Elamur Research
☰
home / papers / 2014 / Aligning English Strings with Abstract Meaning Representation Graphs

Aligning English Strings with Abstract Meaning Representation Graphs

What it is about

The paper aligns the tokens of an English sentence to the concepts and relations of its AMR graph, as a step toward learning semantic interpretation and generation rules. Each AMR is linearized by a depth-first walk that keeps the graph structure for rebuilding it later. Before training, function-like tokens (articles, the copula, role labels, instance markers) are removed, and their positions are remembered. Sense suffixes are stripped. The alignment model is IBM-style: it is trained symmetrically in both directions, with an EM step that pushes the two directional tables to agree. Among the variants compared, the symmetric Model 4 is best, with an F-score of 86.5 on the development set and 83.1 on the test set. Role labels are the hard part. Their F-score is about 60 on development and about 49 on test, against about 92 and 90 for concepts. Removing stop words costs about a third of the recall errors.

Ideas

For Mova

The ideas apply to tools/translate/ibm.rs and hebrew.rs (the IBM aligner that reads Hebrew onto meta tokens), tools/translate/measure.rs (the per-class report), and tools/roundtrip (the tree and frame graph whose order the linearizer must reproduce). The paper's hand-aligned gold set is small. Mova should treat it as a check on a method, not as training data; its licence allows only summaries.

Library summary

The Research library's own summary of the paper, in our words — not the paper's text.

Written for the Research library (claude-opus-5-5) from the paper text, in our own words: the license "CC BY-NC-SA 3.0 (ACL Anthology, before 2016)" does not allow redistributing the text itself. Original — aclanthology.org/D14-1048.pdf

Summary

AMR corpora pair English sentences with semantic graphs but provide no word-to-concept links. Such links are needed to learn rules for parsing into AMR and generating text from it. The authors frame the task as a variant of statistical machine translation word alignment and adapt string-to-string IBM models to graph/string pairs.

Each AMR is turned into a token sequence by depth-first traversal. Both sides are then normalized: frequent unalignable tokens are dropped, sense suffixes are stripped, and tokens are cut to four-letter stems. The corpus is augmented with identically spelled token pairs and with pairs drawn from lexical lists. Training alternates EM runs in the two directions, re-initializing each direction's translation and distortion tables from the other, which pushes the parameters toward agreement. A postprocessing step restores the full graph alignment.

Evaluated on newly hand-aligned data, the system reaches 86.5 F on development and 83.1 F on test. This was the first published learned AMR/English aligner and gold alignment set.

Key points

  • Data: 13,050 sentence pairs from LDC AMR release 1.0 (LDC2014T12). The authors hand-aligned 200 of them, split into 100 dev and 100 test pairs.
  • The symmetric training scheme is implemented with MGIZA++. It runs 5 iterations each of Model 1 and HMM, then 4 iterations of modified Model 4 ("Model 4+").
  • Model 4+ improves F over plain Model 4 by 1.7 points on dev (86.5 vs 84.8) and 3.1 points on test (83.1 vs 80.0). The gain comes mainly from higher precision (94.1 on dev, 92.4 on test).
  • Role tokens (e.g., :arg0, :polarity) are the weak spot:
    • Role-token F is 59.7 on dev and 49.3 on test.
    • Concept-token F is 92.5 on dev and 89.8 on test.
    • Human annotators also align role tokens inconsistently.
  • A heuristic rule links patterns like person/thing + :argN-of + concept to the concept's English word (e.g., "worker").

Topics: Machine translation

Cite

The paper itself, not this card. BibTeX from the library's record of the work. DOI 10.3115/v1/d14-1048

@inproceedings{pourdamghani2014aligning,
  title = {{Aligning English Strings with Abstract Meaning Representation Graphs}},
  author = {Pourdamghani, Nima and Gao, Yang and Hermjakob, Ulf and Knight, Kevin},
  booktitle = {Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP)},
  year = {2014},
  doi = {10.3115/v1/d14-1048},
  url = {https://doi.org/10.3115/v1/d14-1048}
}

References

7 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 Linguistic Annotation Workshop (LAW VII-ID), ACL. find our card
  2. Peter F. Brown, Vincent J. Della Pietra, Stephen A. Della Pietra, and Robert L. Mercer. 1993. The mathematics of statistical machine translation: Parameter estimation. Computational linguistics, 19(2):263–311. find
  3. Jeffrey Flanigan, Sam Thomson, Jaime Carbonell, Chris Dyer, and Noah A. Smith. 2014. A discriminative graph-based parser for the abstract meaning representation. In ACL. find
  4. Michel Galley, Mark Hopkins, Kevin Knight, and Daniel Marcu. 2004. What’s in a translation rule? In HLT-NAACL. find our card
  5. Qin Gao and Stephan Vogel. 2008. Parallel implementations of word alignment tool. In Software Engineering, Testing, and Quality Assurance for Natural Language Processing Workshop, ACL. find
  6. Bevan Jones, Jacob Andreas, Daniel Bauer, Karl Moritz Hermann, and Kevin Knight. 2012. Semantics-based machine translation with hyperedge replacement grammars. In COLING. find our card
  7. Percy Liang, Ben Taskar, and Dan Klein. 2006. Alignment by agreement. In HLT-NAACL. find