Sentence Boundary Detection and the Problem with the U.S.

What it is about

Sentence boundary detection is hard because the period is ambiguous, and most practical tools rely on hand-made abbreviation lists. The author builds a fully statistical system. For each period, the features look only at one token on each side of it, and a linear support vector machine classifies the boundary. On a standard news corpus of about 27,000 examples the system makes 67 errors, which is an error rate of about 0.25%. Of these, 23 involve the word "U.S.". The single feature of the token that ends with the period gives surprisingly good results on its own, and a simple Naive Bayes classifier is also good, so the choice of classifier matters less than the features.

Ideas

• I1. en/src/tok.rs: a period classifier over the token before the period and the token after it, with binary features and a linear score written in Rust. Measure the error count on EWT test per candidate period, and report the errors on "U.S.", "Mr." and similar tokens separately.
• I2. Test the paper's ablation first: keep only the feature of the token ending with the period, then add the next-token feature. Count the errors each step removes on EWT test, to see which feature carries the weight in English.
• I3. Store the abbreviation decisions as a table of learned weights, not a hand list, and print for each decision the feature that drove it, so every boundary can be explained.

For Mova

The code is en/src/tok.rs, and the metric is en tok-eval. The chapter is 07. Punctuation, capital letters, abbreviations and numbers. The paper's corpus is news text, so the 0.25% figure is not a Mova figure. The paper's data is not in Mova's build. Measure first on EWT test: errors per candidate period, with a two-fold check (en folds).

Cite

The paper itself, not this card. BibTeX from the library's record of the work. DOI 10.3115/1620853.1620920

@inproceedings{gillick2009sentence,
  title = {{Sentence Boundary Detection and the Problem with the U.S.}},
  author = {Gillick, Dan},
  booktitle = {Proceedings of Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the Association for Computational Linguistics, Companion Volume: Short Papers},
  year = {2009},
  doi = {10.3115/1620853.1620920},
  url = {https://doi.org/10.3115/1620853.1620920}
}

References

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

1. J. Aberdeen, J. Burger, D. Day, L. Hirschman, P. Robinson, and M. Vilain. 1995. MITRE: description of the Alembic system used for MUC-6. In Proceedings of the 6th conference on Message understanding, pages 141–155. Association for Computational Linguistics Morristown, NJ, USA. find
2. T. Joachims. 1999. Making large-scale support vector machine learning practical, Advances in kernel methods: support vector learning. find
3. T. Kiss and J. Strunk. 2006. Unsupervised Multilingual Sentence Boundary Detection. Computational Linguistics, 32(4):485–525. find our card
4. E. Loper and S. Bird. 2002. NLTK: The Natural Language Toolkit. In Proceedings of the ACL Workshop on Effective Tools and Methodologies for Teaching Natural Language Processing and Computational Linguistics, pages 62–69. find
5. M.P. Marcus, M.A. Marcinkiewicz, and B. Santorini. 1993. Building a large annotated corpus of English: the penn treebank. Computational Linguistics, 19(2):313–330. find our card
6. A. Mikheev. 2002. Periods, Capitalized Words, etc. Computational Linguistics, 28(3):289–318. find
7. D.D. Palmer and M.A. Hearst. 1997. Adaptive Multilingual Sentence Boundary Disambiguation. Computational Linguistics, 23(2):241–267. find
8. J.C. Reynar and A. Ratnaparkhi. 1997. A maximum entropy approach to identifying sentence boundaries. In Proceedings of the Fifth Conference on Applied Natural Language Processing, pages 16–19. find

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/aclbulk-2009-gillick-sentence-boundary-detection-problem-u/