CorefDiffs Co-referential and Differential Knowledge Flow in Document Grounded Conversations Lin Xu1 Qixian Zhou2 Jinlan Fu1 Min-Yen Kan1 See-Kiong Ng1

2025-05-06 0 0 852.19KB 13 页 10玖币

侵权投诉

CorefDiffs: Co-referential and Differential Knowledge Flow in Document

Grounded Conversations

Lin Xu1, Qixian Zhou2, Jinlan Fu1, Min-Yen Kan1, See-Kiong Ng1

1National University of Singapore, 2ByteDance

{cathyxl2016,qixianzhou.mail,jinlanjonna}@gmail.com

{knmnyn,seekiong}@nus.edu.sg

Abstract

Knowledge-grounded dialog systems need to

incorporate smooth transitions among knowl-

edge selected for generating responses, to en-

sure that dialog ﬂows naturally. For document-

grounded dialog systems, the inter- and intra-

document knowledge relations can be used

to model such conversational ﬂows. We de-

velop a novel Multi-Document Co-Referential

Graph (Coref-MDG) to effectively capture

the inter-document relationships based on

commonsense and similarity and the intra-

document co-referential structures of knowl-

edge segments within the grounding doc-

uments. We propose CorefDiffs, a Co-

referential and Differential ﬂow management

method, to linearize the static Coref-MDG into

conversational sequence logic. CorefDiffs per-

forms knowledge selection by accounting for

contextual graph structures and the knowledge

difference sequences. CorefDiffs signiﬁcantly

outperforms the state-of-the-art by 9.5%, 7.4%

and 8.2% on three public benchmarks. This

demonstrates that the effective modeling of co-

reference and knowledge difference for dialog

ﬂows are critical for transitions in document-

grounded conversation1.

1 Introduction

Document-grounded conversations (Moghe et al.,

2018;Dinan et al.,2018;Feng et al.,2021b) is

a core class of knowledge-grounded dialogs that

leverage text-based knowledge segments from doc-

uments to generate informative dialog responses.

This task is typically divided into two sub-tasks,

given the dialog history (Dinan et al.,2018):

namely, knowledge selection and response genera-

tion. Knowledge selection, which determines the

content of the generated responses (Moghe et al.,

2018;Dinan et al.,2018), is the crucial sub-task for

dialog ﬂow management as it leads to the manifes-

The source code has been released at https://github.com/

cathyxl/coref-diffs

Science

fiction film

The empire

strikes back

sent-1

Star wars

(film)

sent-1

sent-2

part_of

overlap

genre_of

Science fiction film

uses speculative,

fictional science ... Star wars is a 1977...

space opera film

It depicts the

adventures of

...in a galaxy

Star wars (film)

starring mark hamill,

harrison ford

The empire strikes

back is a 1980...film.

sent-3

co-ref

Coref-MDG

Conversation:

Usr1: I like science fiction

Bot1: yes. i like movies that use speculative, fictional science...depictions

[Science fiction film, sent-1]

Usr2: I liked the star wars movies and the alien ...

Bot2: I wish i was able to see star wars in 1977. [Star wars (film), sent-1]

Usr3: I loved the tv series.

Bot3: I prefer the film that stars mark hamill and harrison ford.[Star

wars (film), sent-3]

Usr4: I like that one too. the first sci fi film was...

Bot4: Cool. Did you ever see the empire strikes back in 1980? ... [The

empire strikes back, 1]

Topic node

Knowledge node

sent-1

Documents

Figure 1: Co-Referential Multi-Document Graph

(Coref-MDG). Topic vertices correspond to documents

and are connected by commonsense/word overlap rela-

tions. Knowledge vertices are connected with its topic

vertex by its document sentence index, e.g. sent-1, and

connected to each other by co-reference (co-ref) rela-

tions. The Bot’ s utterances are followed by its topic

and knowledge segment, e.g. [Science ﬁction ﬁlm,

sent-1].

tation of knowledge transition (Meng et al.,2020),

essential for naturalistic engaging conversations.

Most existing studies on document-grounded

conversations (Lian et al.,2019;Zheng et al.,2020;

Zhao et al.,2020) treat knowledge selection as

a matching problem between the dialog context

and individual knowledge segments, independently.

However, for document-grounded conversations,

we posit that there is an implicit alignment between

the background knowledge and conversation logic

which can be learned from the underlying structural

relationships of the knowledge segments within and

between the grounding documents. For example,

the conversation in Figure 1exhibits document-

level topic ﬂow, from science ﬁction -> star wars->

the empire strikes back, and deep dives into the

speciﬁcs of the star wars document (Turns 2 to 3).

arXiv:2210.02223v1 [cs.CL] 5 Oct 2022

K1 K2 K3 K4 K5 K6 K7

(a) Independent

(b) Locally correlated (c) Document graph

Doc1 Doc3

Doc2

Doc1 Doc2 Doc3

Figure 2: Exploiting knowledge segment relationships.

k1-7 represent knowledge segments and doc1-3 are the

grounding documents they belong to.

