slides in ppt - NaCTeM

hurriedtinkleΤεχνίτη Νοημοσύνη και Ρομποτική

15 Νοε 2013 (πριν από 3 χρόνια και 11 μήνες)

106 εμφανίσεις

Japanese
-
German Workshop on NLP in Sapporo Jul 4
-
5, 2003

Human
-
human Communication
and Knowledge Management

Kazuo Sumita

Corporate R&D Center

TOSHIBA Corp.

Japanese
-
German Workshop on NLP in Sapporo Jul 4
-
5, 2003

Agenda


Toshiba’s NLP R&D activities


Toshiba’s knowledge management system


Shortage of the current KM system


New approaches


“GroupScribe” for
group communication
management


“MKIDS” for multi
-
modal knowledge sharing


Future work

Japanese
-
German Workshop on NLP in Sapporo Jul 4
-
5, 2003

TOSHIBA NLP R&D activities 1/3


Machine translation


English
-
Japanese MT :

Dictionary tuning using large
corpora


Chinese
-
Japanese MT :

Hybrid framework using rule
-

and statistics
-
based approaches


English
-
Chinese MT :

Prototyping using the framework
for EJMT


Japanese
-
English speech translation :
Prototyping


MT products


The

Honyaku
(PKG software), MT server (Japan
Infoseek, lycos, excite, @nifty, Japan Patent Office’s
IPDL, …), Engine License to other companies

Japanese
-
German Workshop on NLP in Sapporo Jul 4
-
5, 2003

TOSHIBA NLP R&D activities 2/3


Information retrieval and knowledge mining


Natural language based information retrieval


Question answering


Cross language information retrieval


Text mining: document clustering, categorization,
information extraction


Knowledge mining products


KnowledgeMeister

:
KM software which can work

with several other systems (IBM WebSphere portal,
Oracle9iAS Portal, Livelink, Fuji Xerox DocuCentre,
Microsoft SharePointPortal server, Exchange, …)


NewsWatch : Information filtering of news articles

Japanese
-
German Workshop on NLP in Sapporo Jul 4
-
5, 2003

TOSHIBA NLP R&D activities 3/3


Speech processing


Speech synthesis:
Provide human voice quality and
naturalness, multi
-
lingual(American English, British English,
Chinese, Dutch, French, German, Italian, Spanish and
Japanese), small memory and low computational power


Robust speech recognition :
High performance under
noisy environments, multi
-
lingual


Japanese speech dictation :
Speaker independent, high
recognition rate without enrolment


Speech processing products


Middleware for car navigation systems, mobile
equipments, game software, LaLaVoice(PKG
software)

Japanese
-
German Workshop on NLP in Sapporo Jul 4
-
5, 2003

Knowledge sharing system

Features


NLP based information retrieval


Hierarchical clustering of accumulated documents


Categorizing newly input document


Various functions for knowledge sharing

KnowledgeMeister
TM


Chishiki
-
kyouyuu


knowledge sharing

Japanese
-
German Workshop on NLP in Sapporo Jul 4
-
5, 2003

Knowledge sharing system

Questioner

Tell me how to write

an equipment plan.

Language/

intention

understanding



Information

retrieval

Office

knowledge

Personal

know
-
how

Intranet

Experienced person

Japanese
-
German Workshop on NLP in Sapporo Jul 4
-
5, 2003

Semantic Roles: Example 1/2

Examples of search requests from a knowledge
sharing in TOSHIBA corporate R&D center
(English translations)



When

do we have to leave the dormitory?”


Who

can apply for a child
-
care leave?”


Where

can we have Chinese food in Kawasaki?”

Extracted semantic roles :


time, person, place

Japanese
-
German Workshop on NLP in Sapporo Jul 4
-
5, 2003

Semantic Roles: Example 2/2

Example request from customer support
(English translation)


“I am a dynabook XX user.
I’ve just pressed the
power button without shutting it down.

Now it
displays an error message XXX”
.


