PoKE Prior Knowledge Enhanced Emotional Support Conversation with Latent Variable

2025-05-02 0 0 2.32MB 13 页 10玖币

侵权投诉

PoKE: Prior Knowledge Enhanced Emotional Support

Conversation with Latent Variable

Xiaohan Xu

Institute of Information Engineering

Chinese Academy of Sciences

Beijing, China

xuxiaohan@iie.ac.cn

Xuying Meng

Institute of Computing Technology

Chinese Academy of Sciences

Beijing, China

mengxuying@ict.ac.cn

Yequan Wang∗

Beijing Academy of Articial

Intelligence

Beijing, China

tshwangyequan@gmail.com

ABSTRACT

Emotional support conversation (ESC) task can utilize various sup-

port strategies to help people relieve emotional distress and over-

come the problem they face, which has attracted much attention

in these years. The emotional support is a critical communication

skill that should be trained into dialogue systems. Most existing

studies predict the support strategy according to current context to

guide response. However, most state-of-the-art works rely heavily

on external commonsense knowledge to infer the mental state of

the user in every dialogue round.

Although eective, they may suer from signicant human ef-

fort, knowledge update and domain change in a long run.

Therefore, in this article, we focus on exploring the task itself

without using any external knowledge. We nd all existing works

ignore two signicant characteristics of ESC. (a) Abundant prior

knowledge exists in historical conversations, such as the responses

to similar cases and the general order of support strategies, which

has a great reference value for current conversation. (b) There is

aone-to-many mapping relationship between context and support

strategy, i.e.multiple strategies are reasonable for a single context.

It lays a better foundation for the diversity of generations. Taking

into account these two key factors, we propose

nowledge

nhanced emotional support model with latent variable,

PoKE

versations as exemplars to guide generation, applies the rst-order

Markov model to help predict the target strategy, and then utilizes

a latent variable to model the one-to-many relationship of sup-

port strategy. The proposed model fully taps the potential of prior

knowledge in terms of exemplars and strategy sequence instead

of external knowledge, and then utilizes a latent variable to model

the one-to-many relationship of strategy. Furthermore, we intro-

duce a memory schema to incorporate the encoded knowledge into

decoder. Experiment results on benchmark dataset show that our

PoKE outperforms existing baselines on both automatic evaluation

and human evaluation. Compared with the model using external

knowledge, PoKE still can make a slight improvement in some met-

rics. Further experiments prove that abundant prior knowledge is

∗Corresponding author.

Permission to make digital or hard copies of all or part of this work for personal or

classroom use is granted without fee provided that copies are not made or distributed

for prot or commercial advantage and that copies bear this notice and the full citation

on the rst page. Copyrights for components of this work owned by others than ACM

must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,

to post on servers or to redistribute to lists, requires prior specic permission and/or a

fee. Request permissions from permissions@acm.org.

Conference’17, July 2017, Washington, DC, USA

ACM ISBN 978-x-xxxx-xxxx-x/YY/MM. . . $15.00

https://doi.org/10.1145/nnnnnnn.nnnnnnn

Hello, I feel very stressed

[Question]Hello, what’s going on?

I was looking for some assistance.

I lost my job 4 month ago

[Self-disclosure] Oh, I‘ve lost my job

once, but I’ve made it through. How

have you been handling it?

First few month was kind off alright,but

now I feel very depressed and useless.

[Providing suggestions]Maybe you

can consider talking to your co-

workers in asimilar situation.

(a) Conversation

(b) Multiple valid strategies

context

[Affirmation and Assurance]

[Providing suggestions]

[Question]

Retrieved exemplary responses

[Providing suggestions] Maybe you

can find work online doing surveys or

freelance work.

[Providing suggestions] I think if you

talk to your co-workers, they'll have

ideas for upgrading your workspace.

[Question] Are you eligible for

unemployment benefits?

…

Transition prob. of [Self-disclosure]

[Providing suggestions]

Figure 1: (a) An example to illustrate the ESC task. (b) The

one-to-many mapping relationship that there exist multiple

valid strategies for a single context. (c) Retrieved exemplary

responses give supporter more clues to focus on seeker’s

problem and express strategy more accurately. Meanwhile,

transition probability of strategy provides a good bias to

take a correct strategy. Orange text denotes the strategy

taken by supporter.

conducive to high-quality emotional support, and a well-learned

latent variable is critical to the diversity of generations.

CCS CONCEPTS

•Computing methodologies →Discourse, dialogue and prag-

matics;•Information systems →Sentiment analysis.

KEYWORDS

dialogue system, emotional support conversation, prior knowledge,

latent variable

ACM Reference Format:

Xiaohan Xu, Xuying Meng, and Yequan Wang. 2023. PoKE: Prior Knowl-

edge Enhanced Emotional Support Conversation with Latent Variable. In

Proceedings of ACM Conference (Conference’17). ACM, New York, NY, USA,

13 pages. https://doi.org/10.1145/nnnnnnn.nnnnnnn

1 INTRODUCTION

Emotional support conversation (ESC) [

] is an emerging and chal-

lenging task that devotes to coping eectively with help-seeker’s

emotional distress and helping them overcome the challenges they

arXiv:2210.12640v2 [cs.CL] 15 Feb 2023

Conference’17, July 2017, Washington, DC, USA Xiaohan Xu, Xuying Meng, and Yequan Wang

face. In general, a well-designed ESC system is crucial for many

applications, e.g. customer service chats, mental health support,

etc. [

]. Compared to the well-researched emotional and empa-

thetic conversation [

], ESC focuses on reducing users’

emotional stress using various emotional support strategies, such as

Question, Providing Suggestions, etc.

Recently, several works have been proposed to explore the ESC

task. BlenderBot-Joint [

] generates a strategy token as a prompt

to guide the desired response. MISC [

] uses an o-the-shelf gen-

erative commonsense model, called COMET [

], to infer the user’s

mental status, where the COMET can be seen as an external com-

monsense knowledge base. Then, MISC encodes them additionally

and fuses multiple strategies into one response to generate skill-

fully. GLHG [

] also utilizes COMET to generate the local intention

of seeker in each dialogue round, but considers the hierarchical

relationship between the seeker’s global situation (summarizing

the condition of the seeker) and the local intention. Although eec-

tive, the commonsense knowledge in COMET need to be carefully

integrated into these models to realize their best potential, and the

external knowledge base requires a great deal of eort to develop.

Further, their model may not be applicable when knowledge base is

updated or application domain is changed. Therefore, in this article,

we emphasize on exploring the existing knowledge in the dataset

and the characteristics of ESC task under the setting of no external

knowledge.

Due to the characteristics of ESC, all existing works still suer

two key issues. First, all of them are limited to the scope of the

current conversation, but ignore the abundant prior knowledge in

global historical conversations. Moreover, they fail to model the

one-to-many mapping relationship of strategy, i.e. not only one

but multiple strategies could be valid for a single context. These

issues lead to the challenge of generating high-quality and diverse

responses. We next explain these two issues separately.

Generally, when we attempt to solve help-seeker’s problems, we

are adept in drawing on related prior knowledge as reference, e.g.

psychologists would consult many prior classical cases relevant

to current case [

]. In ESC, instead of external knowledge, there

also exists much prior knowledge to rely on, such as the (1) exem-

plary responses to similar cases and (2) the general order of support

strategies. This prior knowledge has a great reference value to help

explore seeker’s problem and decide the target support strategy.

An explanatory example in Figure 1 illustrates how prior knowl-

edge guides and benets emotional support conversation. (1) The

retrieved context-related responses from historical conversations,

called exemplars, can serve as prior knowledge of response. On

the one hand, some exemplars, e.g. “I think if you talk to ...”, guide

supporter to give more emphasis on the key problem “losing job”,

and thus benet supporter to focus on and explore seeker’s problem.

On the other hand, some exemplars, e.g. “Maybe you can nd ...”,

provide a hint to accurately express the target strategy Providing

suggestions in the sentence pattern starting with “Maybe you”. (2) In

addition to prior knowledge of response, the transition probability

of strategy calculated in training set can act as prior knowledge

to help decide the current strategy. This is because the support

strategies in ESC follow the procedure of three stages (Exploration,

Comforting and Action) [

]. Figure 1(c) shows a transition prob-

ability of strategy Self-disclosure. It illustrates that after sharing

the similar diculties they faced, supporters tend to use Providing

suggestions to give advice based on their experience.

Additionally, it is well known that dialogue systems have a one-

to-many problem of generation, i.e. given a single context there

exists multiple valid responses [

]. In ESC, the supporter is re-

quired to take reasonable strategies, so there is also a one-to-many

problem of support strategy. As shown in Figure 1 (b), after the

seeker states his problem, the supporter can also employ other

valid strategies except for the frequently used strategy Providing

suggestions. Taking the strategy Question to take a deeper look at

user’s problem or Armation and Reassurance to comfort the user

is also a decent choice. Moreover, adopting various strategies is

benecial to diverse responses. In a nutshell, incorporating prior

knowledge and modeling the one-to-many mapping relationship

of strategy are critical to provide emotional support in ESC task.

To take into account these two signicant characteristics of ESC,

we propose a novel model called

nowledge

nhanced emo-

tional support conversation with latent variable model (

PoKE

The proposed model could not only fully tap the potential of prior

knowledge in terms of exemplars and strategy sequence, but also

model the one-to-many mapping relationship of strategy. First, we

construct prior knowledge of exemplars and strategy sequence be-

fore training. Then we use a ne-tuned dense passage retrieval

(DPR) [

] to retrieve a set of responses semantically related to

the input context, and build a rst-order Markov transition matrix

of strategy sequence from training set. To model the one-to-many

mapping relationship of strategy, we introduce conditional varia-

tional autoencoder (CVAE) [

] to predict diverse probability distri-

bution of strategy conditioned on current conversation and prior

knowledge of strategy sequence. Furthermore, we assign exemplars

with dierent attentions according to the distribution of strategy

to emphasize those more relevant exemplars. Lastly, we apply the

technique of memory schema to eectively incorporate encoded

prior knowledge and latent variable into decoder for generation.

The key contributions are summarized as follows:

(1) We

explore the emotional support conversation task under the setting

of no external knowledge base and propose a novel model,

PoKE

PoKE can promote emotional support conversation by eectively

modeling the prior knowledge in terms of exemplars and strategy

sequence, and the one-to-many mapping relationship of strategy.

(2) We utilize strategy distribution to denoise the exemplars and

apply a memory schema to eectively incorporate encoded infor-

mation into decoder. (3) Experiments on benchmark dataset (i.e.,

ESConv) of ESC task demonstrate that our method is superior to

existing baselines on both automatic evaluation and human evalua-

tion. Compared with the model using external knowledge, PoKE

still can make a slight improvement in some metrics. (4) Impor-

tantly, we reveal that abundant prior knowledge is conducive to

high-quality emotional support, and a well-learned latent variable

is critical to the diversity of generations.

2 RELATED WORK

In this section, we rst detail some existing proposed methods

for the emotional support conversation. Then, because we utilize

retrieved exemplars to guide generation and take a latent variable to

PoKE: Prior Knowledge Enhanced Emotional Support Conversation with Latent Variable Conference’17, July 2017, Washington, DC, USA

solve the one-to-many issue of strategy, we will elaborate retrieve-

based generation and one-to-many issue in dialogue system.

2.1 Emotional Support Conversation

Before the task ESC is proposed, there are two relevant well re-

searched dialogue systems, i.e. emotional chatting [

] and

empathetic responding [

]. Emotional chatting needs

to respond in appropriate emotion or the given emotion, such as

happy or angry [

]. Empathetic responding needs to understand

and feel what user is experiencing, and respond with empathy [

Compared with them, the emerging task of ESC aims at reducing

help-seeker’s emotional stress and help them explore and overcome

the problem the face. The rst work on ESC task, called BlenderBot-

Joint, adopts a chitchat bot BlenderBot [

] as backbone and takes

emotional support into account in conversation [

]. Specically,

they encode the context history and predict a strategy token. Then,

they concatenate the predicted strategy token to the head of gener-

ation to guide the desired response. Meanwhile, they construct an

Emotional Support Conversation dataset (ESConv) annotated with

support strategies for the ESC task. Based on ESConv, MISC [

]

uses an o-the-shelf commonsense model COMET [

] to infer an

instant mental state of seeker and encodes them additionally. When

predicting strategy, they take the probability of each predicted strat-

egy as weight to get a weighted average representation of strategy,

and utilize it for guiding a skillful generation. GLHG [

] considers

the hierarchical relationship between the seeker’s global situation

(summarizing the condition of the seeker) and the local intention

(inferred by COMET in each dialogue round) in conversation, and

uses a graph neural network to encode their relationship for guid-

ing generation. Note that both MISC and GLHG are constrained by

the external knowledge in COMET, which may not be applicable to

some specic domain. The external knowledge base like COMET

also requires signicant human eort to develop. Meanwhile, all of

them are limited to the scope of current conversation but ignore

abundant prior knowledge existing in the dataset. In contrast, we

focus on exploring the existing knowledge and the characteristics

of the ESC task without using any external knowledge.

2.2 Retrieve-based Generation

There are lots of works for retrieve-based generation. We will de-

tail some classical studies since our main aim is not to compare

with them. Some generative models, like GPT2 [

], perform well

on many tasks such as machine translation and question answer

[

]. However, recent some works have pointed out that in

dialogue system, the generation model just relied on the input con-

text suers from some issues, such as dull generation (e.g. “I don’t

know") and hallucination [

]. To prompt model to generate

more engaging response, RetNRef [

] proposes a simple but ef-

fective retrieve-and-rene strategy. RetNRef appends the retrieved

context-relevant responses to context to guide the generation. Simi-

lar to this approach, Cai et al. [

] retrieves both literally-similar and

topic-related exemplars to guide dialogue generation. Majumder et

al. [

] employs dense passage retrieval and introduce three com-

munication mechanisms of empathy to facilitate the generation

towards empathy. For the ESC task, the abundant prior knowledge

in historical conversations has great reference value for reducing

seeker’s emotional stress. Besides, the responses with the same

strategy are similar in sentence pattern. Thus, we introduce exem-

plars into generation model and denoise exemplars according to the

strategy distribution to emphasize those more relevant exemplars.

2.3 One-to-Many Problem

It is well known that dialogue systems have a one-to-many map-

ping problem that given a single context, there exist multiple valid

responses [

]. To model this one-to-many feature and improve

the diversity of generations, many works introduce latent variable

to model a probability distribution over the potential responses

[

]. DialogVED [

] combines continuous latent variable

into the encoder-decoder pre-training framework to generate more

relevant and diverse responses. Except for continuous represen-

tation of latent variables, some works utilize discrete categorical

variables to promote the interpretability of generation [

]. For

ESC, there also exist several reasonable support strategies and the

corresponding responses at a certain stage. Therefore, it is required

to additionally consider the one-to-many mapping relationship of

strategies. In our work, we introduce a continuous latent variable to

model the distribution over strategy. Furthermore, we employ this

strategy distribution to denoise the exemplars at the sequence-level

to focus on strategy-relevant exemplars.

3 POKE

Problem Denition.

The dialogue context in ESC is an alternat-

ing set of utterances from seeker and supporter. Given a sequence

𝑁

context utterances

𝑐=(𝑢1, 𝑢2,· · · , 𝑢𝑁)

, where each utterance

consists of some words,

𝑢𝑖=(𝑤𝑖

1, 𝑤𝑖

2,· · · , 𝑤𝑖

𝑀)

. In the setting of

ESC, each utterance of supporter is labeled with a support strat-

egy. There are total 8 support strategies, i.e. Question,Reection

of feelings,Information,Restatement or Paraphrasing,Others,Self-

disclosure,Armation and Reassurance, and Providing Suggestions

(for more detail please refer to original paper [

]). We use

𝑚

denote the total number of strategies in the following parts. Except

for the strategy, there is a brief situation

𝑠

ahead of conversation

summarizing the condition of seeker. In this paper, we denote the

previous one support strategy taken by supporter as

𝑦′

, and the

last utterance of seeker (called post) as

𝑝

. Then, our model aims at

using multiple input information and prior knowledge to generate

an emotional support response 𝑟by reasonable support strategies.

PoKE Overview.

Our devised model uses BlenderBot-small [

]

as the backbone. The overview of our method is shown in Figure 2,

which consists of four main parts: (a)

prior knowledge module

to retrieve context-related exemplary responses and build a Markov

transition matrix of strategy sequence from training set, (b)

unied

encoder

to encode multiple input source and exemplars by adding

source tokens, (c)

latent variable module

to model the proba-

bility distribution of strategy and denoise the exemplars and (d)

knowledge-memory decoder

to eectively incorporate encoded

prior knowledge and latent variable into decoder for generation.

3.1 Prior Knowledge Module

Humans tend to use prior knowledge to bias decisions [

], and

there is abundant prior knowledge in historical conversation for

ESC task. Due to the characteristics of ESC, we consider the prior

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：

PoKE:PriorKnowledgeEnhancedEmotionalSupportConversationwithLatentVariableXiaohanXuInstituteofInformationEngineeringChineseAcademyofSciencesBeijing,Chinaxuxiaohan@iie.ac.cnXuyingMengInstituteofComputingTechnologyChineseAcademyofSciencesBeijing,Chinamengxuying@ict.ac.cnYequanWang∗BeijingAcademyofArtif...

展开>> 收起<<

PoKE Prior Knowledge Enhanced Emotional Support Conversation with Latent Variable.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！

PoKE Prior Knowledge Enhanced Emotional Support Conversation with Latent Variable

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: