Improved Sentence Boundary Detection Dan Gillick January 21, 2009 ----- Consider the following text: "On Jan. 20, former Sen. Barack Obama became the 44th President of the U.S. Millions attended the Inauguration." The periods are potentially ambiguous, signifying either the end of a sentence, an abbreviation, or both. The sentence boundary detection (SBD) task involves disambiguating the periods, and in particular, classifying each period as end-of-sentence (<S>) or not. In the example, only the period at the end of U.S. should be classified as <S>: "On Jan. 20, former Sen. Barack Obama became the 44th President of the U.S.<S> Millions attended the Inauguration." Chances are, if you are using some SBD system, it has an error rate of 1%-3% on English newswire text. The system described here achieves the best known error rate on the Wall Street Journal corpus: 0.25% and comparable error rates on the Brown corpus (mixed genre) and other test corpora. ----- SBD is fundamental to many natural language processing problems, but only a few papers describe solutions. A variety of rule-based systems are floating around, and a few semi-statistical systems are available if you know where to look. The most widely cited are: - Alembic (Aberdeen, et al. 1995): Abbreviation list and ~100 hand-crafted regular expressions. - Satz (Palmer & Hearst at Berkeley, 1997): Part of speech features and abbreviation lists as input to a classifier (neural nets and decision trees have similar performance). - mxTerminator (Reynar & Ratnaparkhi, 1997): Maximum entropy classification with simple lexical features. - Mikheev (Mikheev, 2002): Observes that perfect labels for abbreviations and names gives almost perfect SBD results. Creates heuristics for marking these, unsupervised, from held-out data. - Punkt (Strunk and Kiss, 2006): Unsupervised method uses heuristics to identify abbreviations and sentence starters. I have not been able to find publicly available copies of any of these systems, with the exception of Punkt, which ships with NLTK. Nonetheless, here are some error rates reported on what I believe to be the same subset of the WSJ corpus (sections 03-16). - Alembic: 0.9% - Satz: 1.5%; 1.0% with extra hand-written lists of abbreviations and non-names. - mxTerminator: 2.0%; 1.2% with extra abbreviation list. - Mikheev: 1.4%; 0.45% with abbreviation list (assembled automatically but carefully tuned; test-set-dependent parameters are a concern) - Punkt: 1.65% (Though if you use the model that ships with NLTK, you'll get over 3%) All of these systems use lists of abbreviations in some capacity, which I think is a mistake. Some abbreviations almost never end a sentence (Mr.), which makes list-building appealing. But many abbreviations are more ambiguous (U.S., U.N.), which complicates the decision. ----- While 1%-3% is a low error rate, this is often not good enough. In automatic document summarization, for example, including a sentence fragment usually renders the resulting summary unintelligible. With 10-sentence summaries, 1 in 10 is ruined by an SBD system with 99% accuracy. Improving the accuracy to 99.75%, only 1 in 40 is ruined. Improved sentence boundary detection is also likely to help with language modeling and text alignment. ----- I built a supervised system that classifies sentence boundaries without any heuristics or hand-generated lists. It uses the same training data as mxTerminator, and allows for Naive Bayes or SVM models (SVM Light). ---------------------------------------------------------- Corpus SVM Naive Bayes ---------------------------------------------------------- WSJ 0.25% 0.35% Brown 0.36% 0.45% Complete Works of Edgar Allen Poe: 0.52% 0.44% ---------------------------------------------------------- I've packaged this code, written in Python, for general use. Word-level tokenization, which is particularly important for good sentence boundary detection, is included. Note that the included models use all of the labeled data listed here, meaning that the expected results are somewhat better than the numbers reported above. Including the Brown data as training improves the WSJ result to 0.22% and the Poe result to 0.37 (using the SVM). ----- A few other notes on performance. The standard WSJ test corpus includes 26977 possible sentence boundaries. About 70% are in fact sentence boundaries. Classification with the included SVM model will give 59 errors. Of these, 24 (41%) involve the word "U.S.", a particularly interesting case. In training, "U.S." appears 2029 times, and 90 of these are sentence boundaries. Further complicating the situation, "U.S." often appears in a context like "U.S. Security Council" or "U.S. Government", and either "Security" or "Government" are viable sentence starters. Other confusing cases include "U.N.", "U.K.", and state abbreviations like "N.Y." which have similar characteristics as "U.S." but appear somewhat less frequently. ----- Setup: (1) You need Python 2.5 or later. Python 3 does not seem to work. (2) To use SVM models, you'll need SVM Light (http://svmlight.joachims.org/) - once installed, you'll need to modify sbd.py slightly: at the top, change the paths to SVM_LEARN and SVM_CLASSIFY to point to the files you've installed. ----- Example calls: (show command line options) python sbd.py -h (split sentences in sample.txt using the provided Naive Bayes model) python sbd.py -m model_nb sample.txt (now using the provided SVM model) python sbd.py -m model_svm sample.txt (now keeping tokenized output) python sbd.py -m model_nb -t sample.txt (now writing output to sample.sent) python sbd.py -m model_nb -t sample.txt -o sample.sent ----- Note about SVM_LIGHT: The provided SVM model was built with SVM_LIGHT version 6.02. It seems that SVM_CLASSIFY requires a matching version or it will crash. So, you can either try to use version 6.02, or you can make the following quick fix: open model_svm/svm_model change the first line from: SVM-light Version V6.02 to whatever your version is. ----- Dan Gillick January 21, 2009 Berkeley, California dgillick@cs.berkeley.edu