Capturing Translational Divergences with a Statistical Tree-to-Tree Aligner Mary Hearne, John Tinsley, Ventsislav Capturing Translational Divergences with Zhechev & Andy Way a Statistical Tree-to-Tree Aligner Tree Alignments Translational Divergences Automatic Mary Hearne, John Tinsley, Ventsislav Zhechev & Andy Way Tree-to-Tree Alignment Evaluation National Centre for Language Technology Conclusions & School of Computing Future Work Dublin City University
Parallel treebanks Capturing Translational Divergences with a Statistical Tree-to-Tree Aligner Mary Hearne, A parallel treebank comprises: John Tinsley, Ventsislav Zhechev & Andy ◮ sentence pairs Way ◮ parsed Tree Alignments Translational ◮ word-aligned Divergences ◮ tree-aligned Automatic Tree-to-Tree Alignment (Volk & Samuelsson, 2004) Evaluation Conclusions & The role of alignments: Future Work Santos (1996), paraphrasing Lab (1990): Having a linguistic description of two languages is not the same as having a linguistic description of the translation between them.
Parallel treebanks Capturing Translational Divergences with a Statistical Tree-to-Tree Aligner Mary Hearne, John Tinsley, Ventsislav Zhechev & Andy Way ◮ Our work involves automatically obtaining a parallel treebank from a parallel corpus via parsing and tree Tree Alignments alignment . Translational Divergences ◮ Our overall objective is to use the parallel treebank for Automatic Tree-to-Tree inducing a variety of syntax-aware and syntax-driven Alignment models of translation for use in data-driven MT. Evaluation Conclusions & ◮ In this paper/presentation, the focus is on the capture of Future Work translational divergences through the application of a tree-aligner to gold-standard tree pairs.
Capturing translational divergences Capturing Translational Divergences with a Statistical Tree-to-Tree Aligner Mary Hearne, We aim to: John Tinsley, Ventsislav ◮ make explicit the syntactic divergences between source and Zhechev & Andy Way target sentence pairs Tree Alignments ◮ align to express as precisely as possible the translational Translational equivalences between the tree pair Divergences Automatic ◮ constraining phrase-alignments in the data set is a Tree-to-Tree Alignment consequence of aligning trees, but not an objective Evaluation Conclusions & Future Work We remain agnostic with regard to: ◮ which linguistic formalism is most appropriate for the expression of monolingual syntax ◮ how best to exploit parallel treebanks for syntax-aware data-driven MT
Outline Capturing Translational Divergences with a Statistical Tree-to-Tree Aligner Mary Hearne, John Tinsley, Ventsislav Tree Alignments Zhechev & Andy Way Tree Alignments Translational Divergences Translational Divergences Automatic Tree-to-Tree Alignment Automatic Tree-to-Tree Alignment Evaluation Conclusions & Future Work Evaluation Conclusions & Future Work
Outline Capturing Translational Divergences with a Statistical Tree-to-Tree Aligner Mary Hearne, John Tinsley, Ventsislav Tree Alignments Zhechev & Andy Way Tree Alignments Translational Divergences Translational Divergences Automatic Tree-to-Tree Alignment Automatic Tree-to-Tree Alignment Evaluation Conclusions & Future Work Evaluation Conclusions & Future Work
Tree-to-Tree Alignment Capturing Translational Divergences with a Statistical Tree-to-Tree Aligner Mary Hearne, John Tinsley, Links indicate translational equivalence : Ventsislav Zhechev & Andy Way ◮ a link between root nodes indicates equivalence between the sentence pair Tree Alignments ◮ a link between any given pair of source and target nodes Translational Divergences indicates Automatic Tree-to-Tree ◮ equivalence between the substrings they dominate Alignment ◮ equivalence between the substrings they do not dominate Evaluation Conclusions & Future Work S S NP VP NP VP John V NP John V NP sees Mary voit Mary
Tree-to-Tree Alignment Capturing Translational Divergences with a Statistical Tree-to-Tree Aligner Mary Hearne, John Tinsley, Ventsislav Zhechev & Andy Way In the simplest case: Tree Alignments ◮ the sentence lengths are identical Translational Divergences ◮ the word order is identical Automatic Tree-to-Tree ◮ the tree structures are isomorphic Alignment Evaluation Conclusions & Future Work S S NP VP NP VP John V NP John V NP sees Mary voit Mary
Tree-to-Tree Alignment Capturing Translational Divergences with a Statistical Tree-to-Tree Aligner Mary Hearne, John Tinsley, Ventsislav Slightly more complex: Zhechev & Andy Way ◮ not every node in each tree needs to be linked Tree Alignments ◮ each node is linked at most once Translational Divergences ◮ terminal nodes are not linked Automatic Tree-to-Tree Alignment Evaluation VP Conclusions & Future Work VP V PP V NP cliquez P NP click D ADJ N sur D N ADJ the Save Asbutton le bouton Enregistrer Sous
Tree-Alignment vs. Word-Alignment Capturing Translational Divergences with a Statistical Tree-to-Tree Aligner Mary Hearne, Word-alignment : unaligned words are problematic and to be John Tinsley, Ventsislav avoided Zhechev & Andy Way Tree-alignment : unaligned nodes are informative Tree Alignments ... Jacob’s ladder ... − → ... l’´ echelle de Jacob ... Translational Divergences Automatic Word alignment: Tree alignment: Tree-to-Tree Alignment Evaluation Jacob la Conclusions & Future Work NP NP ’s ´ echelle ladder de NP NP NP PP Jacob PN POS N D N P NP Jacob ’s ladder la ´ echelle de PN Jacob
Hierarchical alignments Capturing Translational Divergences with a Statistical Tree-to-Tree Aligner On the relationship between ’s and de in Mary Hearne, John Tinsley, Ventsislav ... Jacob’s ladder ... − → ... l’´ echelle de Jacob ... Zhechev & Andy Way Tree Alignments ’s − → de Translational Divergences X ’s Y − → Y de X Automatic Tree-to-Tree Alignment NP 1 ’s NP 2 − → NP 2 de NP 1 Evaluation Conclusions & NP → NP 1 ’s NP 2 : NP → NP 2 de NP 1 Future Work NP NP NP POS NP NP PP ’s P NP de
Outline Capturing Translational Divergences with a Statistical Tree-to-Tree Aligner Mary Hearne, John Tinsley, Ventsislav Tree Alignments Zhechev & Andy Way Tree Alignments Translational Divergences Translational Divergences Automatic Tree-to-Tree Alignment Automatic Tree-to-Tree Alignment Evaluation Conclusions & Future Work Evaluation Conclusions & Future Work
Nominalisation Capturing Translational Divergences with a Statistical Tree-to-Tree Aligner Mary Hearne, John Tinsley, Ventsislav Zhechev & Andy Way Tree Alignments VP NP Translational Divergences V NP N PP Automatic Tree-to-Tree removing the print head retraite P NP Alignment de la tˆ ete d’impression Evaluation Conclusions & Future Work
Lexical Divergences Capturing Translational Divergences with a Statistical Tree-to-Tree Aligner Mary Hearne, John Tinsley, Ventsislav Zhechev & Andy Way Tree Alignments Translational CONJP CONJP Divergences Automatic CONJ S CONJ S Tree-to-Tree Alignment as au fur et a ` mesure que Evaluation Conclusions & Future Work
Context-Dependent Lexical Selection Capturing Translational Divergences with a Statistical Tree-to-Tree Aligner S S Mary Hearne, John Tinsley, PRON VPv Ventsislav PRON VPverb Zhechev & Andy (1) you V VPinf Way vous V VPverb need PART VPv devez Tree Alignments to Translational Divergences S Automatic S Tree-to-Tree PRON VPv Alignment PRON VPverb (2) you V VPinf Evaluation il V VPverb Conclusions & need PART VPv Future Work faut to CONJPsub CONJsub S PP (3) if PRON VPv P NPdet you V NP pour need
Embedded Complexities Capturing Translational Divergences with a Statistical Tree-to-Tree Aligner Mary Hearne, John Tinsley, Ventsislav Zhechev & Andy Way PP P NP Tree Alignments Translational CONJP pendant DETP NP Divergences Automatic CONJ S PRE D N PP Tree-to-Tree Alignment while NP VP toute la dur´ eeP NP Evaluation the scanner AUX VP de D N PP Conclusions & Future Work is AUX V le ´ etalonnage P NP being calibrated de le scanner
Structural Dissimilarity Capturing Translational Divergences with a Statistical Tree-to-Tree Sadj Aligner Mary Hearne, CONJPsub COMMA S John Tinsley, Ventsislav CONJsub S , NP VPcop Zhechev & Andy Way if NPadj VPaux D NPadj V NP Tree Alignments A N AUX V the N PP is N CONJ N Translational unauthorisedrepair is performed remainder P NP null and void Divergences Automatic of D NPzero Tree-to-Tree Alignment the N N Evaluation warranty period Conclusions & S Future Work NPdet VPv D NPpp V NPdet toute N APvp invaliderait D N intervention Amod V la garantie non autoris´ ee ‘any unauthorised action would invalidate the guarantee’
Recommend
More recommend
