Unsupervised AMR-Dependency Parse Alignment

What it is about

The paper aligns AMR concepts to the nodes of an English dependency tree, not to spans of words as earlier aligners did. The dependency parse comes from ClearNLP and carries named-entity tags and semantic roles. The aligner scores candidate pairs with lexical, named-entity, role and global features, and trains the feature weights with EM. A beam-search decoder then finds the most likely alignment. On the gold alignments the full feature set reaches an F-score of about 87.1. The intro cites 83.1 for the ISI string aligner of Pourdamghani et al. (2014). When its alignments are given to the CAMR parser, parsing Smatch rises. In the error analysis, about 40% of the errors are meaning-coverage errors: the concept is aligned to the head word of its phrase rather than to its own word. Duplicate words account for 17% and long-distance dependencies for 14%. Parse errors are 4%.

Ideas

• I1. Align meta tokens to tree nodes rather than to word spans, and count how often a concept lands on its governor instead of its own word. The paper's largest error class is exactly this. The count is a direct measure for open problem 1 (reader attachment noise).
• I2. Duplicate words cause 17% of the errors, because the lexicon key is the word alone. Adding the tree relation or the role to the key would separate the two occurrences of an ordinal such as first. Test this as a change to the key in meta.rs and measure the residual.
• I3. The EM step, run for a fixed number of iterations (the paper uses 50), can be replaced by a count table over (concept, head relation, tag) triples with the same iteration rule. Keep the table in BTreeMap order and check that two runs give identical output before any gain is claimed.
• I4. The paper ablates feature groups in turn (lexical, then plus named entities, then plus roles, then plus global features). Run the same ablation on the reader: remove one group at a time and report the reader's bits per class. This is the value-ablation rule applied to the reader.

For Mova

The ideas apply to tools/translate/meta.rs (the lexicon and the head of each meta token), tools/translate/hebrew.rs (the Hebrew words are aligned to meta tokens with IBM Model 1; the tree-feature view is the English counterpart of its prior), and tools/roundtrip (the tree and frame stack the reader reads). The paper's aligner is trained with features from a neural tagger and parser. Mova can use the same features from its own parser as counts, and the decoder is a plain argmax over a table.

Topics: Named entity recognition · Dependency parsing · Machine translation

Cite

The paper itself, not this card. BibTeX from the library's record of the work. DOI 10.18653/v1/e17-1053

@inproceedings{chen2017unsupervised,
  title = {{Unsupervised AMR-Dependency Parse Alignment}},
  author = {Chen, Wei-Te and Palmer, Martha},
  booktitle = {Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 1, Long Papers},
  year = {2017},
  doi = {10.18653/v1/e17-1053},
  url = {https://doi.org/10.18653/v1/e17-1053}
}

References

1 references, 1 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. find our card

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/chen-2017-unsupervised-amr-dependency-parse-alignment/