Upload
susanna-pitts
View
216
Download
3
Embed Size (px)
Citation preview
ntcir at clef 2010-09-21 Noriko Kando 1
What is happening at NTCIR
Noriko KandoNational Institute of Informatics, Japan
http://research.nii.ac.jp/ntcir/kando@nii. ac. jp
Thanks for Teruko Mitamura, Tetsuya Sakai, Fred Gey, Yohei Seki, Daisuke Ishikawa, Atsushi Fujii, Hidetsugu Nanba, Terumasa Ehara for preparing slides
ntcir at clef 2010-09-21Noriko Kando 2
NTCIR: NTCIR: NII Test Collection for Information RetrievalNII Test Collection for Information Retrieval
Research Infrastructure for Evaluating IA
Data sets (Test collections or TCs) Scientific, news, patents, and web Chinese, Korean, Japanese, and English
Project started in late 1997
Tasks (Research Areas) IR: Cross-lingual tasks, patents, web, GeoQA : Monolingual tasks, cross-lingual tasks Summarization, trend info., patent maps Opinion analysis, text mining
Once every 18 months
A series of evaluation workshops designed to enhance research in information-access technologies by providing an infrastructure for large-scale evaluations. ■Data sets, evaluation methodologies, and forum
Community-based Research Activities
NTCI R 1 2 3 4 5 6 7 8
' 99'01'02'04'05'07'08'09-
■ Community QA
■ ■ ■ Opinion Analysis
Module- Based ■ ■ Cross-Lingual QA + I R
■ Geo Temporal
■ ■ ■ ■ □ Patent
■ ■ ■ Complex/ Any Types
■ ■ Dialog
■ ■ ■ ■ Cross-Lingual
■ ■ ■ ■ Factoid, List
■ ■ ■ ■ ■ Text Mining / Classification
■ ■ ■ Trend I nf o Visualization
■ ■ ■ Text Summarization
Web ■ ■ ■ Web
■ ■ Statistical MT
■ ■ ■ ■ ■ ■ ■ ■ Cross-Lingual I R
■ ■ ■ ■ ■ ■ ■ ■ Non-English Search
Text Retrieval ■ ■ ■ ■ ■ ■ ■ ■ Ad Hoc I R, I R f or QA
The Years the meetings were held. The tasks started 18 months bef ore
CrosslingualRetrieval
User Generated
Contents
IR for FocusedDomain
Summarization /Consolidation
Question
Answering
Tasks at Past NTCIRs
ntcir at clef 2010-09-21 3Noriko Kando
NTCIR-8 Tasks (2008.07—2009.06)
1. Advanced CL Info Access- QA(CsCtJE->CsCtJ ) ※ Any Types of
Questions- IR for QA (CsCtJE->CsCtJ ) GeoTime (E, J) Geo Temporal Information
3. Focused Domain : Patent-Patent Translation; English -> Japanese ※ Statistical MT, The World-Largest training data (J-E sentence alignment), Summer School, Extrinsic eval by CLIR
-Patent Mining papers -> IPC-Evaluation of SMT
New
Th
e 3
rd In
t’l W o
n E
valu
atin
g
Info
rmatio
n A
ccess (E
VIA
)
refe
reed
New
Newntcir at clef 2010-09-21 4Noriko Kando
NTCIR-7 & -8 Program Committee
Mark Sanderson, Doug Oard, Atsushi Fujii, Tatsunori Mori, Fred Gey, Noriko Kando (and Ellen Voorhees, Sung Hyun Myaeng, Hsin-Hsi Chen, Tetsuya Sakai)
ntcir at clef 2010-09-21 5Noriko Kando
NTCIR-8 Coordination• NTCIR-8 is coordinated by NTCIR Project at NII, Japan.
