li et al.,2020;Hu et al.,

2019;Xu et al.,2020a). Recently, Zhang et al.

(2021a) tackled the full ABSA task, under the

name of Aspect Sentiment Quadruple Prediction

(ASQP). Technically, most existing computational

approaches have used extractive and discrimina-

tive models either in a pipeline or in an end-to-end

framework (Wang et al.,2016;Yu et al.,2019;

Cai et al.,2021) to address ABSA. Generative ap-

proaches have been recently shown to be effective

for the full ABSA task and its sub-tasks (Zhang

et al.,2021a,b;Yan et al.,2021). Most notably,

Zhang et al. (2021a) used a sequence-to-sequence

(seq-to-seq) model to address ASQP as a para-

phrase generation problem. One important con-

sideration is that modeling ABSA in a generative

fashion allows for cross-task knowledge transfer.

We go a step further and propose a uniﬁed model

arXiv:2210.06629v2 [cs.CL] 11 Jun 2023

that can tackle multiple ABSA sub-tasks, including

the ASQP task, and explore its effectiveness for

low data scenarios. Recent work on large language

models relies on the intuition that most natural

language processing tasks can be described via nat-

ural language instructions and that models trained

on these instructions show strong zero-shot perfor-

mance on several tasks (Wei et al.,2021;Sanh et al.,

2022). Based on this success, we propose a uniﬁed

model based on multi-task prompting with instruc-

tional prompts using T5 (Raffel et al.,2020) to

solve the full ABSA task i.e., ASQP (Zhang et al.,

2021a) and several of its associated sub-tasks ad-

dressed in the literature: 1) Aspect term Extraction

(AE) (Jakob and Gurevych,2010); 2) Aspect term

Extraction and Sentiment Classiﬁcation (AESC)

(Yan et al.,2021); 3) Target Aspect Sentiment

Detection (TASD), which aims to extract the as-

pect term, aspect category, and sentiment polarity

(Wan et al.,2020); 4) Aspect Sentiment Triplet Ex-

traction (ASTE), which aims to extract the aspect

term, opinion term, sentiment polarity (Peng et al.,

2020). We conduct an extensive set of experiments

with multiple review datasets. Experimental results

show that our proposed model achieves substan-

tial improvement (

8.29

F1 on average) against the

state-of-the-art in few-shot learning scenario1.

2 Methods

The four elements of ABSA form a quadruple as

the sentiments are associated with both the aspect,

and the opinion terms (cf Figure 1). In this work,

we hypothesize that it is important to capture the

interaction between these components not only at

the quadruple level, but also within a subset of

these four elements.

We consider multiple factorized sub-tasks in-

volving one or more of the four elements to be pre-

dicted. We pose it as a combination of ﬁve Ques-

tion Answering (QA) tasks as illustrated in Figure

2. For each QA task, an instructional prompt is

used to train a seq-to-seq model to learn one or

more ABSA elements – referred to as Instruction

Tuning (IT). Our formulation enables learning all

sub-tasks via Multi-Task Learning (MTL).

Sources available at:

https://github.com/

amazon-science/instruction-tuning-for-absa

Figure 2: Instruction tuning to solve the sub-tasks re-

lated to ABSA. We devise multiple prompts to instruct a

seq-to-seq model to learn in multi-task learning manner.

2.1 Input Transformation

First, we transform each sentence in the corpus

using the instruction templates provided for each

task as shown in Table 1. Furthermore, we use mul-

tiple paraphrased instruction templates as shown

in Table 2for a task, and sample randomly when

preparing a batch during training (and evaluation)

of the seq-to-seq model. However, the target out-

put sequence remains unchanged irrespective of

the template sampled for a task.

2.2 Model Training

Next, we perform IT with the seq-to-seq model.

We train it in a MTL fashion where input-output

combinations are sampled from all tasks simultane-

ously. We use the following loss for model training:

L=−1

t=1

i=1

log pθ(yi|y1, ..., yi−1,xt).(1)

where

is the transformed input sequence (

) for

tth

task.

is the set of model parameters.

is the

length of output sequence.

is the

ith

token in out-

put sequence.

is the number of tasks. The model

parameters are updated using

Adam

optimizer with

weight decay (Loshchilov and Hutter,2019).

2.3 Output Transformation

Finally, we transform the output using the tem-

plates provided in the rightmost column in Table 1.

In case there is more than one quadruple in the out-

put, we use a special separation token

[SSEP]

. We

map sentiment classes positive, negative and neu-

tral to great, bad and ok respectively in the output

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：

InstructionTuningforFew-ShotAspect-BasedSentimentAnalysisSiddharthVaria1∗ShuaiWang1∗KishaloyHalder1∗RobertVacareanu2†∗MiguelBallesteros1YassineBenajiba1NehaAnnaJohn1RishitaAnubhai1SmarandaMuresan1DanRoth11AWSAILabs2UniversityofArizona,Tucson,AZ,USA{siddhvar,wshui,kishaloh,ballemig,benajiy,ranubhai,s...

展开>> 收起<<

Instruction Tuning for Few-Shot Aspect-Based Sentiment Analysis Siddharth Varia1Shuai Wang1Kishaloy Halder1Robert Vacareanu2 Miguel Ballesteros1Yassine Benajiba1Neha Anna John1.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！

Instruction Tuning for Few-Shot Aspect-Based Sentiment Analysis Siddharth Varia1Shuai Wang1Kishaloy Halder1Robert Vacareanu2 Miguel Ballesteros1Yassine Benajiba1Neha Anna John1

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: