Improving Tree-to-Tree Translation with Packed Forests

Transcription

1 Improving Tree-to-Tree Translation with acked Forests Yang Liu, Yajuan Lü, and QunLiu Institute of Computing Technology, Chinese Academy of Sciences STITUTE OF COMUTG 8/10/2009 ACL/IJCNL 2009, Singapore 1

2 Tree-to-Tree Translation Tree-to-Tree (e.g., Eisner 2003, Zhang et al., 2008) STITUTE OF COMUTG I V V-B VS AS NN bushi yu shalong juxing le huitan yu shalong Bush held a talk with Sharon N VBD DT NN N V with Sharon S 8/10/2009 ACL/IJCNL 2009, Singapore 2

3 Tree-to-Tree Translation Tree-to-Tree (e.g., Eisner 2003, Zhang et al., 2008) STITUTE OF COMUTG I V V-B VS AS NN bushi yu shalong juxing le huitan yu X1 Bush held a talk with Sharon N VBD DT NN N with X1 V S 8/10/2009 ACL/IJCNL 2009, Singapore 3

4 Challenges STITUTE OF COMUTG Tree-to-tree approaches face two major challenges: most vulnerable to parsing error poorest rule coverage 8/10/2009 ACL/IJCNL 2009, Singapore 4

5 arsing Error STITUTE OF COMUTG I V-B both trees can be ill-formed! CC VS AS NN bushi yu shalong juxing le huitan Bush held a talk with Sharon N VBD DT NN N S V 8/10/2009 ACL/IJCNL 2009, Singapore 5

6 Rule Coverage STITUTE OF COMUTG I V-B CC VS AS NN bushi yu shalong juxing le huitan Bush held a talk with Sharon N VBD DT NN N S V 8/10/2009 ACL/IJCNL 2009, Singapore 6

7 This Work STITUTE OF COMUTG We replace 1-best trees with packed forests to alleviate the two problems: parsing error rule coverage Our approach outperforms the tree-based system dramatically (+3.6) and achieves comparable performance with Moses. 8/10/2009 ACL/IJCNL 2009, Singapore 7

8 acked Forest I Tomita, 1985; Billot and Lang, 1989 I STITUTE OF COMUTG V V-B VS AS NN bushi yu shalong juxing le huitan V-B CC VS AS NN bushi yu shalong juxing le huitan I V V-B CC VS AS NN bushi yu shalong juxing le huitan 8/10/2009 ACL/IJCNL 2009, Singapore 8

9 Tree-based Rule Extraction STITUTE OF COMUTG I Zhang et al., 2008; Lavieet al., 2008 V V-B VS AS NN bushi yu shalong juxing le huitan yu shalong Bush held a talk with Sharon N VBD DT NN N with Sharon V N S 8/10/2009 ACL/IJCNL 2009, Singapore 9

10 Forest-based Rule Extraction I STITUTE OF COMUTG V V-B CC VS AS NN bushi yu shalong juxing le huitan Bush held a talk with Sharon N VBD DT NN N S V 8/10/2009 ACL/IJCNL 2009, Singapore 10

11 Forest-based Rule Extraction STITUTE OF COMUTG Following GHKM (Galley et al., 2004), our extraction method invovles three steps: identify the correspondence between nodes identify minimal rules get composed rules 8/10/2009 ACL/IJCNL 2009, Singapore 11

12 Frontier Nodes I STITUTE OF COMUTG V V-B CC VS AS NN bushi yu shalong juxing le huitan Bush held a talk with Sharon N VBD DT NN N S V 8/10/2009 ACL/IJCNL 2009, Singapore 12

13 Identify Tree airs I STITUTE OF COMUTG V V-B CC VS AS NN bushi yu shalong juxing le huitan yu shalong Bush held a talk with Sharon N VBD DT NN N with Sharon N V S 8/10/2009 ACL/IJCNL 2009, Singapore 13

14 Decomposition STITUTE OF COMUTG yu shalong yu shalong = with Sharon with Sharon N N N 8/10/2009 ACL/IJCNL 2009, Singapore 14

15 Frontier Trees and Tree airs STITUTE OF COMUTG frontier tree minimal frontier tree frontier tree pair minimal frontier tree pair N 8/10/2009 ACL/IJCNL 2009, Singapore 15

16 Identify Tree airs for the Node step 1: identify corresponding frontier nodes I STITUTE OF COMUTG V V-B CC VS AS NN bushi yu shalong juxing le huitan Bush held a talk with Sharon N VBD DT NN N S V 8/10/2009 ACL/IJCNL 2009, Singapore 16

17 Identify Tree airs for the Node step 2: identify frontier trees for each node I STITUTE OF COMUTG V V-B CC VS AS NN bushi yu shalong juxing le huitan Bush held a talk with Sharon N VBD DT NN N S V 8/10/2009 ACL/IJCNL 2009, Singapore 17

18 Identify Tree airs for the Node step 2: identify frontier trees for each node I STITUTE OF COMUTG V V-B CC VS AS NN bushi yu shalong juxing le huitan Bush held a talk with Sharon N VBD DT NN N S V 8/10/2009 ACL/IJCNL 2009, Singapore 18

19 Identify Tree airs for the Node step 2: identify frontier trees for each node STITUTE OF COMUTG N N 8/10/2009 ACL/IJCNL 2009, Singapore 19

20 Identify Tree airs for the Node step 3: build frontier trees pairs STITUTE OF COMUTG N N N N 8/10/2009 ACL/IJCNL 2009, Singapore 20

21 Identify Tree airs for the Node step 4: identify minimal frontier trees pairs STITUTE OF COMUTG N N N N 8/10/2009 ACL/IJCNL 2009, Singapore 21

22 Difficulty in Finding Tree airs STITUTE OF COMUTG A minimal frontier tree pair is not necessarily a pair of minimal frontier trees N N 8/10/2009 ACL/IJCNL 2009, Singapore 22

23 Difficulty in Finding Tree airs I STITUTE OF COMUTG V V-B CC VS AS NN bushi yu shalong juxing le huitan Bush held a talk with Sharon N VBD DT NN N S V 8/10/2009 ACL/IJCNL 2009, Singapore 23

24 Difficulty in Finding Tree airs I STITUTE OF COMUTG minimal frontier tree V-B CC VS AS le minimal frontier tree pair VBD non-minimal frontier tree S V 8/10/2009 ACL/IJCNL 2009, Singapore 24

25 Get Composed Rules STITUTE OF COMUTG + yu with = yu with 8/10/2009 ACL/IJCNL 2009, Singapore 25

26 Experiments STITUTE OF COMUTG Chinese-to-English translation Training set: 31K sentence pairs with 840K Chinese words and 950K English words Language Model: 3-gram trained on the English side Development set: NIST 2002 Test set: NIST 2005 Metric: case-insensitive BLEU4 8/10/2009 ACL/IJCNL 2009, Singapore 26

27 Tree-based Vs. Forest-based STITUTE OF COMUTG p avg. trees # of rules BLEU K K M 348K M 574K M 743K /10/2009 ACL/IJCNL 2009, Singapore 27

28 Extraction and Decoding Time STITUTE OF COMUTG p avg. trees extraction decoding M M M extraction: milliseconds / sentence pair decoding: seconds / sentence 8/10/2009 ACL/IJCNL 2009, Singapore 28

29 Comparison with Moses STITUTE OF COMUTG training Moses this work 840K+950K M+9.41M /10/2009 ACL/IJCNL 2009, Singapore 29

30 Conclusion STITUTE OF COMUTG acked forests help alleviate two problems that tree-to-tree approaches face: negative impact of parsing mistakes on translation quality poor rule coverage 8/10/2009 ACL/IJCNL 2009, Singapore 30

31 STITUTE OF COMUTG Thanks! 8/10/2009 ACL/IJCNL 2009, Singapore 31