1,a) Graham Neubig 1,b) Sakriani Sakti 1,c) 1,d) 1,e) 2 1. (Statistical Machine Translation: SMT[1]) [2] [3] [4][5][6] 2 1 (a) 3 approach 1 Nara Institute of Science and Technology a) miura.akiba.lr9@is.naist.jp b) neubig@is.naist.jp c) ssakti@is.naist.jp d) tomoki@is.naist.jp e) s-nakamura@is.naist.jp 1 (a) ( - - ) (b) (c) 1(b) 1(c) c 2015 Information Processing Society of Japan 1
(Synchronous Context-free Grammar: SCFG[7]) (Multi-Synchronous CFG: MSCFG[8]) MSCFG SCFG SCFG MSCFG Europarl 4 2. 2.1 (Hierarchical Phrase-Based Translation: Hiero[7]) SCFG SCFG X s, t (1) X s t s t X X 0 of X 1,X 1 X 0 (2) Chiang SCFG SCFG φ(s t), φ(t s) φ lex (s t), φ lex (t s) (t ) ( 1) 6 CKY+ [9] [7] 2.2 MSCFG[8] SCFG SCFG t MSCFG N X s, t 1,..., t N (3) MSCFG SCFG MSCFG SCFG Neubig MSCFG 2 SCFG MSCFG 3. SMT 3 (Cascade): [3] (Synthetic): [3] (Triangulation): [4] [5] c 2015 Information Processing Society of Japan 2
3.1 Cohn [4] ( ) T SP, T PT T SP, T PT s, p, p, t p X s, t (4) φ( ) φ lex ( ) T ST φ ( t s ) = φ ( t p ) φ (p s) (5) φ ( s t ) = φ (s p) φ ( p t ) (6) ( ) ( ) φ lex t s = φ lex t p φlex (p s) (7) ( ) φ lex s t = (8) φ lex (s p) φ lex ( p t ) (5)-(8) φ ( t p, s ) = φ ( t p ) (9) φ ( s p, t ) = φ (s p) (10) 3.2 p SCFG MSCFG X s, t, p (11) (5)-(8) φ(t, p s), φ(s p, t) φ ( t, p s ) = φ ( t p ) φ (p s) (12) φ ( s p, t ) = φ (s p) (13) φ(p s) φ(s p) φ lex (p s) φ lex (s p) T SP 10 t p 2 1 13 MSCFG s, t s, t, p Neubig T 1 T 2 T 1 - [8] s T 1 φ(t 1 s) L t 1 t 1 φ(t 1, t 2 s) t 2 4. 4.1 Europarl [10] (en) (de), (es), (fr), (it) 4 5 Gale-Church [11] 90 1,500 10 10 200 Travatar [12] Hiero SCFG BLEU [13] MERT[14] BLEU MSCFG L = 20 T 1 MSCFG 6 c 2015 Information Processing Society of Japan 3
BLEU Score [%] Source Target Tri. SCFG Tri. MSCFG Tri. MSCFG Tri. MSCFG Direct Cascade (baseline) -PivotLM +PivotLM 100k +PivotLM 2M es 27.10 25.05 25.31 25.38 25.52 25.75 de fr 25.65 23.86 24.12 24.16 24.25 24.58 it 23.04 20.76 21.27 21.42 21.65 22.29 de 20.11 18.52 18.77 18.97 19.08 19.40 es fr 33.48 27.00 29.54 29.87 29.91 29.95 it 27.82 22.57 25.11 25.01 25.18 25.64 de 19.69 18.01 18.73 18.77 18.87 19.19 fr es 34.36 27.26 30.31 30.53 30.73 31.00 it 28.48 22.73 25.31 25.50 25.72 26.22 de 19.09 14.03 17.35 17.99 18.17 18.52 it es 31.99 25.64 28.85 28.83 29.01 29.31 fr 31.39 25.87 28.48 28.40 28.63 29.02 1 BLEU ( : p<0.05, : p<0.01) Direct: SCFG Cascade: SCFG Tri. SCFG: SCFG SCFG Tri. MSCFG: SCFG MSCFG -Pivot +PivotLM 100k/2M 10 200 4.2 1 BLEU 200 0.4 1.2 BLEU BLEU Score [%] Direct Tri. SCFG Tri. MSCFG 23.2 23 22.8 22.6 22.4 22.2 22 21.8 21.6 21.4 21.2 0 500000 1x10 6 1.5x10 6 2x10 6 Pivot-LM Size [sent.] 2 ( - ) MSCFG SCFG 2 c 2015 Information Processing Society of Japan 4
