Dont Lose Yourself Empathetic Response Generation via Explicit Self-Other Awareness Weixiang Zhao Yanyan Zhao Xin Lu Bing Qin

2025-05-01 0 0 1.23MB 13 页 10玖币

侵权投诉

Don’t Lose Yourself!

Empathetic Response Generation via Explicit Self-Other Awareness

Weixiang Zhao, Yanyan Zhao∗, Xin Lu, Bing Qin

Research Center for Social Computing and Information Retrieval

Harbin Institute of Technology, China

{wxzhao, yyzhao}@ir.hit.edu.cn

Abstract

As a critical step to achieve human-like chat-

bots, empathetic response generation has at-

tained increasing interests. Previous attempts

are incomplete and not sufﬁcient enough to

elicit empathy because they only stay on the

initial stage of empathy to automatically sense

and simulate the feelings and thoughts of oth-

ers via other-awareness. However, they ig-

nore to include self-awareness to consider

the own views of the self in their responses,

which is a crucial process to achieve the em-

pathy. To this end, we propose to gener-

ate Empathetic response with explicit Self-

Other Awareness (EmpSOA). Speciﬁcally,

three stages, self-other differentiation, self-

other modulation and self-other generation,

are devised to clearly maintain, regulate and

inject the self-other aware information into

the process of empathetic response generation.

Both automatic and human evaluations on the

benchmark dataset demonstrate the superior-

ity of EmpSOA to generate more empathetic

responses. Our source code is available at

https://github.com/circle-hit/EmpSOA.

1 Introduction

Empathy is a desirable trait of the engaging human

conversation and is considered as the key step to

human-like chatbots. In this paper, we focus on the

task of empathetic response generation (Rashkin

et al.,2019), which understands the feelings and

situations of the user and responses properly.

According to one of the most inﬂuential theories

of empathy proposed by Rogers (1992), empathy

is the ability to sense the others’ private world as

if it were our own,

but without losing the “as if”

condition.

Previous works (Lin et al.,2019;Li

et al.,2020;Sabour et al.,2022) mainly focus on

the prior part of the deﬁnition, which is referred

as emotional contagion (Hatﬁeld et al.,1993), and

∗Corresponding author

Hi , I feel so lonely sometimes because all

my friends live in a different country

I am so sorry to hear that.

Oh, I am sure you are lonely. Maybe

you can join some kind of club that

lets you meet new friends?

(a) Other-Awareness (b) Self-Other Awareness

I feel your sadness. I feel your sadness

but I know that is not my own sadness.

I should cheer you up.

Self-Awareness

Other-Awareness

Figure 1: An example for the self-other awareness dur-

ing an empathetic conversation from the EMPATHETIC-

DIALOGUES (Rashkin et al.,2019) dataset. The user

and the system are referred as the other and the self.

they automatically mimic the thoughts and feelings

of the speaker to converge emotionally.

However, emotional contagion is only the initial

component that precedes empathy and what makes

empathy distinct from emotion contagion is a crit-

ical process called

self-other awareness

(Decety

and Lamm,2006). Thus, it is not sufﬁcient enough

for previous attempts to convey empathy because

they only perform the single

other-awareness

perceive the emotion and situation from the other

and generate responses coupled with the same per-

ceived emotion. As shown in the left part of Fig-

ure 1, with other-awareness, the self succeed to feel

the sadness of the other. But the complete emo-

tional overlap would induce the self to reinforce

such sadness with the response of so sorry to hear

that, which is not the goal of empathy.

And it is of great necessity to take the own view

of the self into consideration to maintain the “as

if” condition with a clear

self-awareness

, which

is a conscious process to maintain and modulate

the own views of the self during the empathetic

interaction. In the right part of Figure 1, with the

incorporation of the self-awareness, the self con-

sciously holds in mind that this it not my own sad-

arXiv:2210.03884v2 [cs.CL] 5 May 2023

ness and I am responsible to cheer you up. Thus,

the explicit self-other awareness plays pivotal roles

in disentangling feelings and views of the self and

the other, which constitutes a crucial perspective of

empathy and largely contributes to generate more

empathetic responses, especially when the other is

in negative emotional states.

