Team OSs System for Dialogue Robot Competition 2022 Yuki Kubo Ryo Yanagimoto Hayato Futase Mikio Nakano Zhaojie Luo Kazunori Komatani Abstract This paper describes our dialogue robot system

2025-05-06 0 0 213.47KB 5 页 10玖币

侵权投诉

Team OS’s System for Dialogue Robot Competition 2022

Yuki Kubo, Ryo Yanagimoto, Hayato Futase, Mikio Nakano, Zhaojie Luo, Kazunori Komatani

Abstract— This paper describes our dialogue robot system,

OSbot, developed for Dialogue Robot Competition 2022. The

dialogue ﬂow is based on state transitions described manually

and the transition conditions use the results of keyword ex-

traction and sentiment analysis. The transitions can be easily

viewed and edited by managing them on a spreadsheet. The

keyword extraction is based on named entity extraction and

our predeﬁned keyword set. The sentiment analysis is text-based

and uses SVM, which was trained with the multimodal dialogue

corpus Hazumi. We quickly checked and edited a dialogue ﬂow

by using a logging function. In the competition’s preliminary

round, our system ended up in third place.

I. INTRODUCTION

We developed a dialogue robot system named OSbot for

Dialogue Robot Competition 2022 [1]. OSbot integrates the

multimodal input/output modules distributed by the compe-

tition organizers [2] and our developed dialogue processing

module. The dialogue processing module is based on Di-

alBB1, a framework for dialogue system development. OSbot

adopts a dialogue strategy similar to that of Team kasuga’s

system, which won second place in the past Dialogue Robot

Competition [3], and reuses some of its modules.

Below are the characteristic features of OSbot:

•OSbot uses state transition network-based dialogue

management to explicitly control the dialogue ﬂow and

avoid unexpected system utterances that do not match

the context at all.

•OSbot uses the results of keyword extraction and senti-

ment analysis for the conditions of state transition. The

results of keyword extraction are also used in system

utterances.

In the following, we explain the architecture of OSbot

and the ideas behind it and show the results obtained in the

preliminary round.

II. OSBOT SYSTEM

In this section, we explain the module components and

module integration of our proposed OSbot system.

A. Module components

As shown in Fig. 1, our OSbot system consists of three

important modules: 1) dialogue management module, 2) sen-

timent analysis module, and 3) keyword extraction module.

1) Dialogue management: The dialogue manager uses a

hand-crafted state transition network. It determines a system

The Institute of Scientiﬁc and Industrial Research (SANKEN),

Osaka University, Osaka, Japan {y-kubo, r-yanagimoto, h-

futase}@ei.sanken.osaka-u.ac.jp, mikio.nakano@c4a.jp, {luo,

komatani}@sanken.osaka-u.ac.jp

1DialBB was developed by C4A Research Institute, Inc. and released for

non-commercial usage at https://github.com/c4a-ri/dialbb

utterance on the basis of the state at that point in time.

It transitions its state to the next state using the results of

keyword extraction and sentiment analysis. It also stores the

keyword extraction results speciﬁed by the state transition

network. The stored keywords are used in generating system

utterance and in deciding the reasons for recommending a

travel destination.

Fig. 2 shows an example dialogue ﬂow. The dialogue

manager uses the results of keyword extraction and sentiment

analysis to determine the state to which to transition. System

utterances are generated depending on the state. In Fig. 2,

OSbot asks about a user’s eating habits. OSbot has three

patterns for the user’s reply. If the user’s reply includes a food

name, OSbot digs deeper into the topic (highlighted in red).

The part of {food}is ﬁlled with a keyword extracted from

the user’s utterance. If the sentiment analysis outputs neutral

or positive, OSbot shows only a comment like “Eating

delicious food makes us happy!” Otherwise, OSbot shows

only backchannels like “Okay.”

2) Keyword extraction: Our keyword extraction module

uses two methods: named entity labeling using GiNZA2

(natural language processing library) and keyword matching

using a predeﬁned keyword set.

First, using GiNZA, the keyword extractor searches for

words labeled “Food,” “Dish,” “First Name,” “Last Name,”

“Noun,” and “Place Name” in a user utterance.

Second, the keyword extractor searches for keywords in

a predeﬁned keyword set. It also tries to ﬁnd a word that

matches or is similar to a keyword in the predeﬁned list. It

measures the similarity by cosine similarity of word vectors.

We used a CBOW model of Word2Vec [4], which was trained

using the Wikipedia corpus and made public by the Shiroyagi

Corporation3.

3) Sentiment analysis: We constructed a sentiment an-

alyzer using the Hazumi corpus [5], which contains multi-

modal dialogues between a system and a user. Our sentiment

analyzer is based on a Support Vector Machine (SVM). Its

output is one of the three sentiment classes (positive, neutral,

and negative). Its input is a 768-dimension vector that is the

output of a pre-trained model (Sentence-BERT [6]) whose

input is a user utterance (in text). We used Sentence-BERT

only as the feature extractor because we are planning to

combine text embedding with features extracted from other

modalities such as audio and vision.

For training the SVM, we used the third-person sentiment

annotations (annotated by multiple people using 7 levels

2https://megagonlabs.github.io/ginza/

3https://github.com/shiroyagicorp/japanese-word2vec-model-builder

arXiv:2210.09928v1 [cs.HC] 18 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：

TeamOS'sSystemforDialogueRobotCompetition2022YukiKubo,RyoYanagimoto,HayatoFutase,MikioNakano,ZhaojieLuo,KazunoriKomataniAbstractThispaperdescribesourdialoguerobotsystem,OSbot,developedforDialogueRobotCompetition2022.Thedialogueowisbasedonstatetransitionsdescribedmanuallyandthetransitionconditionsu...

展开>> 收起<<

Team OSs System for Dialogue Robot Competition 2022 Yuki Kubo Ryo Yanagimoto Hayato Futase Mikio Nakano Zhaojie Luo Kazunori Komatani Abstract This paper describes our dialogue robot system.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！

Team OSs System for Dialogue Robot Competition 2022 Yuki Kubo Ryo Yanagimoto Hayato Futase Mikio Nakano Zhaojie Luo Kazunori Komatani Abstract This paper describes our dialogue robot system

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: