SlideShare a Scribd company logo

Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection

W
Weijing Huang

This document presents a method called TransDetector that facilitates category-level transfer learning from a knowledge base into microblog streams to enhance event detection accuracy. It addresses challenges in detecting events on dynamic and noisy social media platforms like Twitter, particularly for smaller events with limited tweets. The method effectively improves event detection by leveraging rich semantic information from the knowledge base while balancing performance and resource costs.

Data & Analytics
1 of 25
Download to read offline
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection Weijing Huang, Tengjiao Wang, Wei Chen, Yazhou Wang School of Electronics Engineering and Computer Science, Peking University @DASFAA 2017, Suzhou, China Weijing Huang, et,al. Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection@DASFAA 2017, Suzhou, China 1 / 25
Motivation Many Web applications need the accurate event detection technique on microblog stream, including: 1 public opinion analysis [Chen, SIGIR 2013] 2 public security [Li, ICDE 2012], [Imran, WWW 2014] 3 disaster response [Sakaki,WWW 2010] 4 breaking news report1 But detecting events on twitter stream accurately is still challenging. 1 http://www.theverge.com/2016/12/1/13804542/ reuters-algorithm-breaking-news-twitter Weijing Huang, et,al. Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection@DASFAA 2017, Suzhou, China 2 / 25
Challenges (1/2) According to [Huang, WWW 2016], the challenges include, 1 fast changing 2 high noise 3 short length And, we found another key factor, 1 Small events with fewer tweets Hard to trade off between precision and recall. Weijing Huang, et,al. Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection@DASFAA 2017, Suzhou, China 3 / 25
Challenges (2/2) Exploratory study on the Edinburgh twitter corpus: 11/27 events contain less than 50 tweets. Table: Statistics of labeled events. Event Date Event Size S&P downgrade US credit rating 05/08/2011 656 Atlantis shuttle lands 21/07/2011 595 US increases debt ceiling 25/07/2011 485 Plane with Russian hocky team Lokomotiv crashes 07/09/2011 286 Amy Winehouse dies 23/07/2011 283 Gunman opens fire in youth camp in Norway 23/07/2011 260 Earthquake in Virginia 24/08/2011 246 First victim of London riots dies 09/08/2011 174 Explosion in French nuclear plant in Marcoule 12/09/2011 135 Google announces plans to bury Motorola Mobility 15/08/2011 127 NASA announces there might be water on Mars 04/08/2011 124 Car bomb explodes in Oslo, Norway 22/07/2011 114 ... ... ... Indian and Bangladesh sign a border pact 06/09/2011 25 Flight 4896 crash 13/07/2011 21 First aritficial organ transplant 12/07/2011 18 three men die in riots in england 10/08/2011 16 rebels capture interational tripoli airport 21/08/2011 13 Weijing Huang, et,al. Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection@DASFAA 2017, Suzhou, China 4 / 25 11 Small events with fewer tweets
How about existing methods? (1/2) Event detection methods without extra information, such as 1 clustering articles LSH[Petrovic, NAACL 2010] need to set threshold to determine whether new article represents a new event. 2 analyzing word frequencies EDCoW[Weng, ICWSM 2011] treat the word as the basic unit in analysis, without regarding polysemy words (words have different meanings, e.g. “apple”) 3 finding bursty topics via topic modeling TimeUserLDA[Diao, ACL 2012], BurstyBTM[Yan, AAAI 2015] detects the “large” events but may ignore the “small” ones. Weijing Huang, et,al. Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection@DASFAA 2017, Suzhou, China 5 / 25
How about existing methods? (2/2) Event detection methods by leveraging extra information 1 typical one: Twevent[Li, CIKM 2012] divides the tweet into segments according to the Microsoft Web N-Gram service and Wikpedia detects the bursty segments and cluster these segments into candidate events still has to trade off between precision and recall e.g., the bursty segment in the not-so-popular event “first artificial organ transplant” is missed Weijing Huang, et,al. Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection@DASFAA 2017, Suzhou, China 6 / 25
An Example Much easier to detect the event on the time series of word “hood” related to Military, without adjusting the threshold. Jul 2011 Jul 2011 18 19 20 21 22 23 24 25 26 27 28 29 30 31 0 25 50 75 100 Document Frequency Ft. Hood Attack word "hood" word "hood" (about Military) Figure: The comparison of the time series between the raw word hood and the Military related word hood, computed on the Edinburgh twitter corpus. Refer the event to https://en.wikipedia.org/wiki/Fort_Hood#2011_attack_plot. An infant wearing a hood. k and he head, n from orm of hood is are used ses where the bers, (a) Fort Hood From Wikipedia, the free encyclopedia Fort Hood is a U.S. military post located in Killeen, Texas. The post is named after Confederate General John Bell Hood. It is located halfway between Austin and Waco, about 60 miles (b)Weijing Huang, et,al. Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection@DASFAA 2017, Suzhou, China 7 / 25
The insights on the example Knowledge Base Well organized Constructed elaborately full of rich information Microblog stream Short length Fast changing High noise The benefit of enriching the semantics and filtering out noise by Knowledge Base for microblogs is attractive. But it’s expensive to retrieve every word of tweets in the Knowledge Base. Weijing Huang, et,al. Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection@DASFAA 2017, Suzhou, China 8 / 25
Our solution TransDetector: a novel category-level transfer learning method transfer KB’s category-level info into microblog stream balance the performance and the cost of leveraging knowledge base for event detection knowledge base information about Military docs@2011-7-28 in the stream Texture Texture Texture army military … shootings projectiles … killed news #libya libya … hood … analysis on words “killed”, “news”, … category-level topic in microblog stream @2011-07-28, about Military category-level topic in knowledge base, about Military 2011-07-28 event: “attack at Fort Hood”, event: “#libya Libyan rebels chief killed” … 1 2 3 Figure: TransDetector’s processing flow, in 3 phases. Weijing Huang, et,al. Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection@DASFAA 2017, Suzhou, China 9 / 25