To effectively exploit the relationships of the

knowledge segments to guide dialog ﬂows would

require a thorough comprehension of the intra-

document discourse structures and inter-document

relationships for the knowledge selection process.

Existing works either ignore such relations (as illus-

trated in Figure 2(a)), or exploit limited local cor-

relations (as depicted in Figure 2(b)), for example

by encoding knowledge segments within passage

context Wu et al. (2021) . In this work, we propose

to capture both intra- and inter-document relation-

ships of the knowledge segments (Figure 2(c))

in the grounding documents to guide the smooth

and natural knowledge selection and transitions

for document-grounded conversations. However,

how to apply such a static knowledge graph to dia-

log ﬂow management has always been a problem.

Many previous studies (Moon et al.,2019;Xu et al.,

2021a,b) have used graph structures to constrain

search (e.g. conﬁning the next topic to neighboring

areas), but have also ignored deeper integration of

dialog contexts and knowledge graphs, such as op-

timal knowledge representation to capture dialogue

ﬂow information.

Based on the considerations above, we propose

to ﬁrst capture the inter- and intra-document knowl-

edge relationships as a heterogeneous document

graph, and then exploit the graph effectively for di-

alog ﬂow management through ﬁne-to-coarse con-

textualization — from the local word-level knowl-

edge attentions, to knowledge interactions in doc-

ument graphs, and ﬁnally to the knowledge tran-

sition ﬂow along dialogue turns. Speciﬁcally, we

design a two-level document graph consisting of

topic (i.e. document) and knowledge vertices con-

nected by inter- and intra-document relations (Fig-

ure 1). The topic vertices correspond one-to-one

to the grounding documents, while the knowledge

vertices refer to the knowledge segments from each

document. The knowledge vertices are connected

to the corresponding topic vertices they belong

to. Meanwhile, the graph connects the knowl-

edge segments within the same document by their

co-referential mentions, and the documents are

connected based on similarity or commonsense

relationships. Hence we call the graph Multi-

Document Co-referential Graph (Coref-MDG).

We then propose our CorefDiffs method which

leverages Coref-MDG’s graph structure and inte-

grates dialog ﬂow for knowledge contextualization

and selection. CorefDiffs focuses on the inter-turn

knowledge difference ﬂow in the dialog histories by

means of a novel differential linearization module.

Our contributions in this paper can be summa-

rized as follows. 1) We develop Coref-MDG,

a novel multi-document graph structure incorpo-

rating co-referential mentions. When leveraged

in guiding document-grounded conversations in

our CorefDiffs methodology, it empirically outper-

formed alternative graph structures; 2) We system-

atically study the different kinds of inter- and intra-

document relations and show that document-level

semantics, such as co-reference and sentence order,

are signiﬁcant factors for knowledge selection (Sec.

4.4); 3) Our CorefDiffs achieves state-of-the-art

on WoW, Holl-E, multidoc2dial and CMU-DOG

datasets, for both knowledge selection and response

generation tasks.

2 Related Work

Document-grounded dialog Systems.

Early

works on document-grounded dialog sys-

tems (Ghazvininejad et al.,2018) focused on

generating responses directly by copying words

from the external documents. The subsequent

availability of datasets with knowledge an-

notations (Dinan et al.,2018;Moghe et al.,

2018) has led to the separation of the tasks of

knowledge selection and response generation.

For knowledge selection, most works (Dinan

et al.,2018;Lian et al.,2019;Zheng et al.,

2020;Zhao et al.,2020;Meng et al.,2021)

in document-grounded conversations directly

modeled correlations between dialog contexts

and knowledge through independent matching

and optimized the correlations by modeling

knowledge sequence (Kim et al.,2019), increasing

knowledge informativeness (Zheng et al.,2020)

or distinguishing initiative roles (Meng et al.,

2021). A recent work (Wu et al.,2021) boosted

knowledge selection by encoding knowledge

within the passage context, which demonstrates the

importance of exploiting knowledge relations. Our

Dialogue History

𝑢𝑡−1:Blue is always nice. I like royal blue.

𝑟𝑡−1:I once rode on the royal blue train from new york to d.c

𝑢𝑡:Oh that sounds really nice. i bet there was a lot of scenery

and blue skies.

𝑟𝑡: yes, speaking of blue skies, have you seen the 1946 movie

staring bing crosby?

