Elamur Research⌕
☰
home / papers / 2013 / Overcoming the Lack of Parallel Data in Sentence Compression

Overcoming the Lack of Parallel Data in Sentence Compression

Paper · 2013 · Katja Filippova and Yasemin Altun

Summary

Supervised deletion-based sentence compression has long been limited by a lack of training data. Existing parallel corpora such as Ziff-Davis and the two Edinburgh sets hold only about a thousand pairs each, which makes rich or lexicalized features impractical. Filippova and Altun propose an automatic way to harvest a much larger corpus from news on the web. They pair a news headline with the article's first sentence and extract a compression whose dependency tree is a subtree of the source tree. This structural property makes the data usable by tree-pruning compressors. Human raters judged the harvested pairs readable and informative.

The authors also turn the unsupervised ILP dependency-pruning method of Filippova & Strube (2008) into a supervised system. Its hand-built edge weights are replaced by a linear function over many features, learned with structured prediction following Collins (2002). The new model clearly beats the unsupervised baseline in both human and automatic evaluations. It also shows that sparse lexical features help, but only when a large corpus is available to train them.

Key points

  • The automatically built corpus has 250,000 sentence–compression pairs, and each compression is a dependency subtree of its source sentence.

  • The framework prunes edges of a transformed dependency graph with ILP, under tree-structure and character-length constraints. The baseline weights edges by syntactic importance times "headline-likeness," with statistics collected from 9M news articles.

  • In the automatic F1 evaluation on 1,000 test tuples, the systems scored as follows:

    SystemF1
    Unsupervised baseline52.3
    Supervised, syntactic and informativeness features only75.0
    Supervised, 330 non-lexical features79.6
    Supervised, all 27,813 features84.3

  • Training the full-feature model on only 10K tuples lowered F1 to 81.4. Even 10K tuples is more data than all previous compression corpora combined.

  • The test data is demanding: source sentences average 165 characters and the compression rate is about 0.46. Parser errors were the main cause of readability problems.

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

 Institutions

🏢GoogleMountain View · United States353 works · 364,806 citations

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 9,086 worksArtificial Intelligence · 8,889 worksArts and Humanities · 9,296 worksLanguage and Linguistics · 9,206 works

Cite

Cite as Katja Filippova and Yasemin Altun. 2013. Overcoming the Lack of Parallel Data in Sentence Compression. Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing.

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

@inproceedings{filippova2013overcoming,
  title = {{Overcoming the Lack of Parallel Data in Sentence Compression}},
  author = {Katja Filippova and Yasemin Altun},
  booktitle = {Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing},
  year = {2013},
  doi = {10.18653/v1/d13-1155},
  url = {https://doi.org/10.18653/v1/d13-1155}
}

References

10 references, 5 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Bejan, C. & S. Harabagiu (2010). Unsupervised event coreference resolution with rich linguistic features. In Proc. of ACL-10, pp. 1412–1422.
  2. Belz, A., M. White, D. Espinosa, E. Kow, D. Hogan & A. Stent (2011). The first surface realization shared task: Overview and evaluation results. In Proc. of ENLG-11, pp. 217–226. find
  3. Berg-Kirkpatrick, T., D. Gillick & D. Klein (2011). Jointly learning to extract and compress. In Proc. of ACL-11.
  4. Clarke, J. & M. Lapata (2006). Models for sentence compression: A comparison across domains, training requirements and evaluation measures. In Proc. of COLING-ACL-06, pp. 377–385. find
  5. Clarke, J. & M. Lapata (2008). Global inference for sentence compression: An integer linear programming approach. Journal of Artificial Intelligence Research, 31:399–429. find
  6. Cohn, T. & M. Lapata (2009). Sentence compression as tree transduction. Journal of Artificial Intelligence Research, 34:637–674. find
  7. Collins, M. (2002). Discriminative training methods for Hidden Markov Models: Theory and experiments with perceptron algorithms. In Proc. of EMNLP-02, pp. 1–8. de Marneffe, M.-C., B. MacCartney & C. D. Manning (2006). Generating typed dependency parses from phrase structure parses. In Proc. of LREC-…
  8. Dolan, B., C. Quirk & C. Brokett (2004). Unsupervised construction of large paraphrase corpora: Exploiting massively parallel news sources. In Proceedings of the 20th International Conference on Computational Linguistics, Geneva, Switzerland, 23–27 August 2004, pp. 350–356.
  9. Dorr, B., D. Zajic & R. Schwartz (2003). Hedge trimmer: A parse-and-trim approach to headline generation. In Proceedings of the Text Summarization Workshop at HLT-NAACL-03, Edmonton, Alberta, Canada, 2003, pp. 1–8.
  10. S H H* U O S H H* U O S H H* U O S find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/filippova-2013-overcoming-lack-parallel-data-sentence/