TransDetector: Phase 1 (Extracting Category-Level Topics in KB) (1/3) There is a three fold hierarchical structure in Knowledge Base. 1 Taxonomy Graph G(0). The directed edges in G(0) represent the class→subclass relations in KB. e.g., Main topic classifications → Society 2 Category-Page Bipartite Graph G(1). The directed edges in G(1) represent the class→instance relations in KB. e.g., Military → page: Armed Forces 3 Page-Content Map G(2). For a specific Wikipedia dumps version, the edges page →content in G(2) define a one-to-one mapping. e.g., page: Armed Forces → content(20170325version): “The armed forces of a country are its government-sponsored defense, fighting forces, and organizations...” Weijing Huang, et,al. Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection@DASFAA 2017, Suzhou, China 10 / 25
TransDetector: Phase 1 (Extracting Category-Level Topics in KB) (2/3) Algorithm 1: Extraction of Category-Level Topics in Knowledge Base Input: Taxonomy’s Graph G(0), Category-Page Bipartite Graph G(1), Page-Content Bipartite Graph G(2), topic related category node c Output: c’s category-level topic in knowledge base hc 1 Pages(c) ← ∅, hc ← ∅ 2 DAG G(0) ← Remove Cycles of G(0) by nodes’ HITS-PageRank scores. 3 SuccessorNodes(c) ← Breadth-first-traverse(G(0) , c) 4 for node ∈ SuccessorNodes(c) do 5 Pages(c) ← Pages(c) ∪ G(1).neighbours(node) 6 Word frequency table n(c, .) ← do word count on the text contents of Pages(c) 7 Word frequency table n(All, .) ← do word count on the text contents of all pages in G(2). 8 for word w in WordFrequencyTable(All).keys() do 9 chi(c, w) ← w’s chi-square statistics on WordFrequencyTable(c) and WordFrequencyTable(All). 10 hc,w ← chi(c, w) 11 return hc Weijing Huang, et,al. Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection@DASFAA 2017, Suzhou, China 11 / 25
TransDetector: Phase 1 (Extracting Category-Level Topics in KB) (3/3) Taking the category Military as an example, we extract Military’s category-level topic hMilitary . Main topic classifications Culture … Society Politics page: Military page: Armed Forces page: Firearm page: World War II Charity … … … Military … Activism … … World War II… category-level topic in knowledge base, about military army, military, air, command, force, regiment, forces, squadron, infantry, battle, commander, …, shootings, projectiles,… (order by words’ chi-square scores ) … G(0):Taxonomy’s Graph G(1) :Category-Page Bipartite Graph G(2): Page-Content Bipartite Graph Figure: Extracting Category-Level Topics in Knowledge Base via its three fold hierarchical structure, taking Military as an example. Weijing Huang, et,al. Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection@DASFAA 2017, Suzhou, China 12 / 25
TransDetector: Phase 2 (Transferring Category-Level Info into Microblog Stream) (1/2) Transfer KB’s Category-Level Topics {hc}KKB c=1 into microblogs stream: CTrans-LDA. u m θd zdn wdn α β ϕk τk D Nd K Figure: Diagram of CTrans-LDA. In CTrans-LDA, {hc}KKB c=1 is used as prior information: τkv =    λ hkv v∈Sk hkv , v ∈ Sk and k ≤ KKB 0, v /∈ Sk or k > KKB (1) Weijing Huang, et,al. Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection@DASFAA 2017, Suzhou, China 13 / 25
TransDetector: Phase 2 (Transferring Category-Level Info into Microblog Stream) (2/2) We use Gibbs Sampling for solving CTrans-LDA. The initialization probability ˆqk|v makes sure that the learned topics are aligned to the pre-defined category-level topic. ˆqk|v =    τkv K k=1 τkv , k τkv > 0 (a) 0, k τkv = 0 and k ≤ KKB (b) 1/(K − KKB ), k τkv = 0 and k > KKB (c) (2) Conditional probability in gibbs sampling: p(zdn = k|.) ∝ (n (d) dk + αmk)(n (w) kv + τkv + β)/(n (w) k,. + τk,. + V β). Weijing Huang, et,al. Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection@DASFAA 2017, Suzhou, China 14 / 25
TransDetector: Phase 3 (Detecting Events on Category-Level Word Time Series) (1/2) After transfer learning, we conduct analysis on category-level word time series, and detect events in microblog stream. Figure: Visualizing Category-Level Word Time Series in Microblog Stream on Edinburgh Twitter Corpus (20110711-20110915), taking Military as an example. Weijing Huang, et,al. Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection@DASFAA 2017, Suzhou, China 15 / 25
TransDetector: Phase 3 (Detecting Events on Category-Level Word Time Series) (2/2) With richer semantics, and fewer noise, we detect events more accurately, in all the following granularities.(see more technique details in our paper) 1 events’ candidate words e.g., Ft., Hood, attack. 2 event phrases e.g., Ft. Hood attack. 3 events’ representative microblogs e.g., Possible Ft. Hood Attack Thwarted http: // t. co/ BSJ33hk . Weijing Huang, et,al. Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection@DASFAA 2017, Suzhou, China 16 / 25
Experiment Settings Dataset Knowledge Base. Wikipedia dumps23 Microblog Stream. Edinburgh twitter corpus4 Baseline Methods Twevent, BurstyBTM, LSH, EDCoW, and TimeUserLDA Ground Truth Benchmark1 is labeled events in previous study. Benchmark2 is our manually checked events based on Twevent, BurstyBTM, LSH, EDCoW, TimeUserLDA and TransDetector. 2 https: //dumps.wikimedia.org/enwiki/latest/enwiki-latest-categorylinks.sql.gz 3 https: //dumps.wikimedia.org/enwiki/latest/enwiki-latest-pages-articles.xml.bz2 4 http://demeter.inf.ed.ac.uk/cross/docs/fsd_corpus.tar.gz Weijing Huang, et,al. Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection@DASFAA 2017, Suzhou, China 17 / 25
Experimental Results Evaluation on Categroy-Level Topics in Knowledge Base. On Aviation topic, semantic coherence is much better than LightLDA (same as LDA) in terms of NPMI[Roder, WSDM 2015]5. Table: The comparison on the topic coherence(NPMI) between our method and LightLDA, taking Aviation as an example. (NPMI is computed on a group of ten words. ∼ stands for the top five words.) Category-Level Topics extracted from Wikipedia by TransDetector Topics Learned from Wikipedia by LightLDA GID #words* words NPMI GID #words* words NPMI - 1-5 aircraft air airport flight airline - - 1-5 engine aircraft car air power - 0 1-5, 6-10 ∼, airlines aviation flying pilot squadron 0.113 0 1-5, 6-10 ∼, design flight model production speed 0.112 1 1-5, 11-15 ∼, flights pilots raf airways fighter 0.155 1 1-5, 11-15 ∼, system vehicle cars engines mm 0.062 2 1-5, 16-20 ∼, boeing runway force crashed flew 0.092 2 1-5, 16-20 ∼, fuel vehicles designed models type 0.072 3 1-5, 21-25 ∼, airfield landing passengers plane aerial 0.179 3 1-5, 21-25 ∼, version front produced rear electric 0.035 4 1-5, 26-30 ∼, bomber radar wing bombers crash 0.137 4 1-5, 26-30 ∼, space control motor standard development 0.085 5 1-5, 31-35 ∼, airbus airports operations jet helicopter 0.189 5 1-5, 31-35 ∼, film range light using available -0.002 6 1-5, 36-40 ∼, squadrons base flown havilland crew 0.088 6 1-5, 36-40 ∼, wing powered wheel weight launch 0.087 7 1-5, 41-45 ∼, combat luftwaffe aerodrome carrier fokker 0.159 7 1-5, 41-45 ∼, developed low test ford cylinder 0.007 8 1-5, 46-50 ∼, planes fly engine takeoff fleet 0.186 8 1-5, 46-50 ∼, equipment side pilot hp aviation 0.091 9 1-5, 51-55 ∼, fuselage helicopters aviator naval aero 0.157 9 1-5, 51-55 ∼, systems us sold body drive -0.051 10 1-5, 56-60 ∼, glider command training balloon faa 0.166 10 1-5, 56-60 ∼, gear introduced class safety seat 0.069 · · · · · · · · · · · · · · · · · · · · · · · · 18 1-5, 96-100 ∼, scheduled carriers military curtiss biplane 0.131 18 1-5, 96-100 ∼, transmission special replaced limited different 0.059 19 1-5, 101-105 ∼, accident engines iaf albatross rcaf 0.068 19 1-5, 101-105 ∼, features machine nuclear even unit 0.011 5 https://github.com/AKSW/Palmetto Weijing Huang, et,al. Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection@DASFAA 2017, Suzhou, China 18 / 25
Experimental Results Evaluation on Categroy-Level Topics in Knowledge Base. On more topics, semantic coherence is much better than LightLDA (same as LDA) in terms of NPMI[Roder, WSDM 2015]. −0.1 0.0 0.1 0.2 0.3 0 10 20 30 40 50 Group ID NPMI Category−Level Topics in Knowledge Base Topics learned by LightLDA Figure: More topics are compared at the NPMI metrics between our method and LightLDA Weijing Huang, et,al. Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection@DASFAA 2017, Suzhou, China 19 / 25
Experimental Results Effectiveness of transferring category-level topics into the microblog stream, and finding more new words in the stream which is not stored in the knowledge base. Table: Category-Level Topics extracted from knowledge base and the corresponding topics on microblog stream learned from CTrans-LDA. The words in bold font are newly learned on the microblog stream by the transfer learning. Aviation Health Middle East Military Mobile Phones Knowledge Base Microblog Stream Knowledge Base Microblog Stream Knowledge Base Microblog Stream Knowledge Base Microblog Stream Knowledge Base Microblog Stream aircraft air health weight al #syria army killed android iphone air plane patients loss israel #bahrain military news mobile apple airport flight medical diet iran people air #libya nokia android flight time disease health arab israel command libya ios app airline airlines treatment cancer israeli police force rebels phone ipad airlines news hospital lose egypt #libya regiment people samsung samsung aviation boat patient fat egyptian #egypt forces police game mobile flying airport clinical tips ibn news squadron war app blackberry pilot force symptoms treatment jerusalem #israel infantry libyan iphone tablet squadron fly cancer body syria world battle attack htc apps Weijing Huang, et,al. Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection@DASFAA 2017, Suzhou, China 20 / 25
Experimental Results Effectiveness of event detection: 1 TransDetector performs better in terms of the precision and the recall. 2 only sacrificing in the DERate slightly because an event could be grouped into multiple categories. the event “S&P downgrade US credit rating”, related to politics and financial simultaneously. Table: Overall Performance on Event Detection Method Number of Events to be Evaluated Recall@ Benchmark1 Precision@ Benchmark2 Recall@ Benchmark2 F@ Benchmark2 DERatea (Duplicate Event Rate)@ Benchmark2 LSH 500 0.704 0.788 0.651 0.713 0.348 TimeUserLDA 100 0.370 0.790 0.177 0.289 0.114 Twevent 375 0.741 0.808 0.658 0.725 0.142 EDCoW 349 0.556 0.748 0.511 0.607 0.226 BurstyBTM 200 0.667 0.825 0.384 0.497 0.079 TransDetector 457 0.889 0.912 0.876 0.894 0.170 a DERate = (the number of duplicate events) / (the total number of detected realistic events) Weijing Huang, et,al. Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection@DASFAA 2017, Suzhou, China 21 / 25
Experimental Results To understand why TransDetector performs better. Show the relationship between the recall and the event size. 0.00 0.25 0.50 0.75 1.00 >200 (100,200] (50,100] [0,50] Group with different event size (number of tweets) Recall@Benchmark1 TimeUserLDA BurstyBTM EDCoW LSH Twevent TransDetector Figure: The relation between the recall and the event size Weijing Huang, et,al. Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection@DASFAA 2017, Suzhou, China 22 / 25
Experimental Results To understand why TransDetector performs better. Show the relationship between the recall and the event size, taking the military-related events as an example. Table: Events about military detected by systems between 2011-07-22 and 2011-07-28 Date Event key words Representative event tweet Number of event tweet Methodsa L TU TW E B TD 7/22/11 Norway, Oslo, attacks, bombing Terror Attacks Devastate Norway: A bomb ripped through government offices in Oslo and a gunman... http://dlvr.it/cLbk8 557 7/23/11 Gunman, rink Gunman Kills Self, 5 Others at Texas Roller Rink http://dlvr.it/cLcTH 43 - - - 7/26/11 Kandahar, mayor, suicide, attack TELEGRAPH]: Kandahar mayor killed by Afghan suicide bomber: The mayor of Kandahar, the biggest city in south 47 - - 7/28/11 Ft., Hood, attack Possible Ft. Hood Attack Thwarted http://t.co/BSJ33hk 52 - - - - - 7/28/11 Libyan, rebel, gunned Libyan rebel chief gunned down in Benghazi http://sns.mx/prfvy1 44 - - - - - a L=LSH, TU=TimeUserLDA, TW=Twevent, E=EDCoW, B=BurstyBTM, TD=TransDetector. Weijing Huang, et,al. Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection@DASFAA 2017, Suzhou, China 23 / 25
Conclusions Knowledge base is constructed elaborately and contains rich information, which can benefit the not-well-organized microblog stream. We propose TransDetector method, and use category-level topic in knowledge base as the prior knowledge, transfer abundant knowledge from knowledge base into microblog stream enrich the semantics of microblogs and further enhances the accuracy of microblogs event detection Weijing Huang, et,al. Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection@DASFAA 2017, Suzhou, China 24 / 25
Thanks! Q&A 5 This slide and more data are available at http://q-r.to/bajx8I Weijing Huang, et,al. Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection@DASFAA 2017, Suzhou, China 25 / 25

