Linguistically Motivated Complementizer Choice in Surface Realization

What it is about

The optional complementizer "that" (as in "he said that ...") is left out in about two thirds of Penn Treebank cases, and its presence changes fluency and sometimes meaning. The paper asks whether linguistically motivated features can predict whether a realiser should write "that". The baseline is a global realisation ranker for OpenCCG that makes 82 mismatches with the gold choice among 543 "that" cases in section 00. The authors add local features adapted from Jaeger's study of uniform information density: the length and position of the complement clause relative to the matrix verb, matrix-verb properties (part of speech, stem, form, supertag), and information-density features, namely the average log probability of the first words of the clause under two n-gram models. A binary averaged perceptron predicts the presence of "that". The features improve accuracy over the ranker, and the gains hold in cases where the choice matters for fluency. The paper plans to fold these local features into the global ranker.

Ideas

• I1. Treat the optional "that" as a decision with a count table. Key it by the matrix verb and the length class of the clause, and learn the rate from KJV, where the source choice is known. Measure how often the translator's "that" agrees with the source, per author.
• I2. Information density from counts. The uniform-information-density feature is the average surprisal of the first words of the clause. Compute it from target-side n-gram counts (BTreeMap, deterministic) and use it as a feature in the choice, not as a learned weight.
• I3. A per-verb omission rate. Matrix-verb bias is the strongest lexical signal here. On the meta token of the verb (its concept), store the observed "that" rate. This is open problem 4 for complement clauses.
• I4. Log both options with the reason. Each "that" decision should print the count, the verb and the length class behind it, so the choice can be explained in the report, in line with Mova's transparency rule.

For Mova

The encoder in tools/translate/src/profile.rs makes this choice today by context, with no reason written out. The ideas touch open problem 4 (function words as meta-token features) and open problem 6 (the standard leans to the Bibles). A cheap first measure is the agreement of "that" with the source on KJV and WEB, counted twice.

Topics: Machine translation

Cite

The paper itself, not this card. BibTeX from the library's record of the work.

@inproceedings{rajkumar2011linguistically,
  title = {{Linguistically Motivated Complementizer Choice in Surface Realization}},
  author = {Rajkumar, Rajakrishnan and White, Michael},
  booktitle = {Proceedings of the UCNLG+Eval: Language Generation and Evaluation Workshop},
  year = {2011},
  url = {https://aclanthology.org/W11-2706/}
}

References

15 references, 1 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Jason Baldridge and Geert-Jan Kruijff. 2002. Coupling CCG and Hybrid Logic Dependency Semantics. In Proc. ACL-02. find
2. Rena Torres Cacoullos and James A. Walker. 2009. On the persistence of grammar in discourse formulas: A variationist study of “that”. Linguistics, 47(1):1–43. find
3. Stephen Clark and James R. Curran. 2007. Wide-Coverage Efficient Statistical Parsing with CCG and Log-Linear Models. Computational Linguistics, 33(4):493–552. find our card
4. Michael Elhadad. 1991. FUF: The universal unifier user manual version 5.0. Technical Report CUCS-038-91, Dept. of Computer Science, Columbia University. find
5. John A. Hawkins. 2003. Why are zero-marked phrases close to their heads? In Günter Rohdenburg and Britta Mondorf, editors, Determinants of Grammatical Variation in English, Topics in English Linguistics 43. De Gruyter Mouton, Berlin. find
6. T. Florian Jaeger. 2010. Redundancy and reduction: Speakers manage information density. Cognitive Psychology, 61(1):23–62, August. find
7. Mark Johnson. 2009. How the statistical revolution changes (computational) linguistics. In Proceedings of the EACL 2009 Workshop on the Interaction between Linguistics and Computational Linguistics: Virtuous, Vicious or Vacuous?, pages 3–11, Athens, Greece, March. Association for Computational Linguistics. find
8. Roger Levy and T. Florian Jaeger. 2007. Speakers optimize information density through syntactic reduction. Advances in Neural Information Processing Systems, 19:849. find
9. Hiroko Nakanishi, Yusuke Miyao, and Jun’ichi Tsujii. 2005. Probabilistic methods for disambiguation of an HPSG-based chart generator. In Proc. IWPT-05. find
10. Brian Roark, Murat Saraclar, Michael Collins, and Mark Johnson. 2004. Discriminative language modeling with conditional random fields and the perceptron algorithm. In Proc. ACL-04. find
11. Mark Steedman. 2000. The syntactic process. MIT Press, Cambridge, MA, USA. find
12. S. Tagliamonte and J. Smith. 2005. No momentary fancy! the zero ‘complementizer’ in English dialects. English Language and Linguistics, 9(2):289–309. find
13. Erik Velldal and Stephan Oepen. 2005. Maximum entropy models for realization ranking. In Proc. MT Summit X. find
14. Michael White and Rajakrishnan Rajkumar. 2009. Perceptron reranking for CCG realization. In Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 410–419, Singapore, August. Association for Computational Linguistics. find
15. Michael White. 2006. Efficient Realization of Coordinate Structures in Combinatory Categorial Grammar. Research on Language and Computation, 4(1):39–75. find

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/rajkumar-2011-linguistically-motivated-complementizer-choice-surface/