This repository provides labeled data for training homograph disambiguation models, as described in:

Gorman, K., Mazovetskiy, G., and Nikolaev, V. (2018). Improving homograph disambiguation with machine learning. In Proceedings of LREC, 1349-1352.

If you use this data in a publication, we would appreciate if you cite this paper.

Sentences were extracted from English Wikipedia articles. Homograph were initially labeled for the most likely WORDID (as defined below) in context by a team of three annotators. In the case that all three did not agree on the WORDID, a fourth senior annotator resolved the disagreements.

There are now 162 unique homographs and roughly 100 examples per homograph.

The files in the directories data/train and data/eval are TSV files with the following fields:

• homograph: the homograph word itself
• wordid: name of the pronunciation
• sentence: text of the example
• start: the first byte---inclusive--of the target homograph in sentence
• end: the last byte---exclusive---of the target homograph in sentence

These two files represent a suggested 90%/10% train/test split stratified by homograph.

The file data/wordids.tsv is a TSV file which maps from the WORDID field above to information used by the annotator: -a short human-readable description of the WORDID, and a transcription of the WORDID. Note that neither are intended to be authoritative; they are simply to help users distinguish between the various WORDIDs for a homograph. The final two fields have some impressionistic taxonomic information about the nature of the homography itself intended for use during error analysis. The following fields are present:

• homograph: the homograph word itself
• wordid: name of the pronunciation
• label: a short human-readable description of the wordid
• pronunciation: a phonemic transcription of the wordid in US English.
• homograph_type: a binary category describing the broad source of homography: morphosyntactic derivations from the same lemma, or lexically distinct terms.
• fine_homography_type: a more detailed classification of the above.

This data was collected by Kyle Gorman, Vitaly Nikolaev, and Gleb Mazovetskiy, with help from a team of linguists and annotators.

See LICENSE.

See CONTRIBUTING.

This is not an official Google product.