Extracted semantic roles:


Background
Action

Symptom

Japanese
-
German Workshop on NLP in Sapporo Jul 4
-
5, 2003

Insufficiency of the current KM
systems

Treatment of knowledge exchange in human
-
human communication


Knowledge and information exchanged by e
-
mail


GroupScribe


Multi
-
modal knowledge such as video and
speech


MKIDS

Japanese
-
German Workshop on NLP in Sapporo Jul 4
-
5, 2003

Convert human
-
human
communication to sharable
knowledge

Education

Office

Sharable knowledge

Interactive
e
-
Learning

Multi
-
modal
knowledge sharing

Reusable knowledge

Communication

E
-
mail

F2F dialogue

New communication

Japanese
-
German Workshop on NLP in Sapporo Jul 4
-
5, 2003

CIKLE :
Community Knowledge Ware

Community
-
based Interactive Knowledge Leveraging Environment


Drive communication
-
knowledge cycle


Extract and leverage knowledge from/through
communication


Find and recommend knowledge to activate
communication

Extract and

leverage

knowledge

Find and

recommend

knowledge

Knowledgebase (Stock
-
type)

Communication (Flow
-
type)

Comment / Post

Extract / Edit

Message

thread

Linked

document

Flow
-
Stock combination structure

Bind

Japanese
-
German Workshop on NLP in Sapporo Jul 4
-
5, 2003

The CIKLE solution

delivers…


Collaborative knowledge leveraging


Edit knowledge with a community consensus


Create knowledge with dialogue summarization engine


Publish sharable knowledge from even closed
communities


Provide dual view: knowledge and its context


Retrieval of relevant knowledge in a natural way


Accept natural language queries


Give priority to documents than messages

Japanese
-
German Workshop on NLP in Sapporo Jul 4
-
5, 2003

Information extraction (GroupScribe)

Enhancement of the summarization function of CIKLE

Japanese
-
German Workshop on NLP in Sapporo Jul 4
-
5, 2003

Rule based extraction


Surface expressions in each message


Reference relations between messages

Japanese
-
German Workshop on NLP in Sapporo Jul 4
-
5, 2003

Knowledge sharing practice in
TOSHIBA corp.

CIKLE

07/2000


05/2003

CIKLE
gs

05/2003


Japanese
-
German Workshop on NLP in Sapporo Jul 4
-
5, 2003

Experienced person

(Answerer)

Questioner

Question and
Answer

Knowledge DB

○○を××する方法

カドの


=2
-
5

処理方




Authoring tool

How should I
do to manage
the
××

when
○○

?

The way of
managing
××

is…

The way of
managing
××

is


How should I do
to manage the
××

when
○○

?


Accumulation of
the answering
video image


Refinement of the
knowledge





Retrieval and reuse
of the accumulated
knowledge



Multi
-
modal knowledge sharing
system (MKIDS)

Japanese
-
German Workshop on NLP in Sapporo Jul 4
-
5, 2003

Multi
-
modal knowledge sharing

Questioner’s side

Answerer’s side

Japanese
-
German Workshop on NLP in Sapporo Jul 4
-
5, 2003

System configuration

Knowledge DB

Questioner side

Media capture

Distribution

QA retrieval

Answerer side

M
edia capture

Semantic role analysis

Video dialogue

Native XML database

Authoring tool

Japanese
-
German Workshop on NLP in Sapporo Jul 4
-
5, 2003

Snapshot of a dialogue

Semantic role
analysis
result

Speech
recognition
result

Ano kikitai
no desuga

(Ur, I have
a question.)

Hai nande
shou

(Hello! May
I help you?)

Questioner

Answerer

[2]
question

[2]
nodding

Ano hito no tamedakedo

(For that person.)

Haittande nao

(Something entered.)

Japanese
-
German Workshop on NLP in Sapporo Jul 4
-
5, 2003

Future work


Application of the systems to several real
works and the evaluation


Improvement of the scalability and
robustness


Adoption of more natural language
techniques such as IE of named entities for
generating effective summary