To this end, we propose to generate

Emp

athetic

response with explicit

elf-

ther

wareness

(

EmpSOA

). Inspired by the conceptual frame-

work of information ﬂow involved in human em-

pathy (Decety and Lamm,2006), we make such

processes computable and abstract three stages

in EmpSOA, named Self-Other Differentiation

(SOD), Self-Other Modulation (SOM) and Self-

Other Generation (SOG). Speciﬁcally, in SOD, we

construct two heterogeneous graphs with four types

of nodes to maintain the self-awareness represen-

tation and other-awareness representation, respec-

tively. Among them, commonsense knowledge

from COMET (Bosselut et al.,2019) is leveraged

to manifest the ﬁne-grained emotional and cog-

nitive statuses of the self and the other. Further,

we dynamically control the contributions of the

self-other awareness representations in SOM and

inject them into the process of empathetic response

generation in SOG. Experimental results of both au-

tomatic and manual evaluations on the benchmark

dataset demonstrate the superiority of EmpSOA to

generate more empathetic responses.

The main contributions of this work are summa-

rized as follows:

•

We propose to generate empathetic responses

via explicit self-other awareness, which con-

stitutes a critical perspective of empathy.

•

We devise a novel model EmpSOA to clearly

maintain, modulate and inject the self-other

aware information into the process of empa-

thetic response generation.

•

Results of extensive experiments on the bench-

mark dataset demonstrate the effectiveness of

EmpSOA to identify the exact emotion of the

other and generate more empathetic response.

2 Related Work

2.1 Empathetic Response Generation

Endowing empathy to the dialogue systems has

gained more and more attentions recently. For pre-

vious attempts on empathetic response generation,

we divide them into two categories according to

whether they incorporate both affection and cog-

nition aspects of empathy. On the one hand, most

existing works (Alam et al.,2018;Rashkin et al.,

2019;Lin et al.,2019;Majumder et al.,2020;Li

et al.,2020,2022;Wang et al.,2022) only con-

sider the affective aspect of empathy to understand

the emotional state of the other and converge emo-

tionally. On the other hand, Sabour et al. (2022)

propose to comprehensively understand the emo-

tional feelings and cognitive situations of the other

with commonsense knowledge incorporated.

However, all previous methods only perceive

the emotional or cognitive states of the other by

the single other-awareness, ignoring to explicitly

incorporate self-awareness to make an appropriate

empathetic response with own views of the self.

2.2 Emotional Dialogue Generation

Emotion has been proven to be the key factor of

achieving more engaging dialogue systems. Previ-

ous works explore two ways of incorporating emo-

tion into dialogue generation. On the one hand, the

generation-based methods (Zhou et al.,2018;Zhou

and Wang,2018;Shen and Feng,2020) are pro-

posed to generate emotional responses given a spec-

iﬁed emotion label. On the other hand, retrieval-

based (Qiu et al.,2020;Lu et al.,2021) methods

aim to obtain emotional responses from candidates

retrieved from the response repository. However,

expressing the speciﬁed emotion in responses is

merely the fundamental goal to achieve emotional

dialogue systems, which is lack of the understand-

ing for user’s feelings and situations required by

the empathetic response generation.

3 Methodology

3.1 Task Deﬁnition

First, we deﬁne the task of empathetic response

generation. Formally, let

D= [X1, X2,· · · , XN]

denotes a dialogue history with

utterances be-

tween the user (the other) and the system (the self),

where the

-th utterance

Xi= [wi

1, wi

2· · · , wi

a sequence of

words. Besides, each conversation

is provided with an emotion label

from the total

32 available emotions to signal what the emotional

tone that the other is grounded on. The goal is to

generate the next utterance

from the stand of the

self that is coherent to the dialogue history

and

empathetic to the other’s situation and feeling.

Multi-Head Self and Other Awareness Module

Self Awareness

Other Awareness

Agent Utterance Node

User Utterance Node

Emotional Node

Cognitive Node

Emotional State Node

Cognitive State Node

Transformer Encoder

. . . . . .

