Link
Search

Caseless models

Table of contents

If your text is all lowercase, all uppercase, or badly and inconsistently capitalized (many web forums, texts, twitter, etc.) then this will negatively effect the performance of most of our annotators. Most of our annotators were trained on data that is standardly edited and capitalized full sentences.

There are two strategies available to address this that may help. One is to try to first correctly capitalize the text with a truecaser, and then to process the text with the standard models. See the TrueCaseAnnotator for how to do this.

An example command for using regular annotators following truecasing is:

java edu.stanford.nlp.pipeline.StanfordCoreNLP -annotators tokenize,truecase,pos,lemma,ner,depparse -truecase.overwriteText true -file caseless.txt -outputFormat json

The other strategy is to use models more suited to ill-capitalized text.

The GATE folk made an English POS tagger model trained on twitter text. You can get it from the extensions page.

We have made slightly different Stanford CoreNLP models for the tagger, parser, and NER that ignore capitalization. We have only trained such models for English, but the same method could be used for other languages.

To use these models, you need to download a jar file with caseless models. Prior to version 3.6, caseless models were packaged separately as their own jar file (approximately treating “caseless English” like a separate language). Starting with version 3.6, caseless models for English are included in the new comprehensive english jar file. You can find these jar files on the Release history page.

Be sure to include the path to the case-insensitive models jar in the -cp classpath flag and then you can ask for these models to be used like this:

-pos.model edu/stanford/nlp/models/pos-tagger/english-caseless-left3words-distsim.tagger
-parse.model edu/stanford/nlp/models/lexparser/englishPCFG.caseless.ser.gz
-ner.model edu/stanford/nlp/models/ner/english.all.3class.caseless.distsim.crf.ser.gz,edu/stanford/nlp/models/ner/english.muc.7class.caseless.distsim.crf.ser.gz,edu/stanford/nlp/models/ner/english.conll.4class.caseless.distsim.crf.ser.gz

An example

Here is an example text:

% cat lakers.txt
lonzo ball talked about kobe bryant after the lakers game.

With the default English models, no entities (and no proper nouns) are found:

% java edu.stanford.nlp.pipeline.StanfordCoreNLP -file lakers.txt -outputFormat conll -annotators tokenize,pos,lemma,ner
% cat lakers.txt.conll 
1	lonzo	lonzo	NN	O	_	_
2	ball	ball	NN	O	_	_
3	talked	talk	VBD	O	_	_
4	about	about	IN	O	_	_
5	kobe	kobe	NN	O	_	_
6	bryant	bryant	NN	O	_	_
7	after	after	IN	O	_	_
8	the	the	DT	O	_	_
9	lakers	laker	NNS	O	_	_
10	game	game	NN	O	_	_
11	.	.	.	O	_	_

However, if we ask to use the caseless models, then we’re doing pretty well: All the name words are now recognized as proper nouns, and the two person names are recognized. However, the team name is still missed. Correct named entity recognition is just harder for caseless English than for well-edited English!

% java edu.stanford.nlp.pipeline.StanfordCoreNLP -outputFormat conll -annotators tokenize,pos,lemma,ner -file lakers.txt -pos.model edu/stanford/nlp/models/pos-tagger/english-caseless-left3words-distsim.tagger -ner.model edu/stanford/nlp/models/ner/english.all.3class.caseless.distsim.crf.ser.gz,edu/stanford/nlp/models/ner/english.muc.7class.caseless.distsim.crf.ser.gz,edu/stanford/nlp/models/ner/english.conll.4class.caseless.distsim.crf.ser.gz
% cat lakers.txt.conll 
1	lonzo	lonzo	NNP	PERSON	_	_
2	ball	ball	NNP	PERSON	_	_
3	talked	talk	VBD	O	_	_
4	about	about	IN	O	_	_
5	kobe	kobe	NNP	PERSON	_	_
6	bryant	bryant	NNP	PERSON	_	_
7	after	after	IN	O	_	_
8	the	the	DT	O	_	_
9	lakers	lakers	NNPS	O	_	_
10	game	game	NN	O	_	_
11	.	.	.	O	_	_

Training caseless models

To train your own caseless models, you need one additional property, which asks for a function to be called before a token is processed which leads to the case of all words being ignored. We use by default a function that also Americanizes the spelling of certain words:

wordFunction = edu.stanford.nlp.process.LowercaseAndAmericanizeFunction

but there is also simply:

wordFunction = edu.stanford.nlp.process.LowercaseFunction

A note on the version 3.6 model

The caseless NER model edu/stanford/nlp/models/ner/english.all.3class.caseless.distsim.crf.ser.gz released with version 3.6.0 was defective and has very poor performance. Sorry! Stuff happens. If you want good caseless NER, you should either run with caseless models from a 3.5.x series release (all of which are compatible with version 3.6.0) or download the new fixed model from version 3.7.0 or a later version. Since the version 3.5.x releases have a separate caseless jar, it is easy to also specify an additional jar as a dependency; you just have to make sure that it appears on your classpath before other jars which contain models with the same name.