The following organizations contribute to the organization of NTCIR-8 as Task Organizers
- Academia Sinica - Carnegie Mellon Univ- Chinese Academy of Science- Hiroshima City University- Hitachi, Co Ltd.- Hokkai Gakuen University- IBM- Microsoft Research Asia- National Institute of Information and Communication Technology- National Institute of Informatics- National Taiwan Univ
-National Taiwan Ocean Univ- Oki Electonic Co.- Tokyo Institute of Technology- Tokyo Univ- Toyohashi Univ of Technology and Science- Univ of California Barkeley-Univ of Tsukuba- Yamanashi Eiwa College- Yokohama National University
ntcir at clef 2010-09-21 6Noriko Kando
[CCLQA]Carnegie Mellon UnivDalian Univ of TechnologyNational Taiwan Ocean UnivShenyan Institute of
Aeronautical Engineering
Univ of TokushimaWuhan Univ
[IR4QA]Carnegie Mellon UnivChaoyang Univ of
TechnologyDalian Univ of TechnologyDublin City UnivInner Mongolia UnivQueensland Univ of
TechnologyShenyan Inst of Aeronautical
EngineeringTrinity College DublinUniv California, BerkeleyWuhan UnivWuhan Univ (Computer
School)Wuhan Univ of Science and
Technology
[GeoTime]Dublin City UnivHokkaido UnivINESC-ID, PorugalInternational Inst of Technology,
HyderbadKieo UnivNataional Inst of Materials ScienceOsaka Kyoiku UnivUniv California, BerkeleyUniv of IowaUniv of LisbonYokohama City Univ
[MOAT]Beijing Uni of Posts and
TelecommunicationsChaoyang Univ of TechnologyChinese Univ of HK+ Tsinghua UnivCity Univ of Hong Kong (2 groups)Hong Kong Polytechnic UnivKAISTNational Taiwan UnivNEC Laboratories ChinaPeking UnivPohang Univ of Sci and TechSICSToyohashi Univ of TechnologyUniv of AlicanteUniv of NeuchatelYuan Ze Univ
[Patent Mining]Hiroshima City UnivHitachi, Ltd.IBM Japan, Ltd.Institute of Scientific and
Technical Information of China
KAISTNational Univ of SingaporeNECShanghai Jiao Tong UnivShenyang Institute of
Aeronautical EngineeringToyohashi Univ of TechnologyUniv of Applied Sciences - UNIGE
[Patent Translation]Dublin City University, CNGLHiroshima City UniversityKyoto UniversityNiCTPohang Univ of Sci and Techtottori universityToyohashi University of
TechnologyYamanashi Eiwa College
[Community QA]Microsoft Research AsiaNational Institute of InformaticsShirayuri College
NTCIR-8 Active Participantsntcir at clef 2010-09-21 7Noriko Kando
Complex Cross-lingual Question Answering(CCLQA) Task
IR and QA communities can collaborate
Different teamscan exchange and create a “dream-team” QA system
Small teams that do not possess an entire QA system can contribute
ntcir at clef 2010-09-21 8Noriko Kandohttp://aclia.lti.cs.cmu.edu/ntcir8/
9 Noriko Kando
Evaluation Topics – any types of questions -
Type # Example Question Related Past NTCIR Task
DEFINITION 10 What is the Human Genome Project?
ACLIA
BIOGRAPHY 10 Who is Howard Dean? ACLIA
RELATIONSHIP 20 What is the relationship between Saddam Hussein and Jacques Chirac?
ACLIA
EVENT 20 What are the major conflicts between India and China on border issues?
ACLIA
WHY 20 Why doesn't U.S. ratify the Kyoto Protocol?
QAC-4
PERSON 5 Who is the Finland's first woman president?
QAC 1-3, CLQA 1,2
ORGANIZATION 5 What is the name of the company that produced the first Fairtrade coffee?
QAC 1-3, CLQA 1,2
LOCATION 5 What is the name of the river that separates North Korea from China?
QAC 1-3, CLQA 1,2
DATE 5 When did Queen Victoria die? QAC 1-3, CLQA 1,2
ntcir at clef 2010-09-21
CT/JA-T IR4QA run rankings
Mean Q
Mean Q
Mean nDCG
Mean nDCG
ntcir at clef 2010-09-21 10Noriko Kando
11 Noriko Kando
CCLQA Human Evaluation Preliminary Results: JA-JAJA-JA Runs ALL
LTI-JA-JA-01-T 0.1069
LTI-JA-JA-02-T 0.1443
LTI-JA-JA-03-T 0.1438
JA-JA automatic evaluation
JA-JA Runs ALL
LTI-JA-JA-01-T 0.2024
LTI-JA-JA-02-T 0.2259
LTI-JA-JA-03-T 0.2252
ntcir at clef 2010-09-21
12 Noriko Kando
Effect of Combination;IR4QA+CCLQA JA-JA Collaboration Track: F3 score
based on automatic evaluation
CCLQA
LTI
IR4QA
BRKLY-JA-JA-01-DN 0.2934
BRKLY-JA-JA-02-T 0.2686
BRKLY-JA-JA-03-DN 0.2074
BRKLY-JA-JA-04-DN 0.3000
BRKLY-JA-JA-05-T 0.2746
ntcir at clef 2010-09-21
NTCIR-GEOTIMEGEOTEMPORAL INFORMATION RETRIEVAL
(New Track in NTCIR Workshop 8)
Fredric Gey and Ray Larson and Noriko KandoRelevance judgments system by Jorge Machado and Hideki
ShimaEvaluation : Tetsuya Sakai
Judgments: U Iowa, U Lisbon, U California Barkelay, NII Search with a specific focus on Geography + To distinguish from past GIR evaluations, we introduced a temporal component
Asian language geographic search has not previously been evaluated, even though about 50 percent of the NTCIR-6 Cross-Language topics had a geographic component (usually a restriction to a particular country).
ntcir at clef 2010-09-21 13Noriko Kando
• Japanese Runs submitted by eight groups (two anonymous pooling supporters)
NTCIR-GeoTimePARTICIPANTS
Team Name Organization
Anonymous Anonymous
BRKLY University of California, Berkeley, USA
FORST Yokohama National University, JAPAN
HU-KB Hokkaido University, JAPAN
KOLIS Keio University, JAPAN
ANON2 Anonymous submission, group 2
M National Institute of Materials Science, JAPAN
OKSAT Osaka Kyoiku University, JAPAN
ntcir at clef 2010-09-21 14Noriko Kando
• English Runs submitted by six groups
NTCIR-GeoTimePARTICIPANTS
Team Name Organization
BRKLY University of California, Berkeley, USA
DCU Dublin City University, IRELAND
IITH International Institute of Technology, Hyderbad, INDIA
INESC National Institute of Electroniques and Computer Systems, Lisbon, PORTUGAL
UIOWA University of Iowa, USA
XLDB University of Lisbon, PORTUGAL
ntcir at clef 2010-09-21 15Noriko Kando
• BRKLY: baseline approach, probablistic + psued relevance feedback•DCU, IITH, XLDB (U Lisbon) : geographic enhancements•KOLIS (Keio U) : counting the number of geographic and temporal expression in top-ranked docs in initial search, then re-rank•FORST (Yokohama Nat U): utilize factoid QA technique to question decomposition•HU-KB (Hokkaido U), U Iowa: Hybrid approach combinging probablistic model and weighted boolean query formulation
NTCIR-GeoTime Approached
ntcir at clef 2010-09-21 16Noriko Kando
Most Difficult English topic (21): When and where were the 2010 Winter Olympics host city location
announced?
NTCIR-GeoTime: ENGLISH TOPIC DIFFICULTY by Average Precision
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
21 18 22 15 10 12 20 5 11 25 17 4 24 2 14 9 16 23 13 6 8 19 3 7 1
AP
Q
nDCG
Per-topic AP, Q and nDCG averaged over 25 English runs for 25 topics sorted by topic difficulty
(AP ascending)
ntcir at clef 2010-09-21 17Noriko Kando
Most Difficult Japanese topic 18: What date was a country was invaded by the United States in 2002?
NTCIR-GeoTime: JAPANESE TOPIC DIFFICULTY by Average Precision
Per-topic AP, Q and nDCG averaged over 34 Japanes runs for 24 topics sorted by topic difficulty
(AP ascending)
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1814252220121523131124 4 2 8 191016 5 21 6 1 9 3 7
AP
Q
nDCG
ntcir at clef 2010-09-21 18Noriko Kando
NTCIR-GeoTime: TOPIC per TOPIC Analysis on Japanese
ntcir at clef 2010-09-21 19Noriko Kando
0.00
0.20
0.40
0.60
0.80
1.00
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 18 19 20 21 22 23 24 25
HU-KB-JA-JA-03-D
KOLIS-JA-JA-04-D
FORST-JA-JA-04-D
BRKLY-JA-JA-02-D
nD
CG
Topic ID (#17 is missing)
• Geographic reference resolution is difficult enough, but
• More difficult to process temporal expression (“last Wednesday”) references
•Can indefinite answers be accepted (“a few hours”)?
