Question-Interlocutor Scope Realized Graph Modeling over Key Utterances for Dialogue Reading Comprehension Jiangnan Li12 Mo Yu3 Fandong Meng3 Zheng Lin12 Peng Fu1 Weiping Wang1 Jie Zhou3

2025-04-29 0 0 707.56KB 9 页 10玖币

侵权投诉

Question-Interlocutor Scope Realized Graph Modeling over Key Utterances for

Dialogue Reading Comprehension

Jiangnan Li1,2, Mo Yu3, Fandong Meng3, Zheng Lin1,2, Peng Fu1, Weiping Wang1, Jie Zhou3

1Institute of Information Engineering, Chinese Academy of Sciences, Beijing, China

2School of Cyber Security, University of Chinese Academy of Sciences, Beijing, China

3Pattern Recognition Center, WeChat AI, Tencent Inc, China

{lijiangnan, linzheng, fupeng, wangweiping}@iie.ac.cn, {moyumyu,fandongmeng,withtomzhou}@tencent.com

Abstract

In this work, we focus on dialogue reading comprehension

(DRC), a task extracting answer spans for questions from

dialogues. Dialogue context modeling in DRC is tricky due

to complex speaker information and noisy dialogue context.

To solve the two problems, previous research proposes two

self-supervised tasks respectively: guessing who a randomly

masked speaker is according to the dialogue and predict-

ing which utterance in the dialogue contains the answer. Al-

though these tasks are effective, there are still urging prob-

lems: (1) randomly masking speakers regardless of the ques-

tion cannot map the speaker mentioned in the question to

the corresponding speaker in the dialogue, and ignores the

speaker-centric nature of utterances. This leads to wrong an-

swer extraction from utterances in unrelated interlocutors’

scopes; (2) the single utterance prediction, preferring utter-

ances similar to the question, is limited in ﬁnding answer-

contained utterances not similar to the question. To alleviate

these problems, we ﬁrst propose a new key utterances extract-

ing method. It performs prediction on the unit formed by sev-

eral contiguous utterances, which can realize more answer-

contained utterances. Based on utterances in the extracted

units, we then propose Question-Interlocutor Scope Realized

Graph (QuISG) modeling. As a graph constructed on the text

of utterances, QuISG additionally involves the question and

question-mentioning speaker names as nodes. To realize in-

terlocutor scopes, speakers in the dialogue are connected with

the words in their corresponding utterances. Experiments on

the benchmarks show that our method can achieve better and

competitive results against previous works.

Introduction

Beyond the formal forms of text, dialogues are one of the

most frequently used media that people communicate with

others to informally deliver their emotions (Poria et al.

2019), opinions (Cox et al. 2020), and intentions (Qin et al.

2021). Moreover, dialogues are also a crucial type of infor-

mation carriers in literature, such as novels and movies (Ko-

cisk´

y et al. 2018), for people to understand the characters

and plots (Sang et al. 2022) in their reading and entertain-

ment behaviors. Therefore, how to comprehend dialogues is

a key step for machines to act like human.

Joint work with Pattern Recognition Center, WeChat AI, Tencent

Inc, China.

#10

#11

Monica Geller：Anyway ,how did it go with Kate?

Jeoy Tribbiani：[Oh, it was great! Yeah, I-Iwalked

her home, and it was amazing ...... And we stayed up

all night talking, ...... crazy about each other!]

Monica Geller：Joey, you had the night!

Joey Tribbiani：What?

Q1: What Joey did after his date?

…

#12

#13

#14

#15

Rachel Green：You know what?

Jeoy Tribbiani：What?

Rachel Green：There were times when it was not

even me

Scene note：Chandler and Joey laughs, until they

look at each other then recoil in horror.

Q2: Why do Joey and Chandler recoil in horror in response

to Rachel ?

…

text Baseline

prediction

text Gold answer

Figure 1: Two questions with related dialogue clips that the

baseline SelfSuper (Li and Zhao 2021) fails. Utter. #9 is too

long, so we omit some parts of the utterance.

