ãã®è¨äºã¯ 第3回 自然言語処理勉強会@東京 ã§ã®çºè¡¨è³æã§ãã
EMNLP 2010 (The 2010 Conference on Empirical Methods on Natural Language Processing) の論文ãï¼æ¬ï¼ï¼æ¬ç´¹ä»ãã¦ãã¾ãã質ããæ°ã§åè² ã
è«æãåºãæµ ãèªãã¹ã¹ã¡(ä»®)
- ããã¾ãã¾æè¿ã«ãã£ãæ¹æ³è«ãã«åºå·ããå¯è½æ§
- by ææ©ãã (IBIS 2010 ã®ãªã¼ããã³ã°ã»ãã·ã§ã³ã«ã¦)
- ä¾)é¢ä¿ãªããé¢å¿ãªããé£ããã¦ããããããªããã(èªãã§ã|èãã¦ã)ã ã
- ä¾)èªããããªãããããããªè«æã ãèªã
- ä¾)ã¨ãããããã¤ã¼ããã¤ãºããã®æ å ã§ã©ãã¾ã§ãæ³¥èã
- è«æãæå½ãã次第ã«ãåºãæµ
ãèªããããã
- ï¼æ¬ãããï¼ãï¼æé
- ç®å®ã¯ãtwitter ã§ï¼ï¼ï¼åã¤ã¶ããããããã(ã
- ãã¡ããã¡ããã¨èªãè«æã¯ã¡ããã¨èªããã ãã©ã
- ã§ããã¾ã èªç¶è¨èªå¦çã®åå¼·å§ããã°ãã
- ããã£ã¦èªãã§ããã¨ããèªä¿¡ã¯ãªã(ããªã
- ã¿ã¤ãã«ã«ä½¿ããã¦ããç¨èªããç¥ããªãã¨ã
[Boyd-Graber, Resnik] Holistic Sentiment Analysis Across Languages: Multilingual Supervised Latent Dirichlet Allocation
- supervised LDA (Blei+ 2007) ãè¤æ°è¨èªã«æ¡å¼µ
- è¨èªæ¨ªæã§ææ ã®äºæ¸¬ãè¡ã
- ææ äºæ¸¬ã«è¾æ¸ã synsets ãªã©ã®æ å ±ã使ã
supervised LDA (Blei+ 2007)
- Latent Dirichlet Allocation (Blei+ 2003) ï¼ response variable Y_d
- Y ã¯ã¬ãã¥ã¼ã®ç¹æ°ãææ ã»æè¦ãªã©ã表ãå®å¤æ°
- Draw response variable
MLSLDA (Boyd-Graber+ 2010)
- SLDA + 左㮠"Multilingual Topics" ã®é¨å
- h: synset, l: language
- Draw transiton probabilities
- Draw stop probabilities
- Draw emission probabilities for that synset
- LDA ã®åèªã®ä½ç½®ã«ã¯ (synsets ã®) path λ ãæ¥ããpath ããåèªã®ç起確çã¯Ï
- path ã¨ãã£ã¦ãæ½è±¡çãªèå¥åãªã®ã§ãè¾æ¸ã®å ´åã¯é ç®çªå·ãªã©ã«è©²å½
- åèªã®ç起確çã«è¾æ¸ã synsets ã®æ å ±ãåæ ã§ãããã¨ã§äºæ¸¬ç²¾åº¦ãåä¸
å®é¨
çµæ
MLSLDA (Boyd-Graber+ 2010)
- ãã¤ãèª/è±èªã¨ãã¤ãèª/ä¸å½èªã®ããããã«ã¤ãã¦ããããã¯ã®ç¹æ°ã¨ç起確çã®é«ãåèªã¨ãã¦åãæå³ãæã¤åèªãæãããã¦ãã
- åè¨èªãã¨ã« SLDA ããã®ã§ããææ
ã®äºæ¸¬ã¨ããç¹ã§ã¯åæ§ã®çµæãå¾ãããããªæ°ãã
- è¤æ°è¨èªã«ã¾ããã£ã¦ãããã¯ãå ±æãããã¤æå³ã®è¿ãåèªãããªããããã¯ããçèµ·ããã確çãé«ãã¨ä½ãå¬ããã ããï¼
- ç SLDA ã¨ã®äºæ¸¬ç²¾åº¦ã®æ¯è¼ã欲ããã£ãã¨ããã
MLSLDA (Boyd-Graber+ 2010)
- åä¸è¨èªã§ã¯ãè¾æ¸ã synsets ã®æ å ±ã使ããã¨ã§ããããã使ããªã Flat ãã精度ãé«ã
