Elamur Research⌕
☰
home / papers / 2009 / Learning Semantic Correspondences with Less Supervision

Learning Semantic Correspondences with Less Supervision

Summary

The paper addresses grounded language acquisition with very weak supervision. The input is a world state, stored as a set of typed records, and a raw, unsegmented text that describes it. The system must work out how the text splits into utterances, which records and fields are mentioned, and how text pieces align to them. The authors propose one generative model, a hierarchical hidden semi-Markov model trained with EM. It picks a sequence of records, then fields within each record, then emits word spans. Integer, string and categorical values are verbalized in type-specific ways, and a NULL record covers ungrounded text. Evaluation spans three increasingly hard domains: Robocup sportscasting, a newly built weather forecast dataset, and NFL recaps. On Robocup the model beats KRISPER (76.5% vs. 67% F1). As the first unsupervised approach on NFL, it reaches 53.2 F1, closing over half the gap between a heuristic baseline and supervised systems. This shows that detailed text–data alignments can be learned without sentence-level meaning pairs.

Key points

  • Datasets (average scenario size in Table 1):
    • Robocup: 1919 scenarios, about 5.7 words and 2.4 records each.
    • Weather: 22,146 scenarios, about 28.7 words, 36 records and 12 record types each.
    • NFL: 78 annotated games, about 969 words and 329 records each.
  • Weather corpus construction: forecasts were collected from www.weather.gov for 3,753 US cities with at least 10,000 inhabitants over February 7–9, 2009. Each city and date yielded separate day and night scenarios. Records aggregate hourly measurements over time intervals.
  • Three models of increasing structure:
    • Model 1 learns only word–field associations.
    • Model 2 adds segmentation.
    • Model 3 adds Markov transitions between record types, including START and STOP types.
  • NFL results (F1):
    • Random: 12.0.
    • Number/string-matching heuristic: 26.7.
    • Model 1: 32.8.
    • Model 2’: 39.9.
    • Model 2’ with record groups: 53.2.
    • Supervised systems from Snyder and Barzilay (2007): 68.6 and 80.3.
  • NFL candidate pruning: records sharing no numbers or name-like words with a sentence were discarded. Only 1.5% of candidates remained, while the oracle alignment F1 stayed at 88.7.

Interesting ideas

  • I1. Type-specific rendering of field values: numbers may appear slightly perturbed, strings may appear partially, and categorical values get learned word distributions. This gives cheap "anchor points" for alignment and suits any data-to-text task over databases, such as statistics or trade tables.
  • I2. Staged training: the authors first train the simple lexical model, then use its posteriors to initialize richer models. The complex models need this to avoid poor local optima, and the same curriculum trick applies to other latent-structure models trained with EM.
  • I3. A dedicated NULL record: a special record generates text that refers to nothing in the database. This lets the model tolerate noisy, partially grounded text, which is useful whenever only part of a document is supported by structured data.
  • I4. Parallel model copies with shared segmentation: to handle redundant records (play vs. score), records are split into groups, each group gets its own model copy, and all copies share the same segmentation. A coupling potential captures which records co-occur. This is a neat way to allow multi-label alignment without redesigning the core model.
  • I5. Cheap lexical pruning before inference: candidates are filtered by shared numbers and capitalized names, and the oracle score is checked to confirm little is lost. This makes huge candidate sets tractable while measuring the cost of the pruning.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

Fields: Natural language processing

Cited by

158 papers of our library cite it (of the 140884 whose bibliographies we read); 16 of them have a page here, 142 not yet.

Cite

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

@inproceedings{liang2009learning,
  title = {{Learning Semantic Correspondences with Less Supervision}},
  author = {Liang, Percy and Jordan, Michael and Klein, Dan},
  booktitle = {Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language Processing of the AFNLP},
  year = {2009},
  doi = {10.3115/1687878.1687893},
  url = {https://doi.org/10.3115/1687878.1687893}
}

References

20 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. R. Barzilay and M. Lapata. 2005. Collective content selection for concept-to-text generation. In Human Language Technology and Empirical Methods in Natural Language Processing (HLT/EMNLP), pages 331–338, Vancouver, B.C. find
  2. R. Barzilay and M. Lapata. 2008. Modeling local coherence: An entity-based approach. Computational Linguistics, 34:1–34. find our card
  3. P. F. Brown, S. A. D. Pietra, V. J. D. Pietra, and R. L. Mercer. 1993. The mathematics of statistical machine translation: Parameter estimation. Computational Linguistics, 19:263–311. find
  4. D. L. Chen and R. J. Mooney. 2008. Learning to sportscast: A test of grounded language acquisition. In International Conference on Machine Learning (ICML), pages 128– 135. Omnipress. find
  5. J. DeNero, A. Bouchard-Côté, and D. Klein. 2008. Sampling alignment structure under a Bayesian translation model. In Empirical Methods in Natural Language Processing (EMNLP), pages 314–323, Honolulu, HI. find
  6. J. Eisenstein and R. Barzilay. 2008. Bayesian unsupervised topic segmentation. In Empirical Methods in Natural Language Processing (EMNLP), pages 334–343. find
  7. J. Feldman and S. Narayanan. 2004. Embodied meaning in a neural theory of language. Brain and Language, 89:385– 392. find
  8. R. Ge and R. J. Mooney. 2005. A statistical semantic parser that integrates syntax and semantics. In Computational Natural Language Learning (CoNLL), pages 9–16, Ann Arbor, Michigan. find
  9. P. Gorniak and D. Roy. 2007. Situated language understanding as filtering perceived affordances. Cognitive Science, 31:197–231. find
  10. T. Grenager, D. Klein, and C. D. Manning. 2005. Unsupervised learning of field segmentation models for information extraction. In Association for Computational Linguistics (ACL), pages 371–378, Ann Arbor, Michigan. Association for Computational Linguistics. find
  11. R. J. Kate and R. J. Mooney. 2007. Learning language semantics from ambiguous supervision. In Association for the Advancement of Artificial Intelligence (AAAI), pages 895–900, Cambridge, MA. MIT Press. find
  12. P. Liang, B. Taskar, and D. Klein. 2006. Alignment by agreement. In North American Association for Computational Linguistics (NAACL), pages 104–111, New York City. Association for Computational Linguistics. find our card
  13. W. Lu, H. T. Ng, W. S. Lee, and L. S. Zettlemoyer. 2008. A generative model for parsing natural language to meaning representations. In Empirical Methods in Natural Language Processing (EMNLP), pages 783–792. find
  14. R. C. Moore. 2004. Improving IBM word alignment model 1. In Association for Computational Linguistics (ACL), pages 518–525, Barcelona, Spain. Association for Computational Linguistics. find
  15. H. Ney and S. Vogel. 1996. HMM-based word alignment in statistical translation. In International Conference on Computational Linguistics (COLING), pages 836–841. Association for Computational Linguistics. find
  16. J. M. Siskind. 1996. A computational study of cross-situational techniques for learning word-to-meaning mappings. Cognition, 61:1–38. find
  17. B. Snyder and R. Barzilay. 2007. Database-text alignment via structured multilabel classification. In International Joint Conference on Artificial Intelligence (IJCAI), pages 1713–1718, Hyderabad, India. find
  18. C. Yu and D. H. Ballard. 2004. On the integration of grounding language and learning objects. In Association for the Advancement of Artificial Intelligence (AAAI), pages 488– 493, Cambridge, MA. MIT Press. find
  19. L. S. Zettlemoyer and M. Collins. 2005. Learning to map sentences to logical form: Structured classification with probabilistic categorial grammars. In Uncertainty in Artificial Intelligence (UAI), pages 658–666. find
  20. L. S. Zettlemoyer and M. Collins. 2007. Online learning of relaxed CCG grammars for parsing to logical form. In Empirical Methods in Natural Language Processing and Computational Natural Language Learning (EMNLP/CoNLL), pages 678–687. find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/liang-2009-learning-semantic-correspondences-less-supervision/