Efficient Natural Efficient Natural Language Processing Language Processing

scarfpocketΤεχνίτη Νοημοσύνη και Ρομποτική

24 Οκτ 2013 (πριν από 4 χρόνια και 20 μέρες)

67 εμφανίσεις

EfficientNatural
Efficient

Natural
LanguageProcessing
Language

Processing
SeminarWS2011/2012
Seminar

WS

2011

/

2012
Session 1
,
Wednesda
y
October 26
,
2011
,y,
(Introduction, Organization, Topic assignment)
Prof. Dr. Hannah Bast
Chair for Al
g
orithms and Data Structures
g
Department of Computer Science
University of Freiburg
T
o
p
ic of this seminar 1/2
p

Natural Language Processing (NLP)
–Do not just treat text as a string of characters, but try to
"understand" it ... up to a certain level

NLP is needed by lots of applications
–In particular, by a new kind of semantic search we are
currently developing at our chair show Broccoli demo

But NLP is computationally expensive
–Advanced methods take seconds / minutes for a single sentence
–The English Wikipedia alone has ~ 200 million sentences
–Even at1 second / sentence 6 years
–This raises the central (for this seminar) question:
Whtllfdtdiiiliitd
Wh
a
t

l
eve
l
o
f
un
d
ers
t
an
di
ng can we ga
i
n
i
n
li
m
it
e
d
time ... say 10 –50 milliseconds / sentence ?
2
T
o
p
ic of this seminar 2/2
p

Your task in this seminar in some more detail
–NLPnaturally splits into various subfields Elmar's slides
POS tagging, Chunking, Parsing, Question Answering, ...
–Various techniques / implementations for each
–Task 1:Get an overview of existing techniques + code
–Task 2:Evaluate performance of some state-of-the-art code
–Task 3:Identify the bottlenecks wrt speed
3
Or
g
anization of the Seminar
g

One or two presentations per session
–Today: introduction of topics + topic assignment
–After today, we have 12 sessions left
–Next session is on November 16 (threeweeks from now)
–Time schedule for your presentation: next slide
4
T
ime schedule for
y
our
p
resentation
yp

3 weeks before your presentation
–Start collecting material and make a plan of what you
want to talk about

2w敥歳b敦潲e祯畲p牥獥湴慴楯r

2

睥敫w

扥景牥

祯畲

灲敳敮瑡瑩潮
–䵥整⁷楴栠畳
䡡湮慨u⬠䕬浡爩+慮搠灲敳敮琠祯畲⁰污渻a
pleasedonotwasteourtime
bycomingunprepared
please

do

not

waste

our

time

by

coming

unprepared
–In the week that follows, work out all the necessary
details and play around (extensively) with software
–Prepare an outline of your presentation

1 week before your presentation
f
–Meet with us again, and present your
f
indings and the
outline of your presentation (tentative slides)

Intheweekthatfollowsfinishyourworkandthe
In

the

week

that

follows
,
finish

your

work

and

the

presentation
5
Your
p
resentation
p

Guidelines
–You have 30 minutes for your talk + 30 minutes dicussion
–Use slides in PPTor PDF
–Your talk will be recorded
6
Challen
g
es
g

Research
–You have to collect yourself interesting and relevant material,
we only provide the general topic and give some pointers

Understand
–Get a decent overview over your topic and understand what
you will be talking about

Presentation
–Present your material in an interesting manner, don't forget
that you have an audience
Iiht

I
ns
i
g
ht
–To identify the bottlenecks wrt speed it's not enough to just
runthecodeyouhavetounderstandwhatgoeson
inside
run

the

code
,
you

have

to

understand

what

goes

on

inside
7
Our role

We will help you, don't worry
–In your first meeting with us (two weeks before your
talk) we will help you focus on a good selection of
tilftlk
ma
t
er
i
a
l

f
or your
t
a
lk
–We will also give you feedback and advice on the
structureandcontentsofyourslides
structure

and

contents

of

your

slides
–And, of coure, we try to help when you have problems
understandin
g
somethin
g
gg
–However, the initiativehas to come from you !
8
9