Machine-learned contexts for linguistic operations in German sentence realization
What it is about
The Amalgam realiser turns a semantic graph into a German surface tree in eight stages. Linguistic operations are written by hand, but the contexts in which each applies are learned as decision-tree classifiers, trained on 100,000 analysed sentences of technical manuals (70/30 split for training and tuning, then a separate blind set). Each model gets about 600 to 1,170 candidate features taken from the node, its parent and its grandparent, and a small set of computed features such as heaviness. The paper reports four models. Case assignment reaches 93.5% accuracy against a 70.5% baseline. Verb position reaches 94.9% against 42.4% for the most frequent value, the hard part being "undefined". Extraposition, moved one node up at a time, reaches 86.1% against 67.6%, with the largest gain on relative clauses. Syntactic aggregation chooses where a duplicated constituent in a coordination is realised. Trees range from 3 to 447 branching nodes. The point of the paper is that a learned context can be as accurate as the baseline is weak, and that German needs a richer extraposition model than English.
Ideas
- I1. Report every operation against the most-frequent-value baseline, as the paper does. The translator's residual in bits already exists per class; add the baseline column per class, so each counted rule is judged against what it beats.
- I2. Verb position as features for the Hebrew reader. Verb-initial, verb-second and verb-final classes (open problem 7 for word order: verb-subject to subject-verb) can be counted from MACULA trees and used by the normaliser the paper's idea suggests. Measure the order agreement with the English gloss before and after.
- I3. Extraposition by counted position. Relative clauses in WEB sit at different heights in the tree than in KJV. Count the attachment height per author for each clause type, and let the encoder use the author delta for it.
- I4. Aggregation as a counted choice. Where a repeated word in a coordination is realised (first, last or middle) can be a count per construction, not a classifier. Measure on coordinated verses how often the repeated word is dropped in the source and in the output.
For Mova
The relevant code is tools/translate/src/profile.rs (encoder rules), tools/hbo and tools/macula (for I2), and en/src/lin.rs (order). The paper's trees are not Mova's; its method of per-operation contexts with a baseline is. The ideas touch open problems 4, 6 and 7.
Cite
@inproceedings{gamon2002machine,
title = {{Machine-learned contexts for linguistic operations in German sentence realization}},
author = {Gamon, Michael and Ringger, Eric and Corston-Oliver, Simon and Moore, Robert},
booktitle = {Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics},
year = {2002},
doi = {10.3115/1073083.1073089},
url = {https://doi.org/10.3115/1073083.1073089}
}References
- S. Bangalore and O. Rambow 2000. Exploiting a probabilistic hierarchical model for generation. Proceedings of the 18th International Conference on Computational Linguistics (COLING 2000). Saarbrücken, Germany. 42-48. D. M. Chickering. nd. WinMine Toolkit Home Page. http://research.microsoft.com/~dmax/WinMine/Tool doc.htm D. M. Chickering, D. Heckerman and C. Meek. 1997. A Bayesian approach to learning Bayesian networks with local structure. In “Uncertainty in Artificial Intelligence: Proceedings of the Thirteenth Conference”, D. Geiger and P. Punadlik Shenoy, ed., Morgan Kaufman, San Francisco, California, pp. 80-89. ↗
- S. Corston-Oliver, M. Gamon, E. Ringger, and R. Moore. 2002. An overview of Amalgam: A machine-learned generation module. To be presented at INLG 2002. find
- H. Dalianis and E. Hovy 1993 Aggregation in natural language generation. Proceedings of the 4th European Workshop on Natural Language Generation, Pisa, Italy. find
This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/gamon-2002-machine-learned-contexts-linguistic-operations/