- è¤æ°è¨èªã®ãã¼ã¿ã»ãããç¨ããã¨ãè¾æ¸ãªã©ã使ããªãã¦ãåããããã®ç²¾åº¦ã«ãªã
[Zhao, Jiang, Yan, Li] Jointly Modeling Aspects and Opinions with a MaxEnt-LDA Hybrid
- å¾æ¥ææ³ã§ã¯ aspect åºæã® opnion words ãæ±ããªã
- Maximum Entropy (Markov Model) 㨠LDA ãçµ±åããã¢ãã«ãæã¡è¾¼ããã¨ã§ãaspect åºæã® opinion words ãæ±ããã¨ãå¯è½ã«
ã°ã©ãã£ã«ã«ã¢ãã« (Zhao+ 2010)
- ç¢å°ãå ¨é¨ w ã«éã¾ã£ã¦ãâ¦â¦ã観測å¤æ°ãå¤ãâ¦â¦ã
- ãã¡ãã exact ã«ã¯è§£ããªã(Forward-Backword Propagation ã使ããªã)ã®ã§ãMCMC ã§è§£ã
- æ´æ°å¼ã¯è«æè¦ã¦ãï¼
å®é¨
- ãã¼ã¿ã»ããï¼ã¬ã¹ãã©ã³ã¬ãã¥ã¼ãããã«ã¬ãã¥ã¼
- Gibbs sampling 500å iterations
- LocLDA(Brody+ NAACL2010:æªèª) ã¨æ¯è¼
çµæ
ã¬ã¹ãã©ã³ã® aspect 㨠opinion (Zhao+ 2010)
- LocLDA ã¨ã®æ¯è¼ã¯ã¡ãã£ã¨è¦ããã
- aspect 㨠opinion ãåé¢ã§ããªãææ³ããç¡çç¢çåé¢ãããããã«ãã¦è©ä¾¡ãã¦ããã®ã§ãææ¨å¤ãé常ã«ä½ãåºã¦ããããã®æ¹æ³ã§æ¯è¼ããæå³ãããã®ãçåã
[Murata, Ohno, Matsubara] Automatic Comma Insertion for Japanese Text Generation.
- æ¥æ¬èªã®ã«ã³ã(èªç¹)ã¯å¤å½äººã«ã¯é£ãã
- ä½éããã® usage ããã(ç¯ã®åºåãã並åããªã©ãªã©)
- ããã¹ãããã«ã³ããæã¤ã¹ãå ´æãèªåçã«è¦ã¤ãã
- é³å£°èªèã§ã¯æ¯ç¶ãã®éãªã©ããã³ãã«ä½¿ã£ãããããããã®ç 究ã¯ããã¹ãã®ã¿ã対象ã¨ãã¦ãã
ææ³
- ã«ã³ãã®ä½¿ããæ¹ãåæãï¼éãã«åé¡
- commas between clauses
- commas indicating clear dependency relations
- commas for avoiding reading mistakes and reading difficulty
- commas indicating the subject
- commas inserted after a conjunction or adverb at the beginning of a sentence
- commas inserted between parallel words or phrases
- commas inserted after an adverbial phrase to indicate time
- commas emphasizing the adjacent word
- other
ã©ã®ãããªç¯ã®å¾ã«ã«ã³ããæãããã(Murata+ 2010)
- ãããã 20 ãã¿ã¼ã³ã®ç´ æ§ãã³ãã¬ã¼ãã¨ãã¦è¡¨ç¾ãMaxEnt ã§è§£ã
- æç¯ã®å B = b_1...b_n
- ã«ã³ãå R = r_1...r_n (r_i = 1 ãªã b_i ã®å¾ãã«ã«ã³ããå ¥ãã)
- ã«ã³ãã«ãã£ã¦æç¯ã m åã®é¨ååã«åãã
- (j=1, .., m), (k=1, .., n_j-1),
(Murata+ 2010)
å®é¨
- ãã¼ã¿:京é½ããã¹ãã³ã¼ãã¹
- è¨ç·´ã¨ãã¹ãã§ç°ãªãæéã®è¨äºã使ç¨
- recall 69.13%, precision 84.13%, Få¤ 75.90