. . .

Stacked

. . . . . .

Word

Embedding

Role

Embedding

Position

Embedding

(b) Self-Other Modulation

. . .

Cross Attention

Regulation

(a) Self-OtherDifferentiation

Self-Awareness

Graph

Other-Awareness

Graph

Self Utterance Node Emotional Knowledge Node

Cognitive Knowledge Node

Emotional State Node

Other Utterance Node Cognitive State Node

Transformer Decoder

. . .

Self-Other Aware

Empathetic Generator

Figure 2: The overall architecture of our proposed EmpSOA model, which mainly consists of three modules: (a)

Self-Other Differentiation; (b) Self-Other Modulation and (c) Self-Other Generation.

3.2 Overview of the Architecture

We display the overall architecture of EmpSOA

in Figure 2. We abstract three main stages from

the conceptual framework

of information ﬂow

involved in human empathy (Decety and Lamm,

2006) and make them computabel in EmpSOA, in-

cluding (a) Self-Other Differentiation (SOD), (b)

Self-Other Modulation (SOM), and (c) Self-Other

Generation (SOG). We ﬁrst clearly disentangle the

emotional and cognitive states of the self and the

other to maintain the self- and other-awareness rep-

resentations individually in SOD. Then in SOM,

they are dynamically modulated and controlled to

make different contributions to the self-other aware

contextual information obtained from the context

encoder. Finally, such self- and other-awareness

representations are explicitly injected into the gen-

eration process in SOG to obtain the empathetic

responses from views of both the self and the other.

3.3 Self-Other Aware Context Encoder

We adopt Transformer encoder (Vaswani et al.,

2017) to obtain the contextual representations of

the dialogue history. Following previous works (Li

et al.,2022;Sabour et al.,2022), the dialogue is ﬂat-

tened into a word sequence. To make the encoder

aware of the self-other distinction in the encoding

More details about the conceptual framework of human

empathy are provided in appendix ﬁle.

phase, we append two special tokens,

[SLF]

and

[OTH]

, to the beginning of each utterance from

the self and the other, respectively. Further, role

embedding is added to supplement extra self-other

aware information. The ﬁnal input of the self-other

aware context encoder are the sum of word em-

bedding, role embedding and position embedding:

Hso =Encoder(Ew+Er+Ep)(1)

where

Hso ∈RN×dh

and

is the hidden dimen-

sion of the self-other aware context encoder.

3.4 Self-Other Differentiation

As mentioned above, the clear self-other aware-

ness constitutes a crucial perspective of genuine

empathy. To achieve this, we ﬁrst devise self-other

differentiation (SOD). Speciﬁcally, we construct

two heterogeneous graphs, named self-awareness

graph

GSA

and other-awareness graph

GOA

, to dis-

entangle and maintain self- and other-awareness

representations separately. Inspired by Sabour et al.

(2022), both awareness representations consist of

emotional and cognitive aspects. And we leverage

commonsense knowledge from the external knowl-

edge base ATOMIC (Sap et al.,2019) to imply the

ﬁne-grained emotional and cognitive knowledge of

the self and the other at each dialogue turn. Such

knowledge is highly related to the personal mental

states and it has been widely used in many emo-

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：

Don'tLoseYourself!EmpatheticResponseGenerationviaExplicitSelf-OtherAwarenessWeixiangZhao,YanyanZhao,XinLu,BingQinResearchCenterforSocialComputingandInformationRetrievalHarbinInstituteofTechnology,China{wxzhao,yyzhao}@ir.hit.edu.cnAbstractAsacriticalsteptoachievehuman-likechat-bots,empatheticrespons...

展开>> 收起<<

Dont Lose Yourself Empathetic Response Generation via Explicit Self-Other Awareness Weixiang Zhao Yanyan Zhao Xin Lu Bing Qin.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！

Dont Lose Yourself Empathetic Response Generation via Explicit Self-Other Awareness Weixiang Zhao Yanyan Zhao Xin Lu Bing Qin

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: