N-best. N-best Reranking Using Optimal Phrase Alignment for Statistical Machine Translation

Transcription

1 Vol. 51 No (Aug. 2010) N-best N-best Reranking Using Optimal Phrase Alignment for Statistical Machine Translation Mitsuru Koshikawa, 1 Masao Utiyama, 2 Shunji Umetani, 3 Tomomi Matsui 4 and Mikio Yamamoto 1 Phrase-based statistical machine translation system outputs the candidate having the highest probability based on the probabilistic phrase translation rules. However, there exist a huge number of translation candidates and ambiguities on phrase segmentations/alignments for source and target sentences. Therefore, the current statistical translation systems use various heuristics for reducing the number of translation candidates and approximating phrasealignment probabilities, in order to narrow the search space. This paper proposes the formulation to strictly maximize the phrase-alignment probability computed from all features which most phrase-based statistical machine translation systems use within. We also propose a reranking method based on the proposed phrase alignment optimization. In evaluation experiments, our system improved significantly the translation quality. The experimental results also suggested that a variety of translation candidates are more important for increasing accuracy than exact phrase alignments ) 1 2),3) 2) 5) 2 6) 1 Graduate School of Systems and Information Engineering, University of Tsukuba 2 MASTAR MASTAR Project, National Institute of Informaiton and Communications Technology 3 Graduate School of Information Science and Technology, Osaka University 4 Department of Information and System Engineering, Faculty of Science and Engineering, Chuo University 1443 c 2010 Information Processing Society of Japan

2 1444 N-best phrase alignment 7) 4) phrase aligner f e ê ê 1) ê =argmax P (e) P (f, c e) e e arg e arg e max e c P (e)max c P (f, c e) max P (e)p (f, c e) (1) e,c (1) P (e) e c f e (1) 2 maxc c arg e max e,c f e max max 2) (1) (2) 3) ê =arg e max e,c Fig. 1 1 An example of phrase-based translation. λ k h k (f, c, e) (2) k h k (f, c, e) k λ k (1) (1) ) 1 2 2) P (f, c e) =P (c e) P (f c, e) P ( c I 1 e ) I P ( ) fci ē i i=1 f ē c I 1 = c 1,c 2,...,c I c i i ē i ē i c i f ci P ( f ci ē i) P (c I 1 e) 2.3 Lexicalized Block Orientation 4) 2) (3)

3 1445 N-best i (i +1) Lexicalized Block Orientation Lexicalized Block Orientation LBO i (i +1) ē i, ē i+1 f ci f ci+1 3 4) monotone (c i+1 = c i +1 ), class (c i,c i+1) = swap (c i+1 = c i 1 ), (4) discontinuous ( ). monotone swap discontinuous 4) 1 today rainy swap LBO 3 (3) 1 4) P ( c I 1 e ) I P (class (c i,c i+1) ē i, ē i+1) (5) i=1 P (class (c i,c i+1) ē i, ē i+1) P ( class (c i,c i+1) f ci, f ) ci+1, ē i, ē i f ci f ci+1 2) monotone 0 f ci end i f ci+1 start i+1 d(end i,start i+1) 2) Fig. 2 2 Examples of correct phrase alignment (left) and incorrect phrase alignment (right). d (end i,start i+1) = end i start i+1 +1 (6) (3) ) 2 f 1 f 4 e 1 e f, e fci, ē i,p( f ci ē i) 7) ˆ f I 1, ˆē I 1, ĉ I 1 =arg max P ( ) I f 1 ē I 1,c I (ēi I,c I 1, f 1 I =f,ēi 1 =e 1 P 1,c I 1 e ) (7) (7) 1 2 (3) 2 1 (1) 2 max c

4 1446 N-best ( min ( f, e ) ξ ) I=1 I I ξ ξ 500 min ( f, e ) solver 7) fck, ē k 2 x k {0, 1} F 7) F F E 7) f = f 1,f 2,f 3,f 4, e = e 1,e 2,e F E (8) 1 f 1 f 2 F F = ,E = (8) (7) 2 7) 7) maximize Fig. 3 3 Candidates of phrase pairs for input paralell sentences. x k t k k K subject to F x = 1, (9) Ex = 1, x k {0, 1} ( k K). t k λ tm log P ( f ck ē k ) λ tm K 1 =(1,...,1) x =(x 1,...,x K) F x = 1 F x 1 1 Ex = N N-best (1) arg e max e,c c

5 1447 N-best Fig. 4 4 Flowchart of N-best reranking method using phrase aligner. max c max c LBO x 2 x e 1 e 3 3 (9) F x = 1 s g LBO 5 3 Fig. 5 Directed graph built from the target side phrases in 3. LBO (9) a a 1 0 y a maximize x k t k + y ar a k K subject to F x = 1, My = b, Ny = x, a A x k {0, 1} ( k K), y a {0, 1} ( a A). (10)

6 1448 N-best My = b s g b +1 1 b s = 1 b g =+1 b others =0 b s s b g g b others N Ny = x y x (10) Ny = x F x = 1 My = b F x = 1 My = b x y Ny = x (10) x y Ny = x A r a a r a My = b (11) 1 M s g (11) 5 y 4 + y 5 y 6 =0 4 y 4 + y 5 y y 1 y 2 y 3 y 4 y 5 y = y 4 y x 4 1 y 4 + y 5 = x 4 y x Ny = x y 1 y 2 y 3 y 4 y 5 y 6 = x 1 x 2 x 3 x 4 1 (11) (12) 1 y x N NTCIR-7 8) development Minimum Error-Rate Training 9) NTCIR-7 1 BLEU 10)

7 1449 N-best Table 1 1 NTCIR-7 Description of the data set on NTCIR-7 patent translation task. 1,798,571 59,974, ,435 1,798,571 64,184, ,652 dev ,028 3,986 dev ,427 3,653 1,381 45,334 4,116 1,381 48,737 3,882 2 Table 2 Experimental condition. Moses 08/02/20 release 10, 20, 50, 100, 200, 500, 1,000 ttable-limit 20 msd-bidirectional-fe 5gram, Interpolated Modified Kneser-Ney solver ILOG CPLEX 11.0 Fig. 6 6 Effect of the beam width on translation quality (BLEU). Moses 5) 1 N-best N-best SRI Language Modeling Toolkit 11) 2 Moses 1 N-best solver CPLEX ) 4.2 BLEU 6 Moses rerank Moses Moses BLEU 5% 1 Moses N-best Moses 7 Fig. 7 Effect of the beam width on the average score improvement. 200 Moses 500 Moses Moses N-best BLEU

8 1450 N-best Fig. 8 8 BLEU Effect of the search time on translation quality (BLEU) BLEU 8 Moses rerank Moses Moses Moses 2 5. N-best NTCIR-7 Moses BLEU 1) Brown, P.F., Pietra, V.J.D., Pietra, S.A.D. and Mercer, R.L.: The Mathematics of Statistical Machine Translation: Parameter Estimation, Computational Linguistics, Vol.19, No.2, pp (1993). 2) Koehn, P., Och, F.J. and Marcu, D.: Statistical Phrase-Based Translation, Proc. Human Language Technology and North American Association for Computational Linguistics Conference (2003). 3) Och, F.J. and Ney, H.: The Alignment Template Approach to Statistical Machine Translation, Computational Linguistics, Vol.30, No.4, pp (2004). 4) Tillmann, C. and Zhang, T.: A Localized Prediction Model for Statistical Machine Translation, Proc. 43rd Annual Meeting of the Association for Computational Linguistics, Association for Computational Linguistics, pp (2005). 5) Koehn, P., Hoang, H., Birch, A., Callison-Burch, C., Federico, M., Bertoldi, N., Cowan, B., Shen, W., Moran, C., Zens, R., Dyer, C., Bojar, O., Constantin, A. and Herbst, E.: Moses: Open Source Toolkit for Statistical Machine Translation, Proc. 45th Annual Meeting of the Association for Computational Linguistics Companion Volume Proceedings of the Demo and Poster Sessions, Association for Computational Linguistics, pp (2007). 6) Hasan, S., Zens, R. and Ney, H.: Are Very Large N-Best Lists Useful for SMT?, Proc. Human Language Technologies: The Conference of the North American Chapter of the Association for Computational Linguistics; Companion Volume, Short Papers, Association for Computational Linguistics, pp (2007). 7) DeNero, J. and Klein, D.: The Complexity of Phrase Alignment Problems, Proc. 46th Annual Meeting of the Association for Computational Linguistics: Human Language Technology, Short Papers, Association for Computational Linguistics, pp (2008).

9 1451 N-best 8) Fujii, A., Utiyama, M., Yamamoto, M. and Utsuro, T.: Overview of the Patent Translation Task at the NTCIR-7 Workshop, Proc. 7th NTCIR Workshop Meeting on Evaluation of Information Access Technologies: Information Retrieval, Question Answering and Cross-lingual Information Access, NTCIR, pp (2008). 9) Och, F.J.: Minimum Error Rate Training in Statistical Machine Translation, Proc. 41st Anuual Meeting on Association for Computational Linguistics, Association for Computational Linguistics, pp (2003). 10) Papineni, K., Roukos, S., Ward, T. and Zhu, W.J.: BLEU: A Method for Automatic Evaluation of Machine Translation, Proc. 40th Annual Meeting of the Association for Computational Linguistics, Association for Computational Linguistics, pp (2002). 11) Stolcke, A.: SRILM An Extensible Language Modeling Toolkit, Proc. International Conference on Spoken Language Processing, pp (2002). 12) ILOG: ILOG CPLEX 11.0 User s Manual, ILOG (2007). ( ) ( ) OR INFORMS MPS ACL ACL