[Zhao, Gildea] A Fast Fertility Hidden Markov Model for Word Alignment Using MCMC
- çµ±è¨çæ©æ¢°ç¿»è¨³ã¢ãã«ã®ææ¡
- IBM Model 1 ã HMM ã«ããåèªã¢ã©ã¤ã¡ã³ãã¢ãã«ã« fertility ãå°å
¥
- Gibbs sampling ã§æ¨è«
- é«éã㤠IBM Model 4 ã«è¿«ãæ§è½
çµ±è¨çæ©æ¢°ç¿»è¨³ã£ã¦ï¼
- æ¥æ¬èªã®æç« ãè±èªã«ç¿»è¨³ãããã¨ããçã®ãè±èªã®å ããã¹ãããéé³ã®ããéä¿¡è·¯ãçµã¦ãæå ã«ããæ¥æ¬èªã®æç« ãã«ãªã£ã¦ãããã¨èããäºå¾ç¢ºç P(è±èª|æ¥æ¬èª) ãæ大åãããã¨ã§ãè±èªã®å ããã¹ãããæ¨å®ãã
- çµ±è¨çæ©æ¢°ç¿»è¨³ãã¨ã¯ãã
- http://www.kecl.ntt.co.jp/icl/kpro/taro/papers/smt-tutorial.pdf
ã¢ãã«
ä¸è¬ã« f_1^J:ã½ã¼ã¹, e_1^I:ã¿ã¼ã²ãã (ã¤ã¾ãè±èª e ããã©ã³ã¹èª f ã«ç¿»è¨³), åèªã¢ã©ã¤ã¡ã³ã a_1^J ã«å¯¾ãã
(Zhao and Gildea 2010)
- ãã ãã¿ã¼ã²ãã e_1^I ã«ã¯ I+1 åã® "empty words" ã追å ãã¦èãã
- å¿ ããã翻訳èªãï¼å¯¾ï¼ã«å¯¾å¿ããããã§ã¯ãªãã®ããã¾ãã¢ãã«åããããï¼
- 翻訳ã¢ãã«ã« fertility Ï_1^I 㨠Ï_ε ãå°å
¥
- fertility Ï_i 㯠target e_i ã«ã²ãã¥ã source f_j ã®åæ°
- empty words e_i (i=I+1, .., 2I+1) ã®åã¯åè¨ã㦠Ï_ε ã«
(Zhao and Gildea 2010)
- HMM 翻訳ã¢ãã«ã«ã¤ãã¦ãfertility å°å
¥åã¨å¾
- P(f|e) ã®è¨ç®ã«ã¯ã¢ã©ã¤ã¡ã³ã, fertility ã¨ãã«å¹ãã¦ãã
- ããã Gibbs sampling ã§æ¨è«
å·¦:HMM, å³:HMM with fertilities(Zhao and Gildea 2010)
- IBM Model 1 ã«ã¤ãã¦ã¯çç¥
å®é¨
å·¦:Allignment Error Rate, å³:Training Time(Zhao and Gildea 2010)
- fertility ã®å°å
¥ã«ãã精度ã®åä¸ã確èª
- ã·ã³ãã«ã§é«éãªã¢ãã«ã§ãIBM Model 4 ã«å¹æµãã精度
- Fertility ä»ã HMM ããç HMM ããéãã®ã¯ åè ã Gibbs samplingãå¾è ã Viterbi ãªãã
ãã®ä»ãçãç´¹ä»
- [Mejer+ 2010] Confidence in Structured-Prediction using Confidence-Weighted Models
- CW linear classification ãç³»åã©ããªã³ã°ã«é©ç¨ãã¦CRFã¨æ¯ã¹ã
- NP chunking ã¿ã¹ã¯ã§æ¯è¼
- ã ããã CRF ã«è² ãã¦ããããNER Spanish ã§ã¯ãããã«åã£ã¦ãã(ãã¾ãã¾ç·å½¢åé¢ããããã£ãã¨ãï¼)
- ããªã³ã©ã¤ã³ãªã®ã§éãï¼èéãå©ãå²ãã«ç²¾åº¦ãããããæªããªããã¨ãããã¨ããª
- [Navigli+ 2010] Inducing Word Senses to Improve Web Search Result Clustering.
- Web æ¤ç´¢çµæã®ã¯ã©ã¹ã¿ãªã³ã°ã«ãã Word Sense Induction(åèªæå³æ¨è«?)
