The Penn Treebank

The University of Pennsylvania annotated more than 4.5 million words of American English for part of speech, over half of them also for skeletal syntactic structure, and made the corpus available through the Linguistic Data Consortium.

Why it matters

The field got a shared ruler: different methods could be compared on the same data and the same number for the first time.

By the report's table as of November 1992, the annotated material was Department of Energy abstracts, Dow Jones Newswire stories, Department of Agriculture bulletins, Library of America texts, MUC-3 messages, sentences from IBM manuals, WBUR radio transcripts, ATIS sentences and a retagged Brown Corpus. The corpus became the basis both for training and for evaluation: parser improvements started being measured as a percentage on this set. The effect ran wider than the corpus itself; it set the habit of measuring progress in language by a number on a shared set rather than by a demonstration.

Event record

Event date

June 1993

Timeline date

Event date

Verification

Sources gathered automatically · September 17, 2026

Lines

ID

evt-0169

Computational Linguistics volume 19, issue 2, June 1993.

Sources

primary Building a Large Annotated Corpus of English: The Penn Treebank
Computational Linguistics · Published June 1993
archive Building a Large Annotated Corpus of English: The Penn Treebank (technical report MS-CIS-93-87)
University of Pennsylvania, Department of Computer and Information Science (ScholarlyCommons) · Published October 1993

Related events

Earlier

Related The TIMIT corpus 1990links to this record
The same move as the Penn Treebank: a shared annotated set as the condition for comparison.
Related Statistical machine translation 1988
Both works move language processing from rules onto data.
Enables Statistical part-of-speech tagging 1988links to this record
The Penn Treebank's authors write that in the project's early stages the initial automatic part-of-speech assignment came from Church's PARTS program, whose output the annotators corrected.
- Building a Large Annotated Corpus of English: The Penn Treebank · Computational Linguistics

Later

Builds on Lexicalised statistical parsing 1997links to this record
The parser was trained on sections 02-21 of the Wall Street Journal part of the Penn Treebank and tested on section 23.
- Three Generative, Lexicalised Models for Statistical Parsing (Michael Collins), ACL/EACL 1997, pp. 16-23 · Association for Computational Linguistics
Related LeNet-5 and MNIST 1998links to this record
The same principle: a shared labelled set makes comparison possible.
Builds on CoNLL-2000: eleven systems on one split 2000links to this record
The task data was extracted from the parsed Wall Street Journal portion of Penn Treebank II: sections 15 to 18 for training, section 20 for testing.
- Introduction to the CoNLL-2000 Shared Task: Chunking · Association for Computational Linguistics
Related Data outweighs the algorithm 2001links to this record
The experiment became possible because large corpora already existed.
Related ImageNet 2009
The same move ImageNet later repeated: a shared labelled set as the condition for progress.