More Related Content

PDF
Stretching the Life of Twitter Classifiers with Time-Stamped Semantic Graphs
Amparo Elizabeth Cano Basave
 
PPTX
Kdd 2014 tutorial bringing structure to text - chi
Barbara Starr
 
DOCX
(Christopher)With packet-switched networks its services allow mult
SilvaGraf83
 
PPTX
Generating Storylines (Literature Survey)
Anunaya
 
PDF
eventdemo2016
Rachel Guan
 
PDF
On the ethical challenges in the use of AI/ML for research in science and eng...
Xi'an Jiaotong-Liverpool University
 
PDF
CV
Xiang Lian
 
PDF
A history of clu
sugeladi
 
Stretching the Life of Twitter Classifiers with Time-Stamped Semantic Graphs
Amparo Elizabeth Cano Basave
 
Kdd 2014 tutorial bringing structure to text - chi
Barbara Starr
 
(Christopher)With packet-switched networks its services allow mult
SilvaGraf83
 
Generating Storylines (Literature Survey)
Anunaya
 
eventdemo2016
Rachel Guan
 
On the ethical challenges in the use of AI/ML for research in science and eng...
Xi'an Jiaotong-Liverpool University
 
CV
Xiang Lian
 
A history of clu
sugeladi
 

Similar to Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection (20)

PPT
AI Science
Melanie Swan
 
PPTX
Data and science
Anand Deshpande
 
PDF
Pre-defense_talk
aphex34
 
PDF
A&S Chairs Hybrid Courses Project Presentation
Christopher Rice
 
PDF
Kid171 chap0 english version
Frank S.C. Tseng
 
PDF
Evaluation Initiatives for Entity-oriented Search
krisztianbalog
 
PDF
Knowledge Graph Futures
Paul Groth
 
PDF
파이콘 한국 2019 튜토리얼 - 설명가능인공지능이란? (Part 1)
XAIC
 
PPTX
Semantic Web in Physical Science
petermurrayrust
 
PPTX
Processing Events in Probabilistic Risk Assessment
Haystax Technology
 
PPT
Software Design Patterns Lecutre (Intro)
VikramJothyPrakash1
 
PDF
Hala skafkeynote@conferencedata2021
hala Skaf
 
PPT
Semantic Wiki Based Collaborative Scientific Modeling Infrastructure
Jie Bao
 
PPTX
From data portal to knowledge portal: Leveraging semantic technologies to sup...
Xiaogang (Marshall) Ma
 
PDF
Mining Social Media with Linked Open Data, Entity Recognition, and Event Extr...
Leon Derczynski
 
PPTX
Contextual Ontology Alignment - ESWC 2011
Mariana Damova, Ph.D
 
PDF
Deep Context-Awareness: Context Coupling and New Types of Context Information...
Hong-Linh Truong
 
DOCX
College of Doctoral StudiesRES-845 Module 2 Problem.docx
adkinspaige22
 
DOCX
College of Doctoral StudiesRES-845 Module 2 Problem.docx
ShiraPrater50
 
PPTX
Interactive Visualization Systems and Data Integration Methods for Supporting...
Don Pellegrino
 
AI Science
Melanie Swan
 
Data and science
Anand Deshpande
 
Pre-defense_talk
aphex34
 
A&S Chairs Hybrid Courses Project Presentation
Christopher Rice
 
Kid171 chap0 english version
Frank S.C. Tseng
 
Evaluation Initiatives for Entity-oriented Search
krisztianbalog
 
Knowledge Graph Futures
Paul Groth
 
파이콘 한국 2019 튜토리얼 - 설명가능인공지능이란? (Part 1)
XAIC
 
Semantic Web in Physical Science
petermurrayrust
 
Processing Events in Probabilistic Risk Assessment
Haystax Technology
 
Software Design Patterns Lecutre (Intro)
VikramJothyPrakash1
 
Hala skafkeynote@conferencedata2021
hala Skaf
 
Semantic Wiki Based Collaborative Scientific Modeling Infrastructure
Jie Bao
 
From data portal to knowledge portal: Leveraging semantic technologies to sup...
Xiaogang (Marshall) Ma
 
Mining Social Media with Linked Open Data, Entity Recognition, and Event Extr...
Leon Derczynski
 
Contextual Ontology Alignment - ESWC 2011
Mariana Damova, Ph.D
 
Deep Context-Awareness: Context Coupling and New Types of Context Information...
Hong-Linh Truong
 