•Need Japanese Gazetteers•Need NE annotated corpus for further refinement
NTCIR-GeoTime CHALLENGES
ntcir at clef 2010-09-21 20Noriko Kando
Multilingual Opinion Analysis (MAOT)
CO2 Reduction?CO2 Reduction?Lehman shock?Lehman shock?
ntcir at clef 2010-09-21 21Noriko Kando
Opinion Question ListID Opinion Question
N01 What negative prospects were discussed about the Euro when it was introduced in J anuary of 2002?(N03) (What reasons were discussed about Bomb Terror in Bali Island in October, 2002?)N04 What reasons have been given for the Space Shuttle Columbia accident in February, 2002?N05 What negative comments were discussed about Bush's decision to start Iraq war in March, 2003?N06 What negative prospects and opinions were discussed about SARS which started spreading in March, 2003?N07 What reasons are given for the blackout around North America in August, 2003?N08 What reasons and background information was discussed about the terrorist train bombing that happened in Madrid in March, 2004?N11 Why did supporters want to elect George W. Bush in the November 2004 American Presidential Election?
N13What positive comments were discussed to help the victims from earthquake and tsunami in Sumatera, Indonesia in December,2004?
N14 What objections are given for the US opposition to the Kyoto Protocol that was enacted in February 2005?
N16What reasons have been given for the anti- J apanese demonstrations that took place in April, 2005 in Peking and Shanghai inChina?
N17In J uly 2005 there were terrorist bombings in London. What reasons and background were given, and what controversies werediscussed?
N18 What actions by President George Bush were criticized in response to Hurricane Katrina's August 2005 landing?
N20 What negative opinions and discussion happened about the Bird Flu that started spreading in October, 2005?
N24Identify opinions that indicate that Arnold Schwarzenegger is a bad choice to be elected the new governor of California in theOctober 2003 election.
N26Find positive opinions about the reaction of Nuclear and Industrial Safety Agency officials to the Mihama nuclear powerplantaccident in August 2004.
N27 What were the advantages and disadvantages of the direct flight between Taiwan and Mainland China commercially?N32 What are good and bad approaches to losing weight?N36 What are complaints about XIX Olympic Winter Games that were held in and around Salt Lake City, Utah, United States in 2002?N39 What are the comments about China's first manned space flight which happened successfully in October 2003?
N41What negative comments were discussed when in April 2004 CBS made public pictures showing cruel U.S. military abuse of Iraqiprisoners of war?
ntcir at clef 2010-09-21 22Noriko Kando
Effective approaches
• The following teams attained good results with accomplished feature filtering, hot machine learning, and rich lexicon resources.
TeamID Lang Feature Filtering Macihne Learing Lexicon ResourceUNINE EN Z score logistic regression SentiWordNetPKUTM SC Iterative classifier SVM (better than NB, ME, DT) In House, NTU, and J un LI's lexiconCityUHK TC Supervised Lexicon Ensemble NTUSD, LCPW, LCNW, CPWP, SKPI
ntcir at clef 2010-09-21 23Noriko Kando
Community QA Pilot Task
ntcir at clef 2010-09-21 Noriko Kando24
Yahoo Chiebukuro data version 1.0:
3 million questions
Test collectionfor CQA:
1500 questions
Good Answers:
4 assessors individually assessed
Best Answer:Questioner
selected
1500 questions selected at
random
4 university students
Rank all posted answers by answer quality (as estimated by system) for every question.Training
Test
http://research.nii.ac.jp/ntcir/ntcir-ws8/yahoo/index-en.html
GA-Hit@1not useful
GA-nDCG and Qsimilar
5th run from MSRAused the BA info for
test Qs directly so does
not represent Practical
performance
ntcir at clef 2010-09-21 25Noriko Kando
LOVE is HARD according to BA!Systems can’t
find the asker’s BAs!
ntcir at clef 2010-09-21 26Noriko Kando
#questions GA-nG@1LOVE is EASY according to GA!
Systems can find many good answers that are not BA!(BA-evaluation not good
enough for social questions)
ntcir at clef 2010-09-21 27Noriko Kando
28
Goal: Automatic Creation of technical trend maps from a set of research papers and patents.
Research papers and patents are classified in terms of elemental technologies and their effects.
ntcir at clef 2010-09-21 Noriko Kando
EvaluationSubtask 1 (Research Paper Classification)Metrics: Mean Average Precision (MAP)• k-NN based approach is superior to machine learning
approach.• Re-ranking of IPC codes is effective.
Subtask 2: Technical Trend Map CreationMetrics: Recall, Precision, and F-measure• Top systems employed CRF, and the following features are
effective.– Dependency structure– Document structure– Domain adaptation
29ntcir at clef 2010-09-21 Noriko Kando
30
History of Patent IR at NTCIR
• NTCIR-3 (2001-2002)– Technology survey
• Applied conventional IR problems to patent data
• NTCIR-4 (2003-2004)– Invalidity search
• Addressed patent-specific IR problems
• NTCIR-5 (2004-2005)– Enlarged invalidity search
• NTCIR-6 (2006-2007)– Added English patents
2 years of JPO patent applications
* JPO = Japan Patent Office
5 years of JPO patent applications
10 years of JPO patent applications
10 years of USPTO patents granted
* USPTO = US Patent & Trademark Office
Both document sets were published in 1993-2002
ntcir at clef 2010-09-21 Noriko Kando
31
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
12.00 14.00 16.00 18.00 20.00 22.00 24.00 26.00
BLEU
MAP,R
ecall@
Recall@1000 (R = 0.77)Recall@500 (R = 0.84)Recall@200 (R = 0.86)Recall@100 (R = 0.86)MAP (R = 0.77)
Extrinsic E-J: BLEU & IR measures
MAP & Recall@N
Recall@100,200 are highly correlated with BLEU (R = 0.86)
ntcir at clef 2010-09-21 Noriko Kando
Overall structure
New NTCIR StructureNTCIR general chairs:Noriko Kando (NII), Eiichiro Sumita (NICT), Tsuneaki Kato (U of Tokyo)
NTCIR evaluation chairs:Hideo Joho (U of Tsukuba)Tetsuya Sakai (MSRA)
EVIA chairs:Mark Sanderson (RMIT)William Webber (U of Melbourne) + 1Grand
challenges
Core
ch
alle
ng
e
task
Core
ch
alle
ng
e
task
Pilo
t ta
sk
Pilo
t ta
sk
Task Selection Committee EVIA Program
CommitteeRefereed papers onEvaluating Information Access
Tasks accepted for NTCIR-9CORE TASKS• Intent (with One-Click Access subtask)• Recognizing Natural Language Inference• Geotemporal information retrieval (GeoTime2)• IR for Spoken Documents
PILOT TASKS• Cross-lingual Link Discovery• Evaluation Framework of Interactive Information Access• Patent Machine Translation (PATMT)
bookbook
INTENT (w 1CLICK) task (CS, J)
Organisers: Min Zhang, Yiqun Liu (Tsinghua U), Ruihua Song, Tetsuya Sakai, Youngin Song (MSRA), Makoto Kato (Kyoto U)
• Subtopic miningFind different intents given an
ambiguous/underspecified query• RankingSelectively diversity Web search results• One click access (1CLICK)Satisfy information need with the system’s first
result page (X-byte text, not document list)No need to click after clicking on SEARCH
Harry Potter
bookmovie
Recognizing Natural Language Inference (CS, CT, J)
• YES/NO taskDoes text 1 entail text 2?text 1 ⇒ text 2• Main taskGiven text 1 and text 2, choose from(1)Forward entailment (2) backward entailment(3) equivalent (4) contradict (5) independent• Extrinsic evaluation via CLIA will be
conducted
Organisers: Hideki Shima,Teruko Mitamura (CMU),Chuan-Jie Lin (NTOU),Cheng-Wei Lee (Academia Sinica)
GeoTime2 (E, J, K?)Organisers: Fred Gey, Ray Larson (UCB), Noriko Kando (NII)
• Second round of ad hoc IR for WHEN and WHERE• GeoTime1 topic: How old was Max Schmeling when he
died, and where did he die?
