Upgrade to Pro — share decks privately, control downloads, hide ads and more …

[論文紹介][ACL2014]Don't count, predict! #gunosydm

D490d541e3d1ab04d5203e8b210b2233?s=47 ysekky
September 02, 2014

[論文紹介][ACL2014]Don't count, predict! #gunosydm

D490d541e3d1ab04d5203e8b210b2233?s=128

ysekky

September 02, 2014
Tweet

Transcript

  1. [論文紹介]      Don’t  count,  predict!   A  systema7c  comparison

     of     context-­‐coun7ng  vs.  context-­‐predic7ng   seman7c  vector     Marco  Baroni,  Georgiana  Dinu,  German  Kruszewski   (Center  for  Mind/Brain  Sciences(University  of  Trento,  Italy))   ACL  2014 Yoshifumi  Seki   2014.09.02  @Gunosy研究会
  2. 概要 •  word2vecなどのようなcontext-­‐predic7ng   modelが流行しているが,それらとよく知られ ている頻度ベースのベクトルモデルを体系的 に比較することは行われていない   •  様々なlexical

     seman7cs  taskをいろんなパラ メータを用いて評価した   •  結果としてcontext-­‐predic7ng  modelがcount-­‐ based  modelより優れた結果を示した  
  3. distribu7onal  seman7c  models(DSMs) •  Using    vector  that  keep  track

     of  context   •  for  decades,  raw  co-­‐occurrence  counts  don’t   work  that  well   •  archive  higher  performance,  when  various   transforma7on  are  applied  to  the  raw  vector.   –  ex.  by  reweigh7ng  the  count  for  context   informa7veness  and  smoothing  them  with   dimensionality  reduc7on  techniques.   –  this  vector  op7miza7on  process  is  generally   unsupervised  
  4. new  genera7on  of  DSMs •  The  last  few  years  have

     seen  the  development   of  a  new  genera7on  of  DSMs     – the  vector  directly  es7ma7on  problem   •  the  weights  in  a  word  vector  are  set  to   maximize  the  probability  of  the  contexts  in   the  corpus  
  5. new  genera7on  DSMs  is  aYrac7ve! •  replaces  the  essen7ally  heuris7c

     stacking  of   vector  transforms   •  no  manual  annota7on  cost   •  some  of  the  relevant  methods  can  efficiently   scale  up  to  process  very  large  amounts  of   input  data  
  6. Dataset •  a  corpus  of  about  2.8  billion  tokens  

    – ukWaC   – English  Wikipedia   – Bri7sh  Na7onal  Corpus   •  the  top  300k  most  frequent  words  in  the   corpus  as  target  and  context  elements
  7. Count  models •  using  DISSECT  tookit   –  hYp://clic.cimec.unitn.it/composes/toolkit/  

    •  count  vectors  from  symmetric  context  windows  of  two  and   five  words   •  two  weigh7ng  scheme   –  pointwise  mutual  informa7on(PMI)   –  Local  mutual  Informa7on   •  full  and  compressed  vectors   –  Singular  Value  Decomposi7on   –  Non-­‐nega7ve  Matrix  Factoriza7on   –  ranging  200  to  500  in  steps  of  100   •  In  total,  36  count  model  were  evaluated.  
  8. Predict  models •  using  word2vec  toolkit   •  context  windows

     of  2  and  5  words   •  vector  dimensionality  200  to  500  range  in  steps  of  100   •  k:  number  of  nega7ve  samples     –  5  and  10   •  t:  words  that  occur  with  higher  frequency  than  t  are   aggressively  subsampled   –  without    subsampled   –  t  =  exp(-­‐5)   •  we  evaluate  48  predict  models
  9. Out-­‐of-­‐the-­‐box  models •  Distribu7onal  Memory(dm)   –  Using  “linguis7cally  rich”

     data   –  Baroni  and  Lenci  (2010)   –  hYp://clic.cimec.unitn.it/dm/   •  Collobert  and  Weston  vectors(cw)   –  100  dimensional  vectors  trained  for  two  months  on  the   wikipedia   –  The  vector  were  trained  to  op7mize  the  task  of  choosing   the  right  word  over  a  random  alterna7ve  in  middle  of  an   11  word  context  window   –  Collobert  et  al.  (2011)   –  hYp://ronan.collobert.com/senna/  
  10. Evalua7on  Materials •  Seman7c  relatedness   •  Synonym  detec7on  

    •  Concept  categoriza7on   •  Selec7on  preferences   •  Analogy
  11. Seman7c  relatedness •  Rate  the  degree  of  seman7c  similarity  or

     relatedness  between  two  words   on  numerical  scale   •  Rubenstein  and  Goodenough(1965)(rg)   –  Consists  of  65  noun  pairs   –  state  of  the  art:  Hassan  and  Mihalcea(2011)   •  Exploits  wikipedia  linking  structure  and  word  sense  disambigua7on  technique   •  WordSim353(ws)   –  Finkelstein  et  al.(2002)     –  Consists  of  353  pairs   –  State  of  the  art:  Halawi  et  al.(2014)   •  Predict  models  using  WordNet   •  Agirre  et  al.(2009)  split  ws  set  into  similarity(wss)  and  relatedness(wsr)   •  MEN(men)   –  1000  word  pairs   –  Bruni  et  al.(2014)  
  12. Synonym  detec7on •  TOEFL  set   •  80  mul7ple-­‐choices  ques7on

     that  pair  a  target   term  with  4  synonym  candidate   •  Bullinaria  and  Levy(2012)  archive  100%   accuracy  
  13. Concept  categoriza7on •  The  task  is  to  group  nominal  concepts

     into   natural  categories.   •  Using  CLUTO  toolkit   –  hYp://glaros.dtc.umn.edu/gkhome/views/cluto   •  Almuhared-­‐Poesio  benchmark(ap)   –  Almuhared(2006)   –  492  concepts  into  21  category   •  The  ESLLI  2008  Distribu7onal  Seman7c  Workshop   shared-­‐task  set(esslli)   –  44  concepts  into  6  category
  14. Selec7on  preferences •  Verb-­‐noun  pairs  that  where  rated  by  subject

      for  the  typicality  of  the  noun  as  a  subject  or   object  of  the  verb   •  Ulrike  Pado(2007)  (up)   – 211  pairs   •  Macrae   – 100  noun-­‐verb  pairs  
  15. Analogy •  Specifically  to  test  predict  model   •  9K

     seman7cs  and  10.5K  syntac7c  analogy   ques7on   –  Example   •  Brother-­‐sister,  Grandson-­‐?   –  Ans:  granddaughter   •  work-­‐works,  spreak-­‐?   –  Ans:  speaks   •  En7re  dataset  (an)   –  Syntac7c  subset(ansyn)   –  Seman7c  subset(ansem)
  16. Result

  17. パラメータの選び方に   カウントモデルは強く依存する

  18. Conclusion •  全体として見た時にCount  modelよりpredict   modelのほうがよい   – 精度が全体的に高い   – パラメータ設定・データセットの違いに対してロバ

    ストである   •  Seman7cs,  synonym両方に強い