Meeting Decision Tracker Making Meeting Minutes with De-Contextualized Utterances Shumpei Inoue1 Hy Nguyen1 Pham Viet Hoang1 Tsungwei Liu1 Minh-Tien Nguyen2

2025-05-02 0 0 2.34MB 7 页 10玖币

侵权投诉

Meeting Decision Tracker: Making Meeting Minutes with

De-Contextualized Utterances

Shumpei Inoue1, Hy Nguyen1, Pham Viet Hoang1, Tsungwei Liu1, Minh-Tien Nguyen2,∗

1Cinnamon AI, 10th ﬂoor, Geleximco building, 36 Hoang Cau, Dong Da, Hanoi, Vietnam.

{sinoue, hy, hugo, tsungwei}@cinnamon.is

2Hung Yen University of Technology and Education, Hung Yen, Vietnam.

tiennm@utehy.edu.vn

Abstract

Meetings are a universal process to make de-

cisions in business and project collaboration.

The capability to automatically itemize the de-

cisions in daily meetings allows for extensive

tracking of past discussions. To that end, we

developed Meeting Decision Tracker, a proto-

type system to construct decision items com-

prising decision utterance detector (DUD) and

decision utterance rewriter (DUR). We show

that DUR makes a sizable contribution to im-

proving the user experience by dealing with ut-

terance collapse in natural conversation. An in-

troduction video of our system is also available

at https://youtu.be/TG1pJJo0Iqo.

1 Introduction

Obtaining a brief description and salient contents

of meetings is a functionality that can certainly

help business operations. Although automatic

speech recognition enables us to transcribe meeting

records automatically, its transcription is possibly

much more verbose, noisy, or collapsed, and is far

from being utilized in its raw form. Previous re-

search attempted to extract important information

from dialogue, such as decision-making utterances,

(Bak and Oh,2018;Karan et al.,2021), extrac-

tive summaries of online forums (Tarnpradab et al.,

2017;Khalman et al.,2021), or group chat threads

(Wang et al.,2022). Another study, Lugini et al.

(2020) presented a discussion tracker to facilitate

collaborative argumentation in classroom discus-

sion by visualizing discussion transcription.

However, extracted utterances are usually incom-

plete and difﬁcult to understand due to ellipses and

co-references in conversations (Su et al.,2019).

Figure 1(the right) shows an example of a partial

dialogue ending with a decision-related utterance in

our dataset. This shows that objects or indicatives

in utterances in natural conversations are usually

ambiguous, and the meaning of decision-related

∗∗ Corresponding Author.

utterances has a strong dependency on context. Fur-

thermore, especially in Japanese, the format of the

spoken language is often far apart from the written

language because of frank expressions and many

ﬁller phrases. This nature reduces user experience

with the naive use of utterances extracted from dia-

logues. In response to this, Incomplete Utterance

Restoration (IUR) (Pan et al.,2019;Su et al.,2019;

Huang et al.,2021;Inoue et al.,2022) handles the

problem where the model rewrites and restores in-

complete utterances by considering the dialogue

context with promising results. However, we have

yet to see IUR models applied for practical use in

actual business applications.

This paper presents Meeting Decision Tracker

(MDT), a system that automatically generates the

itemized decision list from meeting transcription.

Given the meeting transcription, MDT detects

decision-making utterances and rewrites them to

the de-contextualized utterance, i.e., the written

form with omissions restoration and ﬁller removal.

Such a capability allows users to look back at the

previous meeting contents quickly and have asyn-

chronous communication with no effort from a

minute taker. The system has three crucial charac-

teristics.

•

By combining modules for extracting and

rewriting decision-related utterances, the sys-

tem has a down-to-earth strategy to generate

itemized decision lists from meeting transcrip-

tion. The combination allows us to investigate

the role of IUR in a bigger context with sig-

niﬁcant impact for real business applications.

•

Besides the ordinary task of IUR, our rewriter

handles the translation from the spoken lan-

guage to written language by ﬁltering ﬁller

phrases. It enables users to understand the

decision item at a glance, which contributes

to improving the user experience.

