projection of trees across parallel texts
play

Projection of Trees across Parallel Texts Daniel Zeman, Rudolf Rosa - PowerPoint PPT Presentation

Projection of Trees across Parallel Texts Daniel Zeman, Rudolf Rosa April 17, 2020 NPFL120 Multilingual Natural Language Processing Charles University Faculty of Mathematics and Physics Institute of Formal and Applied Linguistics unless


  1. Projection of Trees across Parallel Texts Daniel Zeman, Rudolf Rosa April 17, 2020 NPFL120 Multilingual Natural Language Processing Charles University Faculty of Mathematics and Physics Institute of Formal and Applied Linguistics unless otherwise stated

  2. Projection of Trees across Parallel Texts Bootstrapping Parsers via Syntactic Projection across Parallel Texts Projection of Trees across Parallel Texts 1/14 • Rebecca Hwa, Philip Resnik, Amy Weinberg, Clara Cabezas, Okan Kolak (2004). • In Natural Language Engineering 1 (1): 1–15. Cambridge University Press • Source: English • Target: Spanish, Chinese • Dependency trees (not phrase structure)

  3. Projection System Architecture Projection of Trees across Parallel Texts 2/14

  4. • unaligned English – – create new empty word • one-to-many – – then create new empty • many-to-one – , and delete other alignments • many-to-many – decompose: fjrst one-to-many, then many-to-one • unaligned foreign – leave them out of the projected tree Direct Projection Projection of Trees across Parallel Texts aligned to – then keep the head of uniquely aligned to instead to align to , and set aligned with , parent of and , aligned with a unique that for any so not aligned with any word in projected from English as follows: 3/14 Given sentence pair ( E, F ) and a set of syntactic relations for E , where E = e 1 , ..., e n is an English sentence and F = f 1 , ..., f m is its non-English parallel, syntactic relations R ( x, y ) are • one-to-one – e i aligned with a unique f x and e j aligned with a unique f y – then R ( e i , e j ) ⇒ R ( f x , f y )

  5. • one-to-many – – then create new empty • many-to-one – , and delete other alignments • many-to-many – decompose: fjrst one-to-many, then many-to-one • unaligned foreign – leave them out of the projected tree Direct Projection to align to Projection of Trees across Parallel Texts aligned to – then keep the head of uniquely aligned to instead , parent of , and set aligned with projected from English as follows: 3/14 Given sentence pair ( E, F ) and a set of syntactic relations for E , where E = e 1 , ..., e n is an English sentence and F = f 1 , ..., f m is its non-English parallel, syntactic relations R ( x, y ) are • one-to-one – e i aligned with a unique f x and e j aligned with a unique f y – then R ( e i , e j ) ⇒ R ( f x , f y ) • unaligned English – e j not aligned with any word in F – create new empty word f y so that for any e i aligned with a unique f x , R ( e i , e j ) ⇒ R ( f x , f y ) and R ( e j , e i ) ⇒ R ( f y , f x )

  6. • many-to-one – , and delete other alignments • many-to-many – decompose: fjrst one-to-many, then many-to-one • unaligned foreign – leave them out of the projected tree Direct Projection Projection of Trees across Parallel Texts aligned to – then keep the head of uniquely aligned to 3/14 projected from English as follows: Given sentence pair ( E, F ) and a set of syntactic relations for E , where E = e 1 , ..., e n is an English sentence and F = f 1 , ..., f m is its non-English parallel, syntactic relations R ( x, y ) are • one-to-one – e i aligned with a unique f x and e j aligned with a unique f y – then R ( e i , e j ) ⇒ R ( f x , f y ) • unaligned English – e j not aligned with any word in F – create new empty word f y so that for any e i aligned with a unique f x , R ( e i , e j ) ⇒ R ( f x , f y ) and R ( e j , e i ) ⇒ R ( f y , f x ) • one-to-many – e i aligned with f x , ..., f y – then create new empty f z , parent of f x , ..., f y , and set e i to align to f z instead

  7. • many-to-many – decompose: fjrst one-to-many, then many-to-one • unaligned foreign – leave them out of the projected tree Direct Projection Projection of Trees across Parallel Texts 3/14 projected from English as follows: Given sentence pair ( E, F ) and a set of syntactic relations for E , where E = e 1 , ..., e n is an English sentence and F = f 1 , ..., f m is its non-English parallel, syntactic relations R ( x, y ) are • one-to-one – e i aligned with a unique f x and e j aligned with a unique f y – then R ( e i , e j ) ⇒ R ( f x , f y ) • unaligned English – e j not aligned with any word in F – create new empty word f y so that for any e i aligned with a unique f x , R ( e i , e j ) ⇒ R ( f x , f y ) and R ( e j , e i ) ⇒ R ( f y , f x ) • one-to-many – e i aligned with f x , ..., f y – then create new empty f z , parent of f x , ..., f y , and set e i to align to f z instead • many-to-one – e i , ..., e j uniquely aligned to f x – then keep the head of e i , ..., e j aligned to f x , and delete other alignments