Despite the importance and value of dialogues, reading

comprehension over dialogues (DRC) lags behind those over

formal text, e.g., news articles and Wikipedia text.1There

are several reasons for the challenges in dialogue reading

comprehension. (1) As shown in Fig. 1, utterances in dia-

logues involve mostly informal oral language. (2) An ut-

terance itself is usually short and incomplete, and therefore

understanding it highly depends on its dialogue context. Fi-

nally, (3) dialogue texts can be ﬂuctuated. For example, peo-

ple make a slip of language sometimes, and may evoke emo-

tional expression. The aforementioned challenges in (1) and

(3) can be alleviated with pretrained models with domain-

speciﬁc training data and enhancement of robustness. How-

ever, the challenge (2) is a major scientiﬁc problem in DRC.

In previous works, Li and Zhao (2021) (abbreviated as

SelfSuper) point out that dialogue context modeling in

DRC faces two challenges: complex speaker information

and noisy question-unrelated context. For speaker model-

1Note there is a direction of conversational question answer-

ing (Reddy, Chen, and Manning 2018; Choi et al. 2018) differing

from the DRC task here. For the former, QA are formed as a dia-

logue, and the model is required to understand Wikipedia articles

to derive the answers.

arXiv:2210.14456v1 [cs.CL] 26 Oct 2022

ing, SelfSuper design a self-supervised task guessing who a

randomly masked speaker is according to the dialogue con-

text (e.g., masking “Monica Geller” of #10). To reduce

noise, SelfSuper design another task to predict whether an

utterance contains the answer. Although decent performance

can be achieved, several urging problems still exist.

Firstly, speaker guessing does not aware the speaker in-

formation in questions and the interlocutor scope. As ran-

domly masking is independent to the question, it cannot tell

which speaker in the dialogue is related to the speaker men-

tioned in the question, e.g., Joey Tribbiani to Joey in

Q1 of Fig. 1. As for the interlocutor scope, we deﬁne it as

the utterances said by the corresponding speaker. We point

that utterances have a speaker-centric nature: First, each ut-

terance has target listeners. For example, in Utter. #10 of

Fig. 1, it requires to understand that Joey is a listener, so

“you had the night” is making fun of Joey from

Monica’s scope. Second, an utterance reﬂects the mes-

sage of experience of its speaker. For example, to answer

Q1 in Fig. 1, it requires to understand “stayed up all

night talking” is the experience appearing in Joey’s

scope. Due to ignoring the question-mentioned interlocutor

and its scope, SelfSuper provides a wrong answer.

Secondly, answer-contained utterance (denoted as key ut-

terance by SelfSuper) prediction prefers utterances similar

to the question, failing to ﬁnd key utterances not similar to

the question. The reason for this is that answers are likely

to appear in utterances similar to the question. For example,

about 77% questions has answers in top-5 utterances similar

to the question according to SimCSE (Gao, Yao, and Chen

2021) in the dev set of FriendsQA (Yang and Choi 2019).

Furthermore, the utterances extracted by the key utterance

prediction have over 82% overlaps with the top-5 utterances.

Therefore, there are considerable key utterances have been

ignored, leading to overrated attention to similar utterances,

e.g., Q2 in Fig. 1. In fact, answer-contained utterances are

likely to appear as the similar utterances or near the simi-

lar utterances because contiguous utterances in local context

tend to be in a topic relevant to the question. However, the

single utterance prediction cannot realize this observation.

To settle aforementioned problems, so that more answer-

contained utterances can be found and answering process

realizes the question and interlocutor scopes, we propose a

new pipeline framework for DRC. We ﬁrst propose a new

key utterances extracting method. The method slides a win-

dow through the dialogue, where contiguous utterances in

the window are regarded as a unit. The prediction is made

on these units. Once a unit containing the answer, all utter-

ances in it are selected as key utterances. Based on these

extracted utterances, we then propose Question-Interlocutor

Scope Realized Graph (QuISG) modeling. Instead of treat-

ing utterances as a plain sequence, QuISG constructs a graph

structure over the contextualized text embeddings. The ques-

tion and speaker names mentioned in the question are ex-

plicitly present in QuISG as nodes. The question-mentioning

speaker then connects with its corresponding speaker in the

dialogue. Furthermore, to remind the model of interlocu-

tor scopes, QuISG connects every speaker node in the di-

alogue with words from the speaker’s scope. We verify our

model on two representative DRC benchmarks, FriendsQA

and Molweni (Li et al. 2020). Our model achieves better and

competitive performance against baselines on both bench-

marks, and further experiments indicate the efﬁcacy of our

proposed method.