College of Doctoral StudiesRES-845 Module 2 Problem.docx
adkinspaige22
 
College of Doctoral StudiesRES-845 Module 2 Problem.docx
ShiraPrater50
 
Interactive Visualization Systems and Data Integration Methods for Supporting...
Don Pellegrino
 
Ad

Recently uploaded (20)

PPTX
AI Strategy room jwfjksfksfjsjsjsjsjfsjfsj
ankitpratapsingh28
 
PPTX
iec ppt-1 pptx icmr ppt on rehabilitation.pptx
thirishadevan81
 
PDF
Business Analytics and business intelligence.pdf
khalidisah4750
 
PPTX
mbdjdhjjodule 5-1 rhfhhfjtjjhafbrhfnfbbfnb
CHINTU5000
 
PDF
Fluorescence-microscope_Botany_detailed content
dollydoll12
 
PPTX
Introduction to Knowledge Engineering Part 1
busyprogrammersguide
 
PPTX
01_intro xxxxxxxxxxfffffffffffaaaaaaaaaaafg
Vittaya Boon Suwansiri
 
PDF
Mega Projects Data Mega Projects Data
saidsalah8181
 
PPTX
MODULE 8 - DISASTER risk PREPAREDNESS.pptx
AceAquino4
 
PPTX
Introduction-to-Cloud-ComputingFinal.pptx
testnet29393883
 
PPT
Reliability_Chapter_ presentation 1221.5784
abobaker13
 
PPTX
Microsoft-Fabric-Unifying-Analytics-for-the-Modern-Enterprise Solution.pptx
Mahesh Reddy
 
PDF
Lecture1 pattern recognition............
ZafriFarid
 
PPTX
Acceptance and paychological effects of mandatory extra coach I classes.pptx
ZaheerAhmad228692
 
PPTX
Database Infoormation System (DBIS).pptx
TefferiMekonnen2
 
PPT
Miokarditis (Inflamasi pada Otot Jantung)
NandaAndini1
 
PPTX
Market Analysis -202507- Wind-Solar+Hybrid+Street+Lights+for+the+North+Amer...
LEDLUXX Smart Lite TEC
 
PPTX
STUDY DESIGN details- Lt Col Maksud (21).pptx
AyeshaAsha11
 
PDF
[EN] Industrial Machine Downtime Prediction
Mônica N. de Resende
 
PDF
.pdf is not working space design for the following data for the following dat...
jerinjoy242
 
AI Strategy room jwfjksfksfjsjsjsjsjfsjfsj
ankitpratapsingh28
 
iec ppt-1 pptx icmr ppt on rehabilitation.pptx
thirishadevan81
 
Business Analytics and business intelligence.pdf
khalidisah4750
 
mbdjdhjjodule 5-1 rhfhhfjtjjhafbrhfnfbbfnb
CHINTU5000
 
Fluorescence-microscope_Botany_detailed content
dollydoll12
 
Introduction to Knowledge Engineering Part 1
busyprogrammersguide
 
01_intro xxxxxxxxxxfffffffffffaaaaaaaaaaafg
Vittaya Boon Suwansiri
 
Mega Projects Data Mega Projects Data
saidsalah8181
 
MODULE 8 - DISASTER risk PREPAREDNESS.pptx
AceAquino4
 
Introduction-to-Cloud-ComputingFinal.pptx
testnet29393883
 
Reliability_Chapter_ presentation 1221.5784
abobaker13
 
Microsoft-Fabric-Unifying-Analytics-for-the-Modern-Enterprise Solution.pptx
Mahesh Reddy
 
Lecture1 pattern recognition............
ZafriFarid
 
Acceptance and paychological effects of mandatory extra coach I classes.pptx
ZaheerAhmad228692
 
Database Infoormation System (DBIS).pptx
TefferiMekonnen2
 
Miokarditis (Inflamasi pada Otot Jantung)
NandaAndini1
 
Market Analysis -202507- Wind-Solar+Hybrid+Street+Lights+for+the+North+Amer...
LEDLUXX Smart Lite TEC
 
STUDY DESIGN details- Lt Col Maksud (21).pptx
AyeshaAsha11
 
[EN] Industrial Machine Downtime Prediction
Mônica N. de Resende
 
.pdf is not working space design for the following data for the following dat...
jerinjoy242
 
Ad

Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection

  • 1. Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection Weijing Huang, Tengjiao Wang, Wei Chen, Yazhou Wang School of Electronics Engineering and Computer Science, Peking University @DASFAA 2017, Suzhou, China Weijing Huang, et,al. Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection@DASFAA 2017, Suzhou, China 1 / 25
  • 2. Motivation Many Web applications need the accurate event detection technique on microblog stream, including: 1 public opinion analysis [Chen, SIGIR 2013] 2 public security [Li, ICDE 2012], [Imran, WWW 2014] 3 disaster response [Sakaki,WWW 2010] 4 breaking news report1 But detecting events on twitter stream accurately is still challenging. 1 http://www.theverge.com/2016/12/1/13804542/ reuters-algorithm-breaking-news-twitter Weijing Huang, et,al. Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection@DASFAA 2017, Suzhou, China 2 / 25
  • 3. Challenges (1/2) According to [Huang, WWW 2016], the challenges include, 1 fast changing 2 high noise 3 short length And, we found another key factor, 1 Small events with fewer tweets Hard to trade off between precision and recall. Weijing Huang, et,al. Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection@DASFAA 2017, Suzhou, China 3 / 25
  • 4. Challenges (2/2) Exploratory study on the Edinburgh twitter corpus: 11/27 events contain less than 50 tweets. Table: Statistics of labeled events. Event Date Event Size S&P downgrade US credit rating 05/08/2011 656 Atlantis shuttle lands 21/07/2011 595 US increases debt ceiling 25/07/2011 485 Plane with Russian hocky team Lokomotiv crashes 07/09/2011 286 Amy Winehouse dies 23/07/2011 283 Gunman opens fire in youth camp in Norway 23/07/2011 260 Earthquake in Virginia 24/08/2011 246 First victim of London riots dies 09/08/2011 174 Explosion in French nuclear plant in Marcoule 12/09/2011 135 Google announces plans to bury Motorola Mobility 15/08/2011 127 NASA announces there might be water on Mars 04/08/2011 124 Car bomb explodes in Oslo, Norway 22/07/2011 114 ... ... ... Indian and Bangladesh sign a border pact 06/09/2011 25 Flight 4896 crash 13/07/2011 21 First aritficial organ transplant 12/07/2011 18 three men die in riots in england 10/08/2011 16 rebels capture interational tripoli airport 21/08/2011 13 Weijing Huang, et,al. Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection@DASFAA 2017, Suzhou, China 4 / 25 11 Small events with fewer tweets
  • 5. How about existing methods? (1/2) Event detection methods without extra information, such as 1 clustering articles LSH[Petrovic, NAACL 2010] need to set threshold to determine whether new article represents a new event. 2 analyzing word frequencies EDCoW[Weng, ICWSM 2011] treat the word as the basic unit in analysis, without regarding polysemy words (words have different meanings, e.g. “apple”) 3 finding bursty topics via topic modeling TimeUserLDA[Diao, ACL 2012], BurstyBTM[Yan, AAAI 2015] detects the “large” events but may ignore the “small” ones. Weijing Huang, et,al. Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection@DASFAA 2017, Suzhou, China 5 / 25
  • 6. How about existing methods? (2/2) Event detection methods by leveraging extra information 1 typical one: Twevent[Li, CIKM 2012] divides the tweet into segments according to the Microsoft Web N-Gram service and Wikpedia detects the bursty segments and cluster these segments into candidate events still has to trade off between precision and recall e.g., the bursty segment in the not-so-popular event “first artificial organ transplant” is missed Weijing Huang, et,al. Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection@DASFAA 2017, Suzhou, China 6 / 25
  • 7. An Example Much easier to detect the event on the time series of word “hood” related to Military, without adjusting the threshold. Jul 2011 Jul 2011 18 19 20 21 22 23 24 25 26 27 28 29 30 31 0 25 50 75 100 Document Frequency Ft. Hood Attack word "hood" word "hood" (about Military) Figure: The comparison of the time series between the raw word hood and the Military related word hood, computed on the Edinburgh twitter corpus. Refer the event to https://en.wikipedia.org/wiki/Fort_Hood#2011_attack_plot. An infant wearing a hood. k and he head, n from orm of hood is are used ses where the bers, (a) Fort Hood From Wikipedia, the free encyclopedia Fort Hood is a U.S. military post located in Killeen, Texas. The post is named after Confederate General John Bell Hood. It is located halfway between Austin and Waco, about 60 miles (b)Weijing Huang, et,al. Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection@DASFAA 2017, Suzhou, China 7 / 25
  • 8. The insights on the example Knowledge Base Well organized Constructed elaborately full of rich information Microblog stream Short length Fast changing High noise The benefit of enriching the semantics and filtering out noise by Knowledge Base for microblogs is attractive. But it’s expensive to retrieve every word of tweets in the Knowledge Base. Weijing Huang, et,al. Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection@DASFAA 2017, Suzhou, China 8 / 25
  • 9. Our solution TransDetector: a novel category-level transfer learning method transfer KB’s category-level info into microblog stream balance the performance and the cost of leveraging knowledge base for event detection knowledge base information about Military docs@2011-7-28 in the stream Texture Texture Texture army military … shootings projectiles … killed news #libya libya … hood … analysis on words “killed”, “news”, … category-level topic in microblog stream @2011-07-28, about Military category-level topic in knowledge base, about Military 2011-07-28 event: “attack at Fort Hood”, event: “#libya Libyan rebels chief killed” … 1 2 3 Figure: TransDetector’s processing flow, in 3 phases. Weijing Huang, et,al. Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection@DASFAA 2017, Suzhou, China 9 / 25
  • 10. TransDetector: Phase 1 (Extracting Category-Level Topics in KB) (1/3) There is a three fold hierarchical structure in Knowledge Base. 1 Taxonomy Graph G(0). The directed edges in G(0) represent the class→subclass relations in KB. e.g., Main topic classifications → Society 2 Category-Page Bipartite Graph G(1). The directed edges in G(1) represent the class→instance relations in KB. e.g., Military → page: Armed Forces 3 Page-Content Map G(2). For a specific Wikipedia dumps version, the edges page →content in G(2) define a one-to-one mapping. e.g., page: Armed Forces → content(20170325version): “The armed forces of a country are its government-sponsored defense, fighting forces, and organizations...” Weijing Huang, et,al. Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection@DASFAA 2017, Suzhou, China 10 / 25