  8. • unaligned foreign – leave them out of the projected tree Direct Projection Projection of Trees across Parallel Texts 3/14 projected from English as follows: Given sentence pair ( E, F ) and a set of syntactic relations for E , where E = e 1 , ..., e n is an English sentence and F = f 1 , ..., f m is its non-English parallel, syntactic relations R ( x, y ) are • one-to-one – e i aligned with a unique f x and e j aligned with a unique f y – then R ( e i , e j ) ⇒ R ( f x , f y ) • unaligned English – e j not aligned with any word in F – create new empty word f y so that for any e i aligned with a unique f x , R ( e i , e j ) ⇒ R ( f x , f y ) and R ( e j , e i ) ⇒ R ( f y , f x ) • one-to-many – e i aligned with f x , ..., f y – then create new empty f z , parent of f x , ..., f y , and set e i to align to f z instead • many-to-one – e i , ..., e j uniquely aligned to f x – then keep the head of e i , ..., e j aligned to f x , and delete other alignments • many-to-many – decompose: fjrst one-to-many, then many-to-one

  9. Direct Projection Projection of Trees across Parallel Texts 3/14 projected from English as follows: Given sentence pair ( E, F ) and a set of syntactic relations for E , where E = e 1 , ..., e n is an English sentence and F = f 1 , ..., f m is its non-English parallel, syntactic relations R ( x, y ) are • one-to-one – e i aligned with a unique f x and e j aligned with a unique f y – then R ( e i , e j ) ⇒ R ( f x , f y ) • unaligned English – e j not aligned with any word in F – create new empty word f y so that for any e i aligned with a unique f x , R ( e i , e j ) ⇒ R ( f x , f y ) and R ( e j , e i ) ⇒ R ( f y , f x ) • one-to-many – e i aligned with f x , ..., f y – then create new empty f z , parent of f x , ..., f y , and set e i to align to f z instead • many-to-one – e i , ..., e j uniquely aligned to f x – then keep the head of e i , ..., e j aligned to f x , and delete other alignments • many-to-many – decompose: fjrst one-to-many, then many-to-one • unaligned foreign – leave them out of the projected tree

  10. Direct Projection Example det Projection of Trees across Parallel Texts case det obj nsubj det nmod case He took a picture nmod det obj nsubj dceru moji Vyfotil si of my daughter 4/14

  11. Direct Projection Example det Projection of Trees across Parallel Texts case det obj nsubj det nmod case He took a picture nmod det obj nsubj dceru moji Vyfotil si of my daughter 4/14

  12. Direct Projection Example case Projection of Trees across Parallel Texts case det obj nsubj det nmod det nmod He took a picture det obj nsubj dceru of my daughter 4/14 f 1 f 2 f 3 Vyfotil si f 6 moji

  13. Direct Projection Example 2 det Projection of Trees across Parallel Texts nsubj case det det case nmod obj He took nsubj dceru moji si Vyfotil of my daughter picture a 5/14

  14. Direct Projection Example 2 det Projection of Trees across Parallel Texts nsubj case det det case nmod obj He took nsubj dceru moji si Vyfotil of my daughter picture a 5/14

  15. Direct Projection Example 2 det Projection of Trees across Parallel Texts nsubj case det det case nmod obj He took nsubj dceru Vyfotil of my daughter picture a 5/14 f 1 si f 6 moji

  16. Direct Projection Example 2 det Projection of Trees across Parallel Texts nsubj case det det case nmod obj He took nsubj dceru Vyfotil of my daughter picture a 5/14 f 1 si f 6 moji

  17. Direct Projection Example 3 nmod Projection of Trees across Parallel Texts x x obj case nsubj det det case det He took obj nsubj dceru moji si Vyfotil of my daughter picture a 6/14

  18. Direct Projection Example 3 nmod Projection of Trees across Parallel Texts x x obj case nsubj det det case det He took obj nsubj dceru moji si Vyfotil of my daughter picture a 6/14

  19. Direct Projection Example 3 nmod Projection of Trees across Parallel Texts x x obj case nsubj det det case det He took obj nsubj dceru Vyfotil of my daughter picture a 6/14 f 1 f 2 si f 6 moji

  20. Direct Projection Example 3 det Projection of Trees across Parallel Texts x x obj case nsubj det det case nmod obj He took nsubj dceru Vyfotil of my daughter picture a 6/14 f 1 f 2 f 4 si f 6 moji

  21. Many-to-One Assumption: nsubj Projection of Trees across Parallel Texts det det case nmod det obj dceru moji si Vyfotil of my daughter picture a He took 7/14 e i , ..., e j Is a Phrase with One Head

  22. Many-to-One Assumption: nsubj Projection of Trees across Parallel Texts det det case nmod det obj dceru moji si Vyfotil of my daughter picture a He took 7/14 e i , ..., e j Is a Phrase with One Head. What if Not?

Recommend


More recommend