Elamur Research
☰
home / papers / 2006 / Discriminant-Based MRS Banking

Discriminant-Based MRS Banking

What it is about

The paper describes how to build a corpus in which every sentence is paired with a meaning representation (MRS). Sentences are parsed with a broad-coverage grammar, and the right analysis is then chosen by binary decisions. The decisions are on semantic discriminants, which are local properties of the semantic dependency graph. The paper defines three types: presence of a predicate, dependencies between predicates, and properties of a relation's main index. An annotator accepts or rejects each discriminant, and every decision removes analyses from the parse forest. The examples are in Norwegian and English. The paper gives no evaluation numbers.

Ideas

For Mova

The code is annot and en/src/explain.rs; the chapter for the clause decisions is languages/en/21-clause-structure.md. The paper is NC, so it is summary only, and it has no numbers to reuse. Measure on EWT only.

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 4.0 (ELRA, LREC proceedings)" does not allow redistributing the text itself. Original — aclanthology.org/L06-1214.pdf

Summary

The paper describes how to build an MRS bank: a corpus where each sentence is linked to a manually confirmed Minimal Recursion Semantics (MRS) logical form. The work serves the Norwegian LOGON machine translation project, which uses semantic transfer and needs verified analyses to debug the grammar, to support transfer development, and to train parse selection models. The authors take the discriminant-based approach of the LinGO Redwoods treebank and apply it to semantics. Sentences are parsed with a broad-coverage precision grammar, the Norwegian LFG grammar NorGram. Annotators then choose the intended reading by accepting or rejecting small semantic contrasts. To compare semantic structures across competing analyses, each MRS is turned into a variable-free elementary dependency graph built around each relation's distinguished variable (usually ARG0). From this graph come three kinds of discriminants: single relations, relation–role–relation triples, and relation–property–value triples. The paper is a design proposal and reports no evaluation results yet. Its value is that the approach does not depend on a particular grammar framework and works directly on the representation that downstream modules consume.

Key points

  • The approach adapts the LinGO Redwoods environment, which combines a tree comparison tool with the [incr tsdb()] profiling database. Redwoods holds about 25,000 annotated utterances produced with the ERG.
  • In Redwoods, a second-year undergraduate was trained for three to four weeks and then disambiguated about 2,000 sentences per week. A minimal disambiguating set of discriminants is computed using Carter's (1997) inference rules.
  • There are three semantic discriminant types: (a) a relation alone, (b) a relation, a role and another relation, and (c) a relation, a variable property such as NUM or TENSE, and its value. All are extracted from a variable-free elementary dependency graph.
  • The planned first MRS banking target is part of the LOGON development corpus, about 5,000 sentences on back-country activities in Norway. Comparisons with syntactic Redwoods treebanking (learning curve, throughput, consistency) are left for future work.
  • The related TrePil project uses LFG-specific c- and f-structure discriminants for syntactic disambiguation. MRS banking ignores syntactic ambiguities that yield the same MRS.

Interesting ideas

  • Eliminating variables to make analyses comparable. Competing analyses may name equivalent variables differently. Anchoring each relation on a distinguished variable and rewriting the structure as labelled dependencies between predicates makes the analyses directly comparable. This could help with diffing any graph-based meaning representation.
  • Saving annotator decisions, not just final choices. Recording every accept or reject decision lets the treebank be updated semi-automatically when the grammar changes, by replaying those decisions on fresh parser output. This suits any annotation project that depends on a grammar or model that keeps evolving.
  • Annotating in stages with coarser discriminants first. For very ambiguous sentences, annotators could start with simple unlabeled bracketing decisions, possibly seeded by a phrase-boundary detector, and switch to finer contrasts only when needed. This could reduce effort on long, highly ambiguous inputs.
  • Disambiguating at the level the application uses. When selection happens over semantic forms, readings that differ only in syntax collapse mechanically. The ranking model then trains on the actual downstream task, at the cost of having fewer features available. This trade-off is worth weighing whenever one chooses the representation level for a reranker.
  • Partial disambiguation with flagging. Annotators can narrow down a sentence partway and mark it for later completion. This keeps throughput high when an item needs knowledge the annotator does not have at hand.

Cite

The paper itself, not this card. BibTeX from the library's record of the work. DOI 10.63317/2ojuhq5ubcyf

@inproceedings{oepen2006discriminant,
  title = {{Discriminant-Based MRS Banking}},
  author = {Oepen, Stephan and Lønning, Jan Tore},
  booktitle = {Proceedings of the Fifth International Conference on Language Resources and Evaluation (LREC'06)},
  year = {2006},
  doi = {10.63317/2ojuhq5ubcyf},
  url = {https://doi.org/10.63317/2ojuhq5ubcyf}
}

References

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

  1. Bond, F., Fujita, S., Hashimoto, C., Kasahara, K., Nariyama, S., Nichols, E., Ohtani, A., Tanaka, T., & Amano, S. (2004). The Hinoki Treebank. A treebank for text understanding. In Proceedings of the 1st International Joint Conference on Natural Language Processing (pp. 158 – 168). Hainan, China. find
  2. Bos, J. (1995). Predicate logic unplugged. In Proceedings of the tenth Amsterdam colloquium (pp. 133 – 142). Amsterdam, The Netherlands. find
  3. Butt, M., Dyvik, H., King, T. H., Masuichi, H., & Rohrer, C. (2002). The Parallel Grammar project. In Proceedings of the COLING Workshop on Grammar Engineering and Evaluation (pp. 1 – 7). Taipei, Taiwan. find
  4. Carter, D. (1997). The TreeBanker. A tool for supervised training of parsed corpora. In Proceedings of the Workshop on Computational Environments for Grammar Development and Linguistic Engineering. Madrid, Spain. find
  5. Copestake, A., Flickinger, D., Sag, I. A., & Pollard, C. (1999). Minimal Recursion Semantics. An introduction. In preparation, CSLI Stanford, Stanford, CA. find
  6. Dyvik, H. (1999). The universality of f-structure. Discovery or stipulation? The case of modals. In Proceedings of the 4th International Lexical Functional Grammar Conference. Manchester, UK. find
  7. Flickinger, D. (2000). On building a more efficient grammar by exploiting types. Natural Language Engineering, 6 (1), 15 – 28. find our card
  8. Ginzburg, J., & Sag, I. A. (2000). Interrogative investigations. The form, meaning, and use of English interrogatives. Stanford, CA: CSLI Publications. find
  9. Oepen, S., & Carroll, J. (2000). Performance profiling for parser engineering. Natural Language Engineering, 6 (1), 81 – 97. find
  10. Oepen, S., Dyvik, H., Lønning, J. T., Velldal, E., Beermann, D., Carroll, J., Flickinger, D., Hellan, L., Johannessen, J. B., Meurer, P., Nordgård, T., & Rosén, V. (2004). Som å kapp-ete med trollet? Towards MRS-based Norwegian – English Machine Translation. In Proceedings of the 10th International… find
  11. Oepen, S., Flickinger, D., & Bond, F. (2004). Towards holistic grammar engineering and testing. Grafting treebank maintenance into the grammar revision cycle. In Proceedings of the IJCNLP workshop Beyond Shallow Analysis. Hainan,China. find
  12. Oepen, S., Flickinger, D., Toutanova, K., & Manning, C. D. (2004). LinGO Redwoods. A rich and dynamic treebank for HPSG. Journal of Research on Language and Computation, 2(4), 575 – 596. find our card
  13. Oepen, S., Toutanova, K., Shieber, S., Manning, C., Flickinger, D., & Brants, T. (2002). The LinGO Redwoods treebank. Motivation and preliminary applications. In Proceedings of the 19th International Conference on Computational Linguistics. Taipei, Taiwan. find our card
  14. Reyle, U. (1993). Dealing with ambiguity by underspecification. Construction, representation, and deduction. Journal of Semantics, 10, 123 – 179. find
  15. Rosén, V., Smedt, K. D., Dyvik, H., & Meurer, P. (2005). TrePil. Developing methods and tools for multilevel treebank construction. In Proceedings of the 4th Workshop on Treebanks and Linguistic Theories (pp. 161 – 172). Barcelona, Spain. find