Big$DataVersusthe$Crowd · *Mike#...

1
Big Data Versus the Crowd Looking for Rela8onships in All the Right Places Ce Zhang, Feng Niu, Christopher Ré, and Jude Shavlik Execu8ve Summary Big Data vs. the Crowd Methodology: Follow Stateoftheart DS Scheme Feature Extrac8on Distant Supervision To Combat Inaccurate Labels in DS How does increasing the corpus size impact the quality of DS? How does increasing the amount of human feedback impact the quality of DS? Hypothesis: larger the corpus, bePer the quality Men8on Obama Barack Obama B. Obama Men8on1 Men8on2 Feature B. Obama Michelle Men8on En8ty Obama Barack Obama B. Obama Men8on extrac8on Linguis8c paPern extrac8on En8ty Linking Person, LocaDon, OrganizaDon …. Use Stanford CoreNLP Link menDons to Freebase using exact string matching Dependency path between menDons. Word sequence between menDons. PERSON PERSON marry nsubj dobj Distant Supervision (DS) Given Freebase as a knowledge base find menDon pairs supporDng Freebase. En8ty En8ty Human Feedback (HF) Annotate menDon pairs generated by distant supervision with Y/N: 3 turkers for each pair with quality control Hypothesis: more human feedback, bePer the quality Distant Supervision HasSpouse Lilly LedbeZer joined Obama and his wife, Michelle ,… Senator Barack Obama and Michelle Obama En8ty En8ty Use broad coverage and redundancy in the large corpus. Ask humans in the crowd to provide feedbacks. * Mike Mintz et al.. Distant supervision for relaDon extracDon without labeled data. In ACL 2009. Takeaways When developing a DS system, one should first expand the training Corpus in order to improve recall and then worry about the precision of training examples. Mo8va8ng Applica8on: DeepDive DeepDive approaches relaDon extracDon using human analysts and staDsDcal signals from terabytes of data. It applies distant supervision to generate training examples from unlabeled text corpus. DeepDive hZp://research.cs.wisc.edu/hazy/deepdive/ Barack Obama brought his wife Michelle to Kenya three years later… Text Subject Spouse Barack Obama Michelle Obama George W. Bush Laura Welch Bill Clinton Hillary Rodham George H. W. Bush Barbara Pierce HasSpouse Facts Contribu8on Empirically study the factors that contribute to distant supervision quality and their relaDve impact. Future Work Study more sophis8cated models (than LR) and distant supervision scheme. Explore other (more effec8ve) usage of human feedbacks. Experiments Experiment Setup Machine Learning Model Train a logisDc regression model using training data obtained from DS and HF. To Study our Hypotheses… Subsample the document corpus to get different DS training set. Subsample the human annotaDons to get different HF training set. Data sets (another data set reported in paper) TACKBP: 1.8M news arDcles ClueWeb: 500M Web pages 1000 10000 100000 1000000 1800000 Corpus Size Human Feedback Amount F1 Impact of Big Data The larger the corpus size is, the higher the quality we can expect. F1 is a loglinear func8on in the corpus size that we used for distant supervision. Large corpus increases the coverage of linguis8c paPerns in DS. 0 0.1 0.2 0.3 0.4 0.5 Precision 0 0.1 0.2 Recall Corpus Size TACKBP Human Feedback Amount Impact of the Crowd Human feedback has comparable precision as DS in our current protocol. 0 0.1 0.2 0.3 0.4 0.5 Precision 0 0.1 0.2 Recall But the slope is smaller than increasing the corpus size. F1 increases sta8s8cally significantly when we provide more human feedbacks. 0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.35 1.E+2 1.E+3 1.E+4 1.E+5 1.E+6 1.E+7 1.E+8 F1, Recall, or Precision Corpus size F1 Recall Precision Without human feedback With full human feedback 1Kdoc corpus 1.8Mdoc corpus ClueWeb This work is generously supported by the Air Force Research Laboratory (AFRL) under prime contract no. FA875009C0181. Any opinions, findings, and conclusion or recommendaDons expressed in this work are those of the authors and do not necessarily reflect the views of any of the above sponsors including DARPA, AFRL or the US government. Follow stateoftheart approaches in each step, but study them in a new level of scalability (up to 100 million documents). Department of Computer Science, University of WisconsinMadison, USA Distant Supervision(DS)* automaDcally generates labeled training data from a text corpus. However, some labels are not accurate . 0 0.20 0.10 0.15 0.05

Transcript of Big$DataVersusthe$Crowd · *Mike#...

Page 1: Big$DataVersusthe$Crowd · *Mike# Mintz#etal..#Distantsupervision#for#relaon#extracDon#withoutlabeled#data.#In# ACL$2009.$ Takeaways$ When#developing#aDS#system,#one#should#firstexpand#the#training##

Big  Data  Versus  the  Crowd  Looking  for  Rela8onships  in  All  the  Right  Places  

Ce  Zhang,  Feng  Niu,  Christopher  Ré,  and  Jude  Shavlik  

Execu8ve  Summary   Big  Data  vs.  the  Crowd  

Methodology:  Follow  State-­‐of-­‐the-­‐art  DS  Scheme  

Feature  Extrac8on   Distant  Supervision  

To  Combat  Inaccurate  Labels  in  DS  

How  does  increasing  the  corpus  size  impact  the  quality  of  DS?  

How  does  increasing  the  amount  of  human  feedback  impact  the  quality  of  DS?  