•

Although our system is originally built for

arXiv:2210.11374v1 [cs.CL] 20 Oct 2022

Figure 1: The data ﬂow in our system and the conversation example. The red in the dialogue shows informa-

tion omitted in the decision-related utterance. The blue shows information to be restored by Decision Utterance

Rewriter.

decision utterance itemization, the proposed

method can be applied as a general solution

for information extraction from the dialogue.

2 System Design

The overall system architecture of Meeting Deci-

sion Tracker (MDT) is depicted in Figure 1(the

left). The main function of MDT is to generate

decision items with de-contextualized representa-

tions from the transcription of daily business meet-

ings. MDT comprises of two modules: Decision

Utterance Detector (DUD) and Decision Utterance

Rewriter (DUR). The detector extracts a decision

list from meeting transcription and the rewriter

translates (rewrites) the list to the written format.

Figure 1(the right) shows the example pair of the

input and expected output for the system. The

example indicates two points. First, the transcrip-

tion contains decision-related utterances that can

be used to summarize the content of the meeting.

Second, the decision utterance itself is usually not

self-consistent and comprehensible only after the

utterance is restored by DUR. The next sections

introduce the detector and rewriter.

2.1 Decision utterance detector

The ﬁrst step of the detector is to detect decision-

related utterances from transcription. We formulate

the detection as a sequence labeling problem on

the utterance level and describe the detector in two

steps: input representation and classiﬁcation.

Input representation

The input uses the

sequence of utterances

{u1, u2, ..., uw}

for

the sequential classiﬁcation, where

the window size. Following Cohan et al.

(2019), we used the input representation

{[CLS], u1,[SEP], u2,[SEP], ..., uw,[SEP]}

which contains the [CLS] token at the head of the

whole input and [SEP] tokens at the tail for each

utterance. Then the input was encoded by BERT

(Devlin et al.,2019) for contextual representation.

We set the window size as 5 empirically based on

the observation of results.

Classiﬁcation

There are several studies have ad-

dressed the decision utterance detection as a clas-

siﬁcation. Fernández et al. (2008) deﬁned the

decision-making sub-dialogues as being composed

of several dialogue act tags such as the introduc-

tion of issue, decision adopted/proposed/conﬁrmed,

agreement. Murray and Renals (2008) created ab-

stract describing decisions, actions and problems

of meeting and then associated the utterances used

for abstract as the action item utterances. Chen and

Hakkani-Tur (2016) classiﬁed action items in the

token level following the semantic intent schema.

In this study, the task of decision-related utter-

ance extraction was formulated as binary sequence

labeling on the sentence level, different from Fer-

nández et al. (2008). This is because we want to

keep a simple setting to conﬁrm the efﬁciency of

IUR in actual cases. To take advantage of con-

text, we followed Cohan et al. (2019) to jointly en-

code consecutive utterances. Preceding utterances

leading to decision are essential because followed

by Fernández et al. (2008), we hypothesize that

the particular kinds of patterns of conversation co-

occur with decision. Utterances following decision

are also important since afﬁrmative response by

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：

MeetingDecisionTracker:MakingMeetingMinuteswithDe-ContextualizedUtterancesShumpeiInoue1,HyNguyen1,PhamVietHoang1,TsungweiLiu1,Minh-TienNguyen2;1CinnamonAI,10thoor,Geleximcobuilding,36HoangCau,DongDa,Hanoi,Vietnam.{sinoue,hy,hugo,tsungwei}@cinnamon.is2HungYenUniversityofTechnologyandEducation,HungY...

展开>> 收起<<

Meeting Decision Tracker Making Meeting Minutes with De-Contextualized Utterances Shumpei Inoue1 Hy Nguyen1 Pham Viet Hoang1 Tsungwei Liu1 Minh-Tien Nguyen2.pdf

共7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！

Meeting Decision Tracker Making Meeting Minutes with De-Contextualized Utterances Shumpei Inoue1 Hy Nguyen1 Pham Viet Hoang1 Tsungwei Liu1 Minh-Tien Nguyen2

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: