On the Evaluation of the Plausibility and Faithfulness of Sentiment Analysis Explanations Julia El Zini Mohamad Mansour Basel Mousi and Mariette Awad

2025-05-02 0 0 377.73KB 13 页 10玖币

侵权投诉

On the Evaluation of the Plausibility and Faithfulness of

Sentiment Analysis Explanations

Julia El Zini, Mohamad Mansour, Basel Mousi, and Mariette Awad

Electrical and Computer Engineering Department

American University of Beirut, Beirut, Lebanon

{jwe04,mgm35,bam20}@mail.aub.edu, mariette.awad@aub.edu.lb

Abstract.

With the pervasive use of Sentiment Analysis (SA) models in ﬁnancial

and social settings, performance is no longer the sole concern for reliable and

accountable deployment. SA models are expected to explain their behavior and

highlight textual evidence of their predictions. Recently, Explainable AI (ExAI) is

enabling the “third AI wave” by providing explanations for the highly non-linear

black-box deep AI models. Nonetheless, current ExAI methods, especially in the

NLP ﬁeld, are conducted on various datasets by employing different metrics to

evaluate several aspects. The lack of a common evaluation framework is hindering

the progress tracking of such methods and their wider adoption.

In this work, inspired by ofﬂine information retrieval, we propose different metrics

and techniques to evaluate the explainability of SA models from two angles. First,

we evaluate the strength of the extracted “rationales” in faithfully explaining the

predicted outcome. Second, we measure the agreement between ExAI methods

and human judgment on a homegrown dataset

to reﬂect on the rationales plausi-

bility. Our conducted experiments comprise four dimensions: (1) the underlying

architectures of SA models, (2) the approach followed by the ExAI method, (3)

the reasoning difﬁculty, and (4) the homogeneity of the ground-truth rationales.

We empirically demonstrate that anchors explanations are more aligned with the

human judgment and can be more conﬁdent in extracting supporting rationales.

As can be foreseen, the reasoning complexity of sentiment is shown to thwart

ExAI methods from extracting supporting evidence. Moreover, a remarkable

discrepancy is discerned between the results of different explainability methods

on the various architectures suggesting the need for consolidation to observe

enhanced performance. Predominantly, transformers are shown to exhibit better

explainability than convolutional and recurrent architectures. Our work paves the

way towards designing more interpretable NLP models and enabling a common

evaluation ground for their relative strengths and robustness.

The ﬁnal publication is available at Springer via

https://doi.org/10.1007/

978-3-031-08337-2_28

. Citation: El Zini, J., Mansour, M., Mousi, B., Awad, M. (2022).

On the Evaluation of the Plausibility and Faithfulness of Sentiment Analysis Explanations. In

Artiﬁcial Intelligence Applications and Innovations. AIAI 2022. IFIP Advances in Information

and Communication Technology, vol 647. Springer, Cham.

1Dataset and code available at https://gitlab.com/awadailab/exai-nlp-eval

arXiv:2210.06916v1 [cs.CL] 13 Oct 2022

1 Introduction

Sentiment Analysis (SA) is instrumental to the ﬁnancial services industry [

] as

it develops techniques to interpret customer feedback, monitor product reputations,

understand the customers’ needs, and conduct market research. Harnessing the power of

Deep Learning (DL) in understanding general contexts, the performance of SA models

is considerably boosted [

]. However, the non-linearity and the black-box nature

of such models hinder the interpretation of the predictions [

]. Besides providing

guarantees on reliability, generalization, robustness, and fairness, the interpretability of

the SA models can be of service to behavioral marketing and personalized advertisement.

Recently,

plainable

rtiﬁcial

ntelligence (ExAI) algorithms are breathing a

new ﬂexibility in general AI applications by developing methods to explain model’s

prediction [

]. Numerical data frameworks and computer vision applications have

witnessed an explosive growth of ExAI nurtured by the ease of expression of features as

interpretable components [

]. However, only a few ExAI methods are applied

to textual classiﬁers, embeddings, and language models [

]. In the SA framework,

researchers integrated data augmentation techniques to improve the interpretability of

SA models [

], studied attention mechanisms in SA through an explainability lens [

]

and applied ExAI on aspect-based SA models [

]. To date, ExAI methods on Natural

Language Processing (NLP) tasks are not evaluated on standardized benchmarking

datasets through common metrics which hinders the progress and adoption of such

methods in the NLP ﬁeld. Evaluating explainability methods is two-fold. First, it helps

assess the extent to which a deep model can be made explainable. Second, it provides

a common ground to measure the contrast between explanations produced by diverse

ExAI approaches.

