[Corpora-List] Tag sets

maxwell at umiacs.umd.edu maxwell at umiacs.umd.edu
Tue Feb 5 18:36:38 CET 2008


We're looking at annotating a small sample (~5k words) of Bengali text, and later maybe Urdu and Punjabi. The annotation will be the dictionary citation form of each word. The texts are mostly news articles, so there are a fair number of words for which there won't be any dictionary citation form. These include many proper names, numerals, acronyms, and who knows what else. I'll refer to these as "non-dictionary words", whereas "dictionary words" will include words whose citation form is in the dictionary we're using, even if the inflected wordform itself is not. (We're doing this to test a morphological parser.)

This is not quite the same as the inverse of named entity tagging, since some parts of names may have citation forms. For example, in English one would tag "Mississippi River" as a name. But "River" would be found in the dictionary, so for our purposes we would only want to tag "Mississippi" as a non-dictionary word.

The simplest thing for us to do would be to just tag all such non-dictionary words the same way, e.g. with a tag "NOT". However, in the interest of future uses to which we might put such a tagged text, it might be good to differentiate among the various kinds of non-dictionary words.

We could easily make up our own tagset for non-dictionary words, but it strikes me that better would be to use some standard tagset for such words, if such a tagset exists. There is a table of tagsets in Manning and Schutze pg. 141-2, including the Penn Treebank, Brown, and CLAWS. However, the tagsets are English-specific. This is especially noticeable in the punctuation tags for the PTB and Brown sets, but also e.g. in the decision to tag singular and plural proper nouns differently. (Some languages attach case markers to proper nouns.) Also, it appears that none of the tagsets distinguishes between numerals ('3', '4.5') and numbers written out ('three', 'four point five'), which we need to do, nor are acronyms distinguished from "symbols".

Another distinction I thought about making is between "ordinary" Bengali names, and foreign names, since one might later want to develop a transducer to convert the latter into their more common Latin forms. However, I suspect that might be too difficult a distinction for annotators to make, and in any case some well-known Bengali names are likely to have "standard" transliterations.

Does anyone know of a semi-standard tagset that would be less English-specific, and would make the kinds of distinctions among non-dictionary words that we want to (or should) make? Or should we just make up our own set?

Mike Maxwell

CASL/ U MD



More information about the Corpora mailing list