Blue Royal Blue

(train)

Blue Skies

(1946 film)

3x21

Words Overlap, Sub-Class

Transformer Transformer Transformer

… …

…

Words

Overlap

Words

Overlap

diff diffdiff xYES/NO Graph propagation

Diff sequence

Target node

𝒖𝒕𝒓𝒕−𝟏 … 𝒃𝒍𝒖𝒆𝒌𝟏𝒌𝟐𝒌𝟑𝒖𝒕𝒓𝒕−𝟏 … 𝐛𝐥𝐮𝐞 𝐬𝐤𝐢𝐞𝐬𝒌𝟒𝒌𝟓𝒖𝒕𝒓𝒕−𝟏 … 𝐫𝐨𝐲𝐚𝐥 𝐛𝐥𝐮𝐞𝒌𝟔𝒌𝟕𝒌𝟖

III. Linearization

I. Local Association

II. Structural Propagation

Figure 3: CorefDiffs Architecture. Green round rectangles and blue circles are topic and knowledge vertices, re-

spectively. Steps I to III contextualize knowledge in an ﬁne-to-coarse manner: ﬁrst, I) by vertex embedding by

BERT; then, II) by propagating Coref-MDG information, and ﬁnally III) by linearizing the knowledge representa-

tions, according to the dialog’s historical knowledge sequence.

work further explores more effective document

structures and connections for this task. There

is also an unpublished paper that used document

semantic graphs (Li et al.,2022), while our work

considers end-to-end integration of document

graph and dialog ﬂow which gives better result

compared to theirs.

Knowledge Graph for Conversations.

Knowl-

edge graphs were also often used in dialog manage-

ment, such as dialog transition graphs (Xu et al.,

2019,2020a) constructed from common transitions

present in a dialog corpus and off-the-shelf com-

monsense graphs (Zhou et al.,2018a). There were

also some works (Liu et al.,2019;Xu et al.,2021a)

transforming unstructured text into structures or

combining triplets and texts into graphs. For exam-

ple, (Xu et al.,2021a) constructed key phrases into

graphs according to their order in stories. Interest-

ingly, to the best of our knowledge, co-reference

mentions have not been considered in such doc-

ument graph construction although it has been

proved critical in learning language models for rea-

soning intensive NLP tasks (Dasigi et al.,2019;Ye

et al.,2020). To apply knowledge graphs for dialog,

many existing works (Xu et al.,2020b,a) used the

graph structures to conﬁne the search space and

optimized selection through hand-crafted rewards.

In contrast, we incorporate the knowledge graph

into dialog management by learning knowledge

representations from the graph structure.

Sequence Learning in dialog.

Sequence learning

is essential for conversations. Several studies (Kim

et al.,2019;Zhan et al.,2021b) explored the histor-

ical knowledge sequence to select knowledge for

document-grounded dialog. For example, (Kim

et al.,2019) captured knowledge sequence by a

latent variable, while (Zhan et al.,2021b) further

proposed to learn abstract topic sequence to miti-

gate the issues of knowledge sparsity and knowl-

edge transition noise. Inspired by the importance

of exploiting knowledge difference (Zheng et al.,

2020) for informative dialog, we extend the use

of dialog knowledge differences into sequences,

thus capturing the knowledge shift patterns from

turns with longer distances as well as the sequential

patterns of knowledge transitions in a dialogue.

3 Approach

Figure 3shows the overall architecture of our ap-

proach. As shown in the Dialog History part, in

each data sample, given a dialog history

{ut−l, rt−l, ..., rt−1, ut}

turns and a set of

grounding documents

D={d1, .., di, ..., d|D|}

where

u∗

and

r∗

are utterances from the user and

chatbot, respectively.

di={ki

1, ki

2, ..., ki

|di|}

is a

document containing a bunch of knowledge seg-

ments, our task is to select the most appropriate

knowledge segment from the grounding documents

(i.e. the knowledge selection subtask) and gener-

ate the chatbot’s next response

based on the se-

lected knowledge (i.e. the response generation sub-

task). Each grounding document

has a phrase

as its topic. For example, the document of wikipage

blue has the topic phrase blue.

3.1 Coref-MDG Construction

We devise a Multi-Document Co-referential Graph

(Coref-MDG) to capture the inter-document and

the intra-document relations. Each data sample gets

a speciﬁc Coref-MDG, denoted as

G={V,E}

where V,Eare vertices and edges respectively.

Vertices V.

Our Coref-MDG consists of two types

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：

CorefDiffs:Co-referentialandDifferentialKnowledgeFlowinDocumentGroundedConversationsLinXu1,QixianZhou2,JinlanFu1,Min-YenKan1,See-KiongNg11NationalUniversityofSingapore,2ByteDance{cathyxl2016,qixianzhou.mail,jinlanjonna}@gmail.com{knmnyn,seekiong}@nus.edu.sgAbstractKnowledge-groundeddialogsystemsneed...

展开>> 收起<<

CorefDiffs Co-referential and Differential Knowledge Flow in Document Grounded Conversations Lin Xu1 Qixian Zhou2 Jinlan Fu1 Min-Yen Kan1 See-Kiong Ng1.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！

CorefDiffs Co-referential and Differential Knowledge Flow in Document Grounded Conversations Lin Xu1 Qixian Zhou2 Jinlan Fu1 Min-Yen Kan1 See-Kiong Ng1

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: