Generating English from Abstract Meaning Representations

What it is about

The paper treats AMR-to-English generation as phrase-based machine translation, with the AMR first turned into a flat token string in an English-like order. The key component is a learned linearisation: token-aligned AMR and English pairs show where each edge should go. Three methods are compared: a pre-order depth-first baseline, a majority method that memorises the most common order per role set, and a classifier made of three maximum-entropy models (drop the instance-of edge, place each role before or after it, order each pair of roles). Total alignment crossings on dev fall from 46,671 (depth-first) to 33,772 (majority) and 35,603 (classifier), and the classifier cuts adjacent crossings most. With cleaning and name, number and date realisers, the best system reaches about 26.8 to 26.9 BLEU on test, against 22.0 for the tree-transducer system.

Ideas

• I1. Learn an ordering of relation labels from counts: for each pair of relations under the same concept, count how often one precedes the other in aligned data, smooth with add-one, and order children by pairwise score. This is deterministic and can be compared with the side tables in en/src/lin.rs.
• I2. Use alignment crossings as a linearisation residual on dev. Count crossings for each candidate order before any BLEU run. This is a cheap test for changes to the order model that does not depend on the decoder.
• I3. Concept dropping: the classifier drops a concept when it is probably unaligned, and 97% of the concepts it drops are in fact unaligned. In the meta-world, a concept with no source word after reading could be marked droppable. Measure precision and recall on the Bible pairs.
• I4. Name, number and date realisers as a fixed rule table, added as extra training pairs. For Bible text this is the place for proper names; keep it a rule table with a negative control.

For Mova

The alignment-crossings measure fits en/src/lin.rs, and the dropping rule fits tools/translate/meta.rs. Both are count-based, so the result can be checked for identical output in two runs. The maximum-entropy classifiers are not needed; the counts are.

Topics: Machine translation

Cite

The paper itself, not this card. BibTeX from the library's record of the work. DOI 10.18653/v1/w16-6603

@inproceedings{pourdamghani2016generating,
  title = {{Generating English from Abstract Meaning Representations}},
  author = {Pourdamghani, Nima and Knight, Kevin and Hermjakob, Ulf},
  booktitle = {Proceedings of the 9th International Natural Language Generation conference},
  year = {2016},
  doi = {10.18653/v1/w16-6603},
  url = {https://doi.org/10.18653/v1/w16-6603}
}

References

11 references, 3 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Laura Banarescu, Claire Bonial, Shu Cai, Madalina Georgescu, Kira Griffitt, Ulf Hermjakob, Kevin Knight, Philipp Koehn, Martha Palmer, and Nathan Schneider. 2013. Abstract Meaning Representation for sembanking. In Proc. ACL Linguistic Annotation Workshop (LAW). find our card
2. Peter F. Brown, John Cocke, Stephen A. Della Pietra, Vincent J. Della Pietra, Fredrick Jelinek, John D. Lafferty, Robert L. Mercer, and Paul S. Roossin. 1990. A statistical approach to machine translation. Computational linguistics, 16(2):79–85. ↗
3. Simon Corston-Oliver, Michael Gamon, Eric Ringger, and Robert Moore. 2002. An overview of Amalgam: A machine-learned generation module. In Proc. INLG. find
4. Jeffrey Flanigan, Chris Dyer, Noah A. Smith, and Jaime Carbonell. 2016. Generation from abstract meaning representation using tree transducers. In Proc. NAACL. find our card
5. Yuqing Guo, Haifeng Wang, and Josef Van Genabith. 2011. Dependency-based n-gram models for general purpose sentence realisation. Natural Language Engineering, 17(4):455–483. ↗
6. Philipp Koehn, Hieu Hoang, Alexandra Birch, Chris Callison-Burch, Marcello Federico, Nicola Bertoldi, Brooke Cowan, Wade Shen, Christine Moran, Richard Zens, et al. 2007. Moses: Open source toolkit for statistical machine translation. In Proc. ACL Poster and Demonstration Sessions. find
7. Irene Langkilde and Kevin Knight. 1998. Generation that exploits corpus-based statistical knowledge. In Proc. ACL. find
8. Uri Lerner and Slav Petrov. 2013. Source-side classifier preordering for machine translation. In Proc. EMNLP. find
9. François Mairesse, Milica Gašić, Filip Jurčı́ček, Simon Keizer, Blaise Thomson, Kai Yu, and Steve Young. 2010. Phrase-based statistical language generation using graphical models and active learning. In Proc. ACL. find
10. Nima Pourdamghani, Yang Gao, Ulf Hermjakob, and Kevin Knight. 2014. Aligning English strings with Abstract Meaning Representation graphs. In Proc. EMNLP. find our card
11. Le Zhang. 2004. Maximum entropy modeling toolkit for Python and C++. http://bit.ly/1DGnb2p. ↗

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/pourdamghani-2016-generating-english-abstract-meaning-representations/