- ã¯ã¨ãªã¼ã«å¯¾ãã¦åçã«ã°ã©ããæ§ç¯(åæãã¼ããæ¤ç´¢çµæã®ã¹ããããããå¾ã)ãã¨ãã¸ã®å ±èµ·åº¦ã¯ Dice ã
- ã°ã©ãã®éè·¯?(ä¸è§å½¢&åè§å½¢)ãè©ä¾¡ãã¦ã¹ã³ã¢ã®ä½ã辺ãæ¶å»ããã®ããã¤ã³ãã
- å®é¨ã§ã¯ Suffix Tree ãªã©ã使ç¨ããä»ã®ææ³ã¨æ¯ã¹ã¦å§åçã«ããçµæã辺ãæ¨ã¦ã¦ããã¨ãããå¹ãã¦ããã®ãã測ãæ¹ããã£ã¦ããã®ãã
- [Agarwal+ 2010] Automatic Detection and Classification of Social Events.
- ACE(Automated Content Extraction) ã¿ã¹ã¯ http://www.itl.nist.gov/iad/mig/tests/ace/
- ããã¹ããã social events(人ç©ãå«ã対象ãä»ã®å¯¾è±¡ã«å¯¾ãã¦ç¥ã£ã¦ã/ä½ããããã¨)ãæ½åº
- ææ³ã¯â¦â¦å¼ãï¼ã¤ããªãã£ãã®ã§ï¼ããç解ã§ããªãã£ã(è¦ç¬)
- 代åè©ã®åå®ã¨ãã©ããããã ãããããã¯å¥ã®ã¿ã¹ã¯ï¼
- åç §ããã¦ãã Extracting Social Networks from Literary Fiction(Elson+ ACL2010)ããããããããAustinã®ä½åãªã©ããç»å ´äººç©é¢ä¿å³ãæ½åºã
- [Li+ 2010] Negative Training Data can be Harmful to Text Classification.
- ããã¹ãåé¡ã¿ã¹ã¯ã§è² ä¾ã¯å¿ è¦ãªãã©ãããæ害ã ã£ãããã¨ãã話ã
- ï¼å¤åé¡ãåæã¨ãã¦ãSVN ã¨ãã®æ´¾çï¼ãã¤ã¼ããã¤ãº(!?)ãå®é¨ãã¦ããã ãã©ããã¼ã»ãããã³ç³»ãªããã¸ã¹ãã£ãã¯å帰ãªãã対象ã«å ¥ã£ã¦ãªãã®ã¯ãªãã§ã ããã
- è«æã®çµè«ã¯ããã ããè² ä¾ã¯ãã¡ã ããã¨ããæ¹åããããªãå®ã¯ reliable negatives ãè¦ã¤ããããã®ã¢ã«ã´ãªãºã CR-SVM ã®ææ¡ãã¨ããæ¹åãªã®ã§ã¿ã¤ãã«ã¯é£ã(ã ãã can ãå ¥ã£ã¦ãããã)ã
- [Yao+ 2010] Collective Cross-Document Relation Extraction Without Labeled Data.
- ãã¡ããããã¥ã¡ã³ãããã®æ½åºã¿ã¹ã¯ãrelation ã®æ½åºãããã¥ã¡ã³ã横æã§è¡ãã
- mention ã relation ã MEN/JOINT/PAIR ãªã©ã®ãã¤ããªç´ æ§ã§ã¤ãªããCRF ã§äºå¾ç¢ºçãè©ä¾¡ãLinear Chain ã§ã¯ãªãã®ã§ MCMC ã§æ¨è«
- isolated(ç´ æ§ã¯BIASã¨MENTIONã®ã¿), joint(ãã«), pipeline(isoé¨ã§æ¨è«âãããåºå®ãã¦jointã®æ¨è«)ã¨ããï¼ãã¿ã¼ã³ã§å®é¨ãjointãä¸çªè¯ããããªã®ã«å¿ ãããããã§ãªãã®ããããããã
- "Without Labeled Data" ã¨è¨ãã¤ã¤ãopenNLP POS tagger ã§ãã¼ã¿ã«ã¿ã°ä»ããã¦ããã¨ãæ°ã«ãªãã"Labeled" ã£ã¦ä½ãæãã¦ãããã ãããæ£è§£ï¼ãunsupervised ã£ã¦ãã¨ï¼