( ): ich bedaure, daß es keine gemeinsame annäherung gegeben hat. ( ): sono spiacente del mancato approccio comune. Tri. SCFG: mi rammarico per il fatto che non si ravvicinamento comune. (BLEU+1: 13.84) Tri. MSCFG+PivotLM 2M: mi dispiace che non esiste un approccio comune. (BLEU+1: 25.10) i regret that there is no common approach. (Generated English Sentence) Tri. MSCFG+PivotLM 2M MSCFG approccio approach 5. 2 SCFG 1 MSCFG MSCFG Microsoft CORE [6] Xiaoning Zhu, Zhongjun He, Hua Wu, Conghui Zhu, Haifeng Wang, and Tiejun Zhao. Improving Pivot-Based Statistical Machine Translation by Pivoting the Cooccurrence Count of Phrase Pairs. In Proc. EMNLP, 2014. [7] David Chiang. Hierarchical phrase-based translation. Computational Linguistics, Vol. 33, No. 2, pp. 201 228, 2007. [8] Graham Neubig, Philip Arthur, and Kevin Duh. Multi- Target Machine Translation with Multi-Synchronous Context-free Grammars. In Proc. NAACL, 2015. [9] Jean-Cédric Chappelier, Martin Rajman, et al. A Generalized CYK Algorithm for Parsing Stochastic CFG. TAPD, Vol. 98, No. 133-137, p. 5, 1998. [10] Philipp Koehn. Europarl: A parallel corpus for statistical machine translation. In MT summit, Vol. 5, pp. 79 86, 2005. [11] William A Gale and Kenneth W Church. A program for aligning sentences in bilingual corpora. Computational linguistics, Vol. 19, No. 1, pp. 75 102, 1993. [12] Graham Neubig. Travatar: A Forest-to-String Machine Translation Engine based on Tree Transducers. In Proc. ACL Demo Track, pp. 91 96, 2013. [13] Kishore Papineni, Salim Roukos, Todd Ward, and Wei- Jing Zhu. BLEU: a method for automatic evaluation of machine translation. In Proc. ACL, pp. 311 318, 2002. [14] Franz Josef Och. Minimum Error Rate Training in Statistical Machine Translation. In Proc. ACL, pp. 160 167, 2003. [1] Peter F. Brown, Vincent J.Della Pietra, Stephen A. Della Pietra, and Robert L. Mercer. The Mathematics of Statistical Machine Translation: Parameter Estimation. Computational Linguistics, Vol. 19, pp. 263 312, 1993. [2] Christopher Dyer, Aaron Cordova, Alex Mont, and Jimmy Lin. Fast, easy, and cheap: construction of statistical machine translation models with MapReduce. In Proc. WMT, pp. 199 207, 2008. [3] Adrià de Gispert and José B. Mariño. Catalan-English Statistical Machine Translation without Parallel Corpus: Bridging through Spanish. In Proc. of LREC 5th Workshop on Strategies for developing machine translation for minority languages, 2006. [4] Trevor Cohn and Mirella Lapata. Machine Translation by Triangulation: Making Effective Use of Multi-Parallel Corpora. In Proc. ACL, pp. 728 735, June 2007. [5] Masao Utiyama and Hitoshi Isahara. A Comparison of Pivot Methods for Phrase-Based Statistical Machine Translation. In Proc. NAACL, pp. 484 491, 2007. c 2015 Information Processing Society of Japan 5