Hypothesis:  larger  the  corpus,  bePer  the  quality      

Men8on  Obama  Barack  Obama  B.  Obama  

Men8on1   Men8on2   Feature  B.  Obama   Michelle                                                

       

Men8on   En8ty  Obama  Barack  Obama  B.  Obama  

Men8on  extrac8on  

Linguis8c  paPern  extrac8on  

En8ty  Linking  

•  Person,  LocaDon,  OrganizaDon  ….  

•  Use  Stanford  CoreNLP  

•  Link  menDons  to  Freebase  using  exact  string  matching  

•  Dependency  path  between  menDons.  

•  Word  sequence  between  menDons.  

PERSON  

PERSON  

marry  nsubj  

dobj  

Distant  Supervision  (DS)  Given  Freebase  as  a  knowledge  base  

find  menDon  pairs  supporDng  Freebase.  

En8ty   En8ty  

Human  Feedback  (HF)  Annotate  menDon  pairs  generated  by  distant  supervision  with  Y/N:              -­‐  3  turkers  for  each  pair              -­‐  with  quality  control  

Hypothesis:  more  human  feedback,  bePer  the  quality      

Distant  Supervision  HasSpouse  

Lilly  LedbeZer  joined  Obama  and  his  wife,  Michelle,…  

Senator  Barack  Obama  and  Michelle  Obama  …  

En8ty   En8ty  

Use  broad  coverage  and  redundancy  in  the  large  corpus.  

Ask  humans  in  the  crowd  to  provide  feedbacks.  

*  Mike  Mintz  et  al..  Distant  supervision  for  relaDon  extracDon  without  labeled  data.  In  ACL  2009.  

Takeaways  When  developing  a  DS  system,  one  should  first  expand  the  training    Corpus  in  order  to  improve  recall  and  then  worry  about  the  precision  of  training  examples.  

Mo8va8ng  Applica8on:  DeepDive  DeepDive  approaches  relaDon  extracDon  using  human  analysts  and  staDsDcal  signals  from  terabytes  of  data.  It  applies  distant  supervision  to  generate  training  examples  from  unlabeled  text  corpus.      

DeepDive

hZp://research.cs.wisc.edu/hazy/deepdive/  

Barack  Obama  brought  his  wife  Michelle    to  Kenya  three  years  later…  Text  

Subject   Spouse  Barack  Obama   Michelle  Obama  George  W.  Bush   Laura  Welch  Bill  Clinton   Hillary  Rodham  George  H.  W.  Bush   Barbara  Pierce  

HasSpouse  

Facts  

Contribu8on  Empirically  study  the  factors  that  contribute  to  distant  supervision  quality  and  their  relaDve  impact.  

ClueWeb   Future  Work  •  Study  more  sophis8cated  models  (than  LR)  

and  distant  supervision  scheme.  •  Explore  other  (more  effec8ve)  usage  of  

human  feedbacks.    

Experiments  

Experiment  Setup  

Machine  Learning  Model  Train  a  logisDc  regression  model  using  training  data  obtained  from  DS  and  HF.  

To  Study  our  Hypotheses…  •  Subsample  the  document  corpus  to  get  

different  DS  training  set.  •  Subsample  the  human  annotaDons  to  get  

different  HF  training  set.  

Data  sets  (another  data  set  reported  in  paper)  TAC-­‐KBP:  1.8M  news  arDcles  ClueWeb:  500M  Web  pages  

1000  

10000  

100000  

1000000  

1800000  

Corpus  Size  

Human  Feedback    Amount  

F1  

Impact  of  Big  Data  

The  larger  the  corpus  size  is,  the  higher  the  quality  we  can  expect.  

F1  is  a  log-­‐linear  func8on  in  the  corpus  size  that  we  used  for  distant  supervision.  

Large  corpus  increases  the  coverage  of  linguis8c  paPerns  in  DS.  

0  0.1  0.2  0.3  0.4  0.5  

Precision  

0  

0.1  

0.2  

Recall  

Corpus  Size  

TAC-­‐KBP  

Human  Feedback  Amount  

Impact  of  the  Crowd  

Human  feedback  has  comparable  precision  as  DS  in  our  current  protocol.  

0  0.1  0.2  0.3  0.4  0.5  

Precision  

0  

0.1  

0.2  

Recall  

But  the  slope  is  smaller  than  increasing  the  corpus  size.  

F1  increases  sta8s8cally  significantly  when  we  provide  more  human  feedbacks.    

0.00  0.05  0.10  0.15  0.20  0.25  0.30  0.35  

1.E+2   1.E+3   1.E+4   1.E+5   1.E+6   1.E+7   1.E+8  

F1,  R

ecall,  or  

Precision  

Corpus  size  

F1  Recall  Precision  

Without  human  feedback   With  full  human  feedback   1K-­‐doc  corpus   1.8M-­‐doc  corpus  

ClueWeb                                                This  work  is  generously  supported  by  the                                                Air  Force  Research  Laboratory  (AFRL)  under  prime  contract  no.  FA8750-­‐09-­‐C-­‐0181.  Any  opinions,  findings,  and  conclusion  or  recommendaDons  expressed  in  this  work  are  those  of  the  authors  and  do  not  necessarily  reflect  the  views  of  any  of  the  above  sponsors  including  DARPA,  AFRL  or  the  US  government.  

Follow  state-­‐of-­‐the-­‐art  approaches  in  each  step,  but  study  them  in  a  new  level  of  scalability  (up  to  100  million  documents).  

Department  of  Computer  Science,  University  of  Wisconsin-­‐Madison,  USA  

Distant  Supervision(DS)*  automaDcally  generates  labeled  training  data  from  a  text  corpus.  However,  some  labels  are  not  accurate.  

0  

0.20  

0.10  

0.15  

0.05