Related Work

Dialogue Reading Comprehension. Unlike traditional

Machine Reading Comprehension (Rajpurkar et al. 2016),

Dialogue Reading Comprehension (DRC) aims to answer

a question according to the given dialogue. There are sev-

eral related but different types of conversational question

answering: CoQA (Reddy, Chen, and Manning 2018) con-

versationally asks questions after reading Wikipedia arti-

cles. QuAC (Choi et al. 2018) forms a dialogue of QA

between a student and a teacher about Wikipedia arti-

cles. DREAM (Sun et al. 2019) tries to answer multi-

choice questions over dialogues of English exams. To un-

derstand characteristics of speakers, Sang et al. (2022) pro-

pose TVShowGuess in a multi-choice style to predict un-

known speakers in dialogues. Conversely, we focus on DRC

extracting answer spans from a dialogue for an independent

question (Yang and Choi 2019). For our focused DRC, as

dialogues are a kind of special text, Li and Choi (2020) pro-

pose several pretrained and downstream tasks based on ut-

terances in dialogues. To consider coreference and relation-

ship of speakers, Liu et al. (2020) introduce the two types

of knowledge from other dialogue-related tasks. Besides, Li

et al. (2021); Ma, Zhang, and Zhao (2021) model the knowl-

edge of discourse structure for dialogues. To model the com-

plex speaker information and noisy dialogue context, Li and

Zhao (2021) propose two self supervised tasks, i.e., masked-

speaker guessing and key utterance prediction. However, ex-

isting work ignores explicitly modeling the question and

speaker scopes and suffers from low key-utterance coverage.

Dialogue Modeling with Graph Representations. In

many QA tasks (Yang et al. 2018; Talmor et al. 2019), graphs

are the main carrier for reasoning (Qiu et al. 2019; Fang et al.

2020; Yasunaga et al. 2021). As for dialogue understanding,

graphs are still a hotspot for various purposes. In dialogue

emotion recognition, graphs are constructed to consider the

interactions between different parties of speakers (Ghosal

et al. 2019; Ishiwatari et al. 2020; Shen et al. 2021). In di-

alogue act classiﬁcation, graphs model the cross-utterances

and cross-tasks information (Qin et al. 2021). In dialogue se-

mantic modeling, Bai et al. (2021) extend AMR (Banarescu

et al. 2013) to construct graphs for dialogues. As for our

focused DRC, graphs are constructed for knowledge propa-

gation between utterances by works (Liu et al. 2020; Li et al.

2021; Ma, Zhang, and Zhao 2021) mentioned above.

Framework

Task Deﬁnition

Given a dialogue consisting of Nutterances: D=[utter1

, utter2, ..., utterN], the task aims to extract the answer span

afor a question q= [qw1, qw2, ..., qwLq]from D, where

qwiis the i-th word in qand Lqis the length of q. In D,

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：

Question-InterlocutorScopeRealizedGraphModelingoverKeyUtterancesforDialogueReadingComprehensionJiangnanLi1,2,MoYu3,FandongMeng3,ZhengLin1,2,PengFu1,WeipingWang1,JieZhou31InstituteofInformationEngineering,ChineseAcademyofSciences,Beijing,China2SchoolofCyberSecurity,UniversityofChineseAcademyofScience...

展开>> 收起<<

Question-Interlocutor Scope Realized Graph Modeling over Key Utterances for Dialogue Reading Comprehension Jiangnan Li12 Mo Yu3 Fandong Meng3 Zheng Lin12 Peng Fu1 Weiping Wang1 Jie Zhou3.pdf

共9页,预览2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！

Question-Interlocutor Scope Realized Graph Modeling over Key Utterances for Dialogue Reading Comprehension Jiangnan Li12 Mo Yu3 Fandong Meng3 Zheng Lin12 Peng Fu1 Weiping Wang1 Jie Zhou3

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: