SUMBot Summarizing Context in Open-Domain Dialogue Systems Rui Ribeiro Lu ısa Coheur INESC-ID Lisboa Portugal

2025-04-24 0 0 2.03MB 5 页 10玖币

侵权投诉

SUMBot: Summarizing Context in Open-Domain Dialogue Systems

Rui Ribeiro, Lu´

ısa Coheur

INESC-ID, Lisboa, Portugal

Instituto Superior T´

ecnico, Universidade de Lisboa, Portugal

rui.m.ribeiro@tecnico.ulisboa.pt, luisa.coheur@tecnico.ulisboa.pt

Abstract

In this paper, we investigate the problem of including relevant

information as context in open-domain dialogue systems. Most

models struggle to identify and incorporate important knowl-

edge from dialogues and simply use the entire turns as context,

which increases the size of the input fed to the model with un-

necessary information. Additionally, due to the input size limi-

tation of a few hundred tokens of large pre-trained models, re-

gions of the history are not included and informative parts from

the dialogue may be omitted. In order to surpass this problem,

we introduce a simple method that substitutes part of the context

with a summary instead of the whole history, which increases

the ability of models to keep track of all the previous relevant

information. We show that the inclusion of a summary may im-

prove the answer generation task and discuss some examples to

further understand the system’s weaknesses.

Index Terms: dialogue systems, summarization, dealing with

context, open-domain

1. Introduction

Chit-chat systems have become more and more prominent with

the emergence of large pre-trained models and the increased ac-

cess to public libraries [1, 2, 3] that allow to easily train and

deploy these models. Speciﬁcally, new advances have shown

promising progress in the dialogue generation task, as these

systems became more competent at providing human-like an-

swers. However, these deep-learning systems tend to generate

generic responses which are repetitive or incoherent with the

context, particularly when conversations attain many interac-

tions and contain long turns.

Recent approaches have studied the ability of deep gener-

ative models to capture relevant information from the dialogue

context [4, 5]. They have found that these models do not efﬁ-

ciently make use of all parts from the dialogue history and tend

to ignore relevant turn information. Other approaches [6, 7, 8, 9]

have attempted to represent the context and leverage the result-

ing representations to various dialogue tasks. However, none of

these approaches has studied the substitution of the context with

a summary.

In this paper, we investigate the importance of encapsulat-

ing complete dialogue utterances into a summary and reducing

the context size in the open-domain dialogue task. We attempt

to answer the following question: can a summary of the pre-

vious context include all the important information and also

decrease the input size fed to a model? To answer this ques-

tion, we propose a simple yet effective method that incorpo-

rates summaries of the previous turns that are not included as

input. More speciﬁcally, apart from the user request, we only

include a few complete speaker turns, and the remaining turns

are compiled into a summary that describes succinctly the omit-

ted utterances. We train different versions of the model where

Figure 1: Example of a dialogue between two speakers and the

respective summary on the SAMSum dataset.

we change the number of complete utterances provided, which

may vary between 0 and 10. This procedure allows us to ana-

lyze if the inclusion of summaries is an effective strategy and if

the summaries become a valuable choice as substitutes for the

complete turns.

The training is divided into two independent stages: ﬁrst,

we ﬁne-tune BART [10] in the SAMSum corpus [11] and use

it to generate summaries for the dialogue context. Figure 1

shows an example of a dialogue from this dataset. Then, we

ﬁne-tune DialoGPT decoder [12] with the summaries from the

previous stage by incorporating them with the dialogue between

both speakers.

We evaluate our model on the open-domain Persona-Chat

dataset [13] and observe that the inclusion of the summaries

may improve the overall results. We also analyze if the sum-

maries are proper substitutes for the dialogue history and dis-

cuss possible ﬂaws that can decrease the performance of the

generation model.

2. Related Work

Since the introduction of encoder-decoder models [14, 15], chit-

chat dialogue systems have been in constant evolution and are

more capable of generating ﬂuent and human-like sentences. In

these systems, the encoder extracts important features from the

utterances and passes that information to a decoder that gener-

ates a response.

Considering that our approach attempts to provide a proper

substitute for the dialogue history, the related work that be-

comes more relevant focuses on studying and representing the

context in the dialogue task. [4] study the aptitude of encoder-

decoder models based on RNNs and Transformers to interpret

and understand the dialogue context. The authors introduce

arXiv:2210.06496v1 [cs.CL] 12 Oct 2022

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：

SUMBot:SummarizingContextinOpen-DomainDialogueSystemsRuiRibeiro,Lu´saCoheurINESC-ID,Lisboa,PortugalInstitutoSuperiorT´ecnico,UniversidadedeLisboa,Portugalrui.m.ribeiro@tecnico.ulisboa.pt,luisa.coheur@tecnico.ulisboa.ptAbstractInthispaper,weinvestigatetheproblemofincludingrelevantinformationascontex...

展开>> 收起<<

SUMBot Summarizing Context in Open-Domain Dialogue Systems Rui Ribeiro Lu ısa Coheur INESC-ID Lisboa Portugal.pdf

共5页,预览1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！

SUMBot Summarizing Context in Open-Domain Dialogue Systems Rui Ribeiro Lu ısa Coheur INESC-ID Lisboa Portugal

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: