Cross-linguistic Attribute Selection for REG: Comparing Dutch and English

What it is about

In referring-expression generation (REG) the task is to choose a set of attributes that picks out one target object among distractors ("the red sofa"). The paper asks whether this attribute-selection step can be learned in one language and used in another. It uses the graph-based algorithm of Krahmer et al. (2003): a scene is a labelled graph, and the algorithm searches for the cheapest subgraph that identifies the target. Costs and the search order come from attribute frequencies in human descriptions. Eight cost functions are compared: stochastic costs (from -log2 of attribute probability) and coarser free-naive costs (free, 1 or 2), each trained on English (TUNA 2008) and Dutch (D-TUNA), for two domains (furniture and people). Each function is tested on 38 or 40 items per domain. Free-naive costs do better than stochastic ones overall (Dice F(1,76) = 34.85, p < .001). Costs trained in one language give almost the same results on the other language's data, which the authors take as evidence that attribute selection is largely language independent.

Ideas

• I1. Attribute selection over the world: for an entity table in world, search for the cheapest set of properties that picks out one entity, with costs from counts. This gives a deterministic answer to "which one?" questions and a short description for the referent.
• I2. Free-naive costs as a table: three cost levels from counts (free, 1, 2), stored in a BTreeMap. Easy to explain: each cost is a threshold on how often an attribute appears in the training descriptions.
• I3. Search order as a separate, explicit list. The paper shows that free attributes are only used if they are tried before the search stops, so the order must be fixed and tested, not left to chance.
• I4. Cross-domain test: derive costs from one English domain (people) and test on another (furniture), and check whether the result matches the within-domain score. This is the paper's cross-language design, adapted to English domains; the negative control is a shuffled attribute order, which must score worse.

For Mova

The chapters are 12. Pronouns, determiners and numerals (the definite article and "which one" determiners), 15. Adjectives and adverbs (attribute adjectives and their order) and 22. Phrases: noun, adjective, adverb and prepositional phrases (the modifier order inside a noun phrase). The code is coref/src/mention.rs (definite descriptions), world/src/qa.rs (which-questions) and world/src/knowledge.rs (entities with properties). The measures: Dice and exact-match accuracy of the chosen attribute set against human or gold answers on held-out English items, run twice for identical output, with the shuffled-order negative control. The cost data are evaluation-only, so Mova should use its own counts from its own English texts.

Cite

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

@inproceedings{theune2010cross,
  title = {{Cross-linguistic Attribute Selection for REG: Comparing Dutch and English}},
  author = {Theune, Mariët and Koolen, Ruud and Krahmer, Emiel},
  booktitle = {Proceedings of the 6th International Natural Language Generation Conference},
  year = {2010},
  url = {https://aclanthology.org/W10-4221/}
}

References

10 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. R. Dale and E. Reiter. 1995. Computational interpretation of the Gricean maxims in the generation of referring expressions. Cognitive Science, 19(2):233– 263. find
2. A. Gatt, I. van der Sluis, and K. van Deemter. 2007. Evaluating algorithms for the generation of referring expressions using a balanced corpus. In Proceedings of the 11th European Workshop on Natural Language Generation (ENLG 2007), pages 49–56. find
3. A. Gatt, A. Belz, and E. Kow. 2008. find
4. The TUNA Challenge 2008: Overview and evaluation results. In Proceedings of the 5th International Natural Language Generation Conference (INLG 2008), pages 198–206. find
5. R. Koolen and E. Krahmer. 2010. The D-TUNA corpus: A Dutch dataset for the evaluation of referring expression generation algorithms. In Proceedings of the 7th international conference on Language Resources and Evaluation (LREC 2010). find
6. R. Koolen, A. Gatt, M. Goudbeek, and E. Krahmer. 2010. Overspecification in referring expressions: Causal factors and language differences. Submitted. find
7. E. Krahmer, S. van Erk, and A. Verleg. 2003. Graph-based generation of referring expressions. Computational Linguistics, 29(1):53–72. find
8. E. Krahmer, M. Theune, J. Viethen, and I. Hendrickx. 2008. Graph: The costs of redundancy in referring expressions. In Proceedings of the 5th International Natural Language Generation Conference (INLG 2008), pages 227–229. find
9. K. van Deemter, I. I. van der Sluis, and A. Gatt. 2006. Building a semantically transparent corpus for the generation of referring expressions. In Proceedings of the 4th International Natural Language Generation Conference (INLG 2006), pages 130–132. find
10. J. Viethen, R. Dale, E. Krahmer, M. Theune, and P. Touset. 2008. Controlling redundancy in referring expressions. In Proceedings of the Sixth International Conference on Language Resources and Evaluation (LREC 2008), pages 239–246. find

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/theune-2010-cross-linguistic-attribute-selection-reg/