  • 11. TransDetector: Phase 1 (Extracting Category-Level Topics in KB) (2/3) Algorithm 1: Extraction of Category-Level Topics in Knowledge Base Input: Taxonomy’s Graph G(0), Category-Page Bipartite Graph G(1), Page-Content Bipartite Graph G(2), topic related category node c Output: c’s category-level topic in knowledge base hc 1 Pages(c) ← ∅, hc ← ∅ 2 DAG G(0) ← Remove Cycles of G(0) by nodes’ HITS-PageRank scores. 3 SuccessorNodes(c) ← Breadth-first-traverse(G(0) , c) 4 for node ∈ SuccessorNodes(c) do 5 Pages(c) ← Pages(c) ∪ G(1).neighbours(node) 6 Word frequency table n(c, .) ← do word count on the text contents of Pages(c) 7 Word frequency table n(All, .) ← do word count on the text contents of all pages in G(2). 8 for word w in WordFrequencyTable(All).keys() do 9 chi(c, w) ← w’s chi-square statistics on WordFrequencyTable(c) and WordFrequencyTable(All). 10 hc,w ← chi(c, w) 11 return hc Weijing Huang, et,al. Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection@DASFAA 2017, Suzhou, China 11 / 25
  • 12. TransDetector: Phase 1 (Extracting Category-Level Topics in KB) (3/3) Taking the category Military as an example, we extract Military’s category-level topic hMilitary . Main topic classifications Culture … Society Politics page: Military page: Armed Forces page: Firearm page: World War II Charity … … … Military … Activism … … World War II… category-level topic in knowledge base, about military army, military, air, command, force, regiment, forces, squadron, infantry, battle, commander, …, shootings, projectiles,… (order by words’ chi-square scores ) … G(0):Taxonomy’s Graph G(1) :Category-Page Bipartite Graph G(2): Page-Content Bipartite Graph Figure: Extracting Category-Level Topics in Knowledge Base via its three fold hierarchical structure, taking Military as an example. Weijing Huang, et,al. Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection@DASFAA 2017, Suzhou, China 12 / 25
  • 13. TransDetector: Phase 2 (Transferring Category-Level Info into Microblog Stream) (1/2) Transfer KB’s Category-Level Topics {hc}KKB c=1 into microblogs stream: CTrans-LDA. u m θd zdn wdn α β ϕk τk D Nd K Figure: Diagram of CTrans-LDA. In CTrans-LDA, {hc}KKB c=1 is used as prior information: τkv =    λ hkv v∈Sk hkv , v ∈ Sk and k ≤ KKB 0, v /∈ Sk or k > KKB (1) Weijing Huang, et,al. Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection@DASFAA 2017, Suzhou, China 13 / 25
  • 14. TransDetector: Phase 2 (Transferring Category-Level Info into Microblog Stream) (2/2) We use Gibbs Sampling for solving CTrans-LDA. The initialization probability ˆqk|v makes sure that the learned topics are aligned to the pre-defined category-level topic. ˆqk|v =    τkv K k=1 τkv , k τkv > 0 (a) 0, k τkv = 0 and k ≤ KKB (b) 1/(K − KKB ), k τkv = 0 and k > KKB (c) (2) Conditional probability in gibbs sampling: p(zdn = k|.) ∝ (n (d) dk + αmk)(n (w) kv + τkv + β)/(n (w) k,. + τk,. + V β). Weijing Huang, et,al. Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection@DASFAA 2017, Suzhou, China 14 / 25
  • 15. TransDetector: Phase 3 (Detecting Events on Category-Level Word Time Series) (1/2) After transfer learning, we conduct analysis on category-level word time series, and detect events in microblog stream. Figure: Visualizing Category-Level Word Time Series in Microblog Stream on Edinburgh Twitter Corpus (20110711-20110915), taking Military as an example. Weijing Huang, et,al. Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection@DASFAA 2017, Suzhou, China 15 / 25
  • 16. TransDetector: Phase 3 (Detecting Events on Category-Level Word Time Series) (2/2) With richer semantics, and fewer noise, we detect events more accurately, in all the following granularities.(see more technique details in our paper) 1 events’ candidate words e.g., Ft., Hood, attack. 2 event phrases e.g., Ft. Hood attack. 3 events’ representative microblogs e.g., Possible Ft. Hood Attack Thwarted http: // t. co/ BSJ33hk . Weijing Huang, et,al. Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection@DASFAA 2017, Suzhou, China 16 / 25
  • 17. Experiment Settings Dataset Knowledge Base. Wikipedia dumps23 Microblog Stream. Edinburgh twitter corpus4 Baseline Methods Twevent, BurstyBTM, LSH, EDCoW, and TimeUserLDA Ground Truth Benchmark1 is labeled events in previous study. Benchmark2 is our manually checked events based on Twevent, BurstyBTM, LSH, EDCoW, TimeUserLDA and TransDetector. 2 https: //dumps.wikimedia.org/enwiki/latest/enwiki-latest-categorylinks.sql.gz 3 https: //dumps.wikimedia.org/enwiki/latest/enwiki-latest-pages-articles.xml.bz2 4 http://demeter.inf.ed.ac.uk/cross/docs/fsd_corpus.tar.gz Weijing Huang, et,al. Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection@DASFAA 2017, Suzhou, China 17 / 25
  • 18. Experimental Results Evaluation on Categroy-Level Topics in Knowledge Base. On Aviation topic, semantic coherence is much better than LightLDA (same as LDA) in terms of NPMI[Roder, WSDM 2015]5. Table: The comparison on the topic coherence(NPMI) between our method and LightLDA, taking Aviation as an example. (NPMI is computed on a group of ten words. ∼ stands for the top five words.) Category-Level Topics extracted from Wikipedia by TransDetector Topics Learned from Wikipedia by LightLDA GID #words* words NPMI GID #words* words NPMI - 1-5 aircraft air airport flight airline - - 1-5 engine aircraft car air power - 0 1-5, 6-10 ∼, airlines aviation flying pilot squadron 0.113 0 1-5, 6-10 ∼, design flight model production speed 0.112 1 1-5, 11-15 ∼, flights pilots raf airways fighter 0.155 1 1-5, 11-15 ∼, system vehicle cars engines mm 0.062 2 1-5, 16-20 ∼, boeing runway force crashed flew 0.092 2 1-5, 16-20 ∼, fuel vehicles designed models type 0.072 3 1-5, 21-25 ∼, airfield landing passengers plane aerial 0.179 3 1-5, 21-25 ∼, version front produced rear electric 0.035 4 1-5, 26-30 ∼, bomber radar wing bombers crash 0.137 4 1-5, 26-30 ∼, space control motor standard development 0.085 5 1-5, 31-35 ∼, airbus airports operations jet helicopter 0.189 5 1-5, 31-35 ∼, film range light using available -0.002 6 1-5, 36-40 ∼, squadrons base flown havilland crew 0.088 6 1-5, 36-40 ∼, wing powered wheel weight launch 0.087 7 1-5, 41-45 ∼, combat luftwaffe aerodrome carrier fokker 0.159 7 1-5, 41-45 ∼, developed low test ford cylinder 0.007 8 1-5, 46-50 ∼, planes fly engine takeoff fleet 0.186 8 1-5, 46-50 ∼, equipment side pilot hp aviation 0.091 9 1-5, 51-55 ∼, fuselage helicopters aviator naval aero 0.157 9 1-5, 51-55 ∼, systems us sold body drive -0.051 10 1-5, 56-60 ∼, glider command training balloon faa 0.166 10 1-5, 56-60 ∼, gear introduced class safety seat 0.069 · · · · · · · · · · · · · · · · · · · · · · · · 18 1-5, 96-100 ∼, scheduled carriers military curtiss biplane 0.131 18 1-5, 96-100 ∼, transmission special replaced limited different 0.059 19 1-5, 101-105 ∼, accident engines iaf albatross rcaf 0.068 19 1-5, 101-105 ∼, features machine nuclear even unit 0.011 5 https://github.com/AKSW/Palmetto Weijing Huang, et,al. Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection@DASFAA 2017, Suzhou, China 18 / 25
  • 19. Experimental Results Evaluation on Categroy-Level Topics in Knowledge Base. On more topics, semantic coherence is much better than LightLDA (same as LDA) in terms of NPMI[Roder, WSDM 2015]. −0.1 0.0 0.1 0.2 0.3 0 10 20 30 40 50 Group ID NPMI Category−Level Topics in Knowledge Base Topics learned by LightLDA Figure: More topics are compared at the NPMI metrics between our method and LightLDA Weijing Huang, et,al. Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection@DASFAA 2017, Suzhou, China 19 / 25
  • 20. Experimental Results Effectiveness of transferring category-level topics into the microblog stream, and finding more new words in the stream which is not stored in the knowledge base. Table: Category-Level Topics extracted from knowledge base and the corresponding topics on microblog stream learned from CTrans-LDA. The words in bold font are newly learned on the microblog stream by the transfer learning. Aviation Health Middle East Military Mobile Phones Knowledge Base Microblog Stream Knowledge Base Microblog Stream Knowledge Base Microblog Stream Knowledge Base Microblog Stream Knowledge Base Microblog Stream aircraft air health weight al #syria army killed android iphone air plane patients loss israel #bahrain military news mobile apple airport flight medical diet iran people air #libya nokia android flight time disease health arab israel command libya ios app airline airlines treatment cancer israeli police force rebels phone ipad airlines news hospital lose egypt #libya regiment people samsung samsung aviation boat patient fat egyptian #egypt forces police game mobile flying airport clinical tips ibn news squadron war app blackberry pilot force symptoms treatment jerusalem #israel infantry libyan iphone tablet squadron fly cancer body syria world battle attack htc apps Weijing Huang, et,al. Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection@DASFAA 2017, Suzhou, China 20 / 25
  • 21. Experimental Results Effectiveness of event detection: 1 TransDetector performs better in terms of the precision and the recall. 2 only sacrificing in the DERate slightly because an event could be grouped into multiple categories. the event “S&P downgrade US credit rating”, related to politics and financial simultaneously. Table: Overall Performance on Event Detection Method Number of Events to be Evaluated Recall@ Benchmark1 Precision@ Benchmark2 Recall@ Benchmark2 F@ Benchmark2 DERatea (Duplicate Event Rate)@ Benchmark2 LSH 500 0.704 0.788 0.651 0.713 0.348 TimeUserLDA 100 0.370 0.790 0.177 0.289 0.114 Twevent 375 0.741 0.808 0.658 0.725 0.142 EDCoW 349 0.556 0.748 0.511 0.607 0.226 BurstyBTM 200 0.667 0.825 0.384 0.497 0.079 TransDetector 457 0.889 0.912 0.876 0.894 0.170 a DERate = (the number of duplicate events) / (the total number of detected realistic events) Weijing Huang, et,al. Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection@DASFAA 2017, Suzhou, China 21 / 25
  • 22. Experimental Results To understand why TransDetector performs better. Show the relationship between the recall and the event size. 0.00 0.25 0.50 0.75 1.00 >200 (100,200] (50,100] [0,50] Group with different event size (number of tweets) Recall@Benchmark1 TimeUserLDA BurstyBTM EDCoW LSH Twevent TransDetector Figure: The relation between the recall and the event size Weijing Huang, et,al. Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection@DASFAA 2017, Suzhou, China 22 / 25
  • 23. Experimental Results To understand why TransDetector performs better. Show the relationship between the recall and the event size, taking the military-related events as an example. Table: Events about military detected by systems between 2011-07-22 and 2011-07-28 Date Event key words Representative event tweet Number of event tweet Methodsa L TU TW E B TD 7/22/11 Norway, Oslo, attacks, bombing Terror Attacks Devastate Norway: A bomb ripped through government offices in Oslo and a gunman... http://dlvr.it/cLbk8 557 7/23/11 Gunman, rink Gunman Kills Self, 5 Others at Texas Roller Rink http://dlvr.it/cLcTH 43 - - - 7/26/11 Kandahar, mayor, suicide, attack TELEGRAPH]: Kandahar mayor killed by Afghan suicide bomber: The mayor of Kandahar, the biggest city in south 47 - - 7/28/11 Ft., Hood, attack Possible Ft. Hood Attack Thwarted http://t.co/BSJ33hk 52 - - - - - 7/28/11 Libyan, rebel, gunned Libyan rebel chief gunned down in Benghazi http://sns.mx/prfvy1 44 - - - - - a L=LSH, TU=TimeUserLDA, TW=Twevent, E=EDCoW, B=BurstyBTM, TD=TransDetector. Weijing Huang, et,al. Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection@DASFAA 2017, Suzhou, China 23 / 25
  • 24. Conclusions Knowledge base is constructed elaborately and contains rich information, which can benefit the not-well-organized microblog stream. We propose TransDetector method, and use category-level topic in knowledge base as the prior knowledge, transfer abundant knowledge from knowledge base into microblog stream enrich the semantics of microblogs and further enhances the accuracy of microblogs event detection Weijing Huang, et,al. Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection@DASFAA 2017, Suzhou, China 24 / 25
  • 25. Thanks! Q&A 5 This slide and more data are available at http://q-r.to/bajx8I Weijing Huang, et,al. Category-Level Transfer Learning from Knowledge Base to Microblog Stream for Accurate Event Detection@DASFAA 2017, Suzhou, China 25 / 25