At GeoTime1, docs that contain both WHEN and WHERE were treated as relevant;
Those that contain only WHEN or only WHERE info were treated as partially relevant.
Can we do better?
Can we create more realistic
GeoTime topics?
IR for Spoken Documents (J)Organisers: Tomoyosi Akiba, Seiichi Nakagawa, Kiyoaki Aikawa (Toyohashi U of Technology),
Yoshiaki Itoh (Iwate Prefectural U), Tatsuya Kawahara (Kyoto U), Xinhui Hu (NICT), Hiroaki Nanjo (Ryukoku University), Hiromitsu Nishizaki (U of Yamanashi), Tomoko Matsui (Institute of Statistical Mathematics), Yoichi Yamashita (Ritsumeikan U)
• Handling spontaneous speech data (spoken lectures)• Spoken term detection
Find the position of a given query term within SD• Spoken document retrieval
Find passages in SD data that match a given query
Reference speech recognition results
provided (non-speech people can
easily participate)
Cross-lingual Link Discovery (E->C,K,J)
• Given an English Wikipedia page,
(1)Identify anchors; and
(2)For each anchor, provide a list of C/K/J documents to be linked
Solar eclipse 日食
太陽
月食
太陽
月食CorrespondingW entry
sun
Lunar eclipse
Organisers: Ling-Xiang Tang, Shlomo Geva, Andrew Trotman, Yue Xu, Darren Huang (Queesland U of Technology), Andrew Trotman (U of Otago)
(1)
(2)
Linking INEXand NTCIR!
Evaluation Framework of Interactive Information Access (J)
Organisers: Tsuneaki Kato (U Tokyo) and Mitsunori Matsushita (Kansai U)
• Explore evaluation of interactive and exploratory information access
• Shared modules, shared interfaces between modules
• Interactive complex question answering subtask (visual/textual output)
• Trend information summarization subtask (e.g. cabinet support rate change)
PATMT (C-E, J-E, E-J)Organisers: Benjamin Tsou, Bin Lu (CUHK), Isao Goto
(NICT)
• C-E MT (new at NTCIR-9), J-E MT, E-J MT
Manual and automatic evaluation (BLEU, NIST)
• PATMT@NTCIR-10 will include extrinsic evaluation by CLIR
• Lessons from PATMT@NTCIR-7,8:
BLEUMAPHuma
n rating
High Low
with multiple reference translations (RTs)
High
TALK OUTLINE• About myself• Looking Back- NTCIR- NTCIR-5,6 CLIR (as a participant)- NTCIR-7,8 ACLIA- NTCIR-8 GeoTime• Looking Forward- NTCIR Grand Challenges- NTCIR-9 Tasks• Summary
TALK SUMMARY• Looking back (1999-2010)- More teams are now using straight MT for CLIR.
Often over 90% of monolingual performance.What are CLIR researchers doing now? CLIA?• Looking forward (2010-)- Grand challenges [tentative] = NN2S + E2E =
No Need To Search + Easy To Explore.- Multilinguality will still be important especially in
the system2user direction.Please participate in NTCIR-9 tasks!
NTCIR-9 Important Dates (tentative)
1st Oct 2010 first call for participation
Winter-Autumn 2011
tasks run
November 2011
camera-ready papers due
December 2011
NTCIR-9@NII, Tokyo
Watch http://research.nii.ac.jp/ntcir/
ntcir at clef 2010-09-21 Noriko Kando 45
Thanks MerciDanke schön Gracie
Gracias Ta! Tack Köszönöm KiitosTerima Kasih Khap Khun
Ahsante Tak 謝謝 ありがとう
Thanks MerciDanke schön Gracie
Gracias Ta! Tack Köszönöm KiitosTerima Kasih Khap Khun
Ahsante Tak 謝謝 ありがとう
http://research.nii.ac.jp/ntcir/http://research.nii.ac.jp/ntcir/
Will be moved to: http://ntcir.nii.ac.jp/Will be moved to: http://ntcir.nii.ac.jp/