In this work, we inspect two human aspects of explainability methods: (1) faithfulness

to the model being explained and (2) plausibility from a human lens. For this purpose,

we select eight state-of-the-art SA models with underlying architectures of recurrent,

convolutional, and attention layers. We generate explanations of the predictions of

these models on three ExAI methods that can be applied in NLP; mainly LIME [

anchors [

] and SHAP [

]. The generated explanations are then evaluated through

two procedures. First, faithfulness

is evaluated by examining the degradation in the

model’s performance when only extracted rationales are fed to the model. Second, the

plausibility

of extracted rationales is evaluated via comparison to the human judgment of

what a sufﬁcient explanation is. This experiment entails a homegrown dataset of manually

labeled explanations on SA data aggregated through conjunction and disjunction means.

The comparison is achieved on six proposed metrics, inspired by information retrieval,

to evaluate the precision and fallout of exAI methods on the SA models. Hence, our

evaluation is carried out over four different dimensions: (1) SA model, (2) ExAI method,

(3) reasoning complexity, and (4) human judgment homogeneity.

The contributions of this work are: (1) a dataset for SA explainability labeled on

different dimensions (2) the ﬁrst faithfulness and plausibility evaluation inspired from

information retrieval (3) a thorough four-dimensional ExAI evaluation on SA models.

2refers to the metric hereafter

Our empirical analysis allows us to draw conclusions on the faithfulness of LIME

rationales and the plausibility of the anchors model which is found to be more conﬁdent

in extracting supporting evidence. Moreover, we highlight the consistency of different

attention architectures in deriving relatively more plausible explanations.

Next, we provide a general background on the sentiment analysis and ExAI models

used in this work in Section 2. Then, we present our evaluation dataset and framework

in Sections 3 and 4. We report our comparative analysis in Section 5 before concluding

with ﬁnal remarks in Section 6.

2 Background

Little has been done on the evaluation of ExAI in NLP settings. Recently, a framework

to evaluate rationalized explanations is introduced in ERASER [

]. ERASER provides

benchmarking data for 7 NLP tasks and suggests sufﬁciency and comprehensiveness as

evaluation metrics. While ERASER considers a wider range of NLP tasks; it is narrow

in terms of the deep architectures and angles that it considers. In contrast, we study the

explainability of SA from four different perspectives. Prior to ERASER, the work of [

]

evaluates the explainability of SVMs compared to CNNs to ﬁnd that the latter models

yield more interpretable decisions. Other attempts only consider the attention mechanism

and the debate concerning its inherent interpretability [

]. Additionally, the

concept of explanation faithfulness has been introduced before [

] with no explicit

evaluation of textual classiﬁers.

Next, we provide the background on sentiment analysis and explainability methods.

2.1 Sentiment Analysis Models

SA models currently exploit deep architectures, word embeddings, transfer learning,

and attention mechanisms [

]. In this work, we experiment with eight state-of-the-art

SA models with different architectures that vary between convolution, recurrent, and

attention networks. The models are chosen with a diversity of word embeddings and

some of the models leverage transfer learning techniques during training.

First, CNN-MC, a CNN for sentiment classiﬁcation [

], consists of a simple CNN

trained on top of static word2vec word embeddings. To experiment with different em-

bedding models, the Universal Sentence Encoder (USE)[

], is used to train additional

convolution layers. USE obtains sentence embeddings through a deep averaging network.

The recurrent architecture is tested on the Byte-multiplicative LSTMs (bmLSTM), used

in [27] to generate reviews and discover sentiments.

The attention architecture [

] is studied in ﬁve transformer models. We ﬁrst study

the parent transformer, BERT [

] which pre-trains deep bidirectional representations by

jointly conditioning on the left as well as the right context extracted by each layer on

unlabeled data. Then, we study an optimized version of BERT, RoBERTa [

], which is

trained longer and on prolonged sequences. We further study a different optimization

of BERT, ALBERT [

], that factorizes the embedding matrix into smaller matrices,

separating thus the size of vocabulary from the size of the hidden layers. Moreover, we

consider DistilBERT [

], a distilled version of BERT trained on very large batches

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：

OntheEvaluationofthePlausibilityandFaithfulnessofSentimentAnalysisExplanationsJuliaElZini,MohamadMansour,BaselMousi,andMarietteAwadElectricalandComputerEngineeringDepartmentAmericanUniversityofBeirut,Beirut,Lebanonfjwe04,mgm35,bam20g@mail.aub.edu,mariette.awad@aub.edu.lbAbstract.Withthepervasiveuseo...

展开>> 收起<<

On the Evaluation of the Plausibility and Faithfulness of Sentiment Analysis Explanations Julia El Zini Mohamad Mansour Basel Mousi and Mariette Awad.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！

On the Evaluation of the Plausibility and Faithfulness of Sentiment Analysis Explanations Julia El Zini Mohamad Mansour Basel Mousi and Mariette Awad

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: