NERDemo.javaincluded in the 2008-05-07 version of the software?
On Unix/Linux (or command-line Mac OS X), use GNU tar, if you're not already. (If you're using
Linux, you're almost certainly using GNU tar.) For some reason we don't
understand, it doesn't seem to unpack with classic Unix tar. Make sure you
-z option if you are not gunzipping it in
tar -xzf stanford-ner-2009-01-16.tgz.
On Windows, it unpacks fine with most common tools, such as WinZip or 7-Zip. The latter is open source. (As of Sep 2007, WinRAR doesn't work: it apparently does not handle tar files correctly.)
On the Mac, just double-click it to unpack. The default unarchiver (BOMArchiveHelper) works fine.
If it won't unpack, you normally have either a corrupted download (try downloading it again), a defective tool for unpacking tar files (normally evidenced by Unix file protection bits like 0664 appearing in file names), or there is some configuration error on your system, which we can't help with.
The documentation for training your own classifier is certainly
somewhere between bad and non-existent. But nevertheless, everything
you need is in the box, and you should
look through the Javadoc for at least the classes
Basically, the training data should be in tab-separated columns, and you define the meaning of those columns via a map. One column should be called "answer" and has the NER class, and existing features know about names like "word" and "tag". You define the data file, the map, and what features to generate via a properties file. There is considerable documentation of what features different properties generate in the Javadoc of NERFeatureFactory, though ultimately you have to go to the source code to answer some questions....
Here's a sample NER properties file:
trainFile = training-data.col serializeTo = ner-model.ser.gz map = word=0,answer=1 useClassFeature=true useWord=true useNGrams=true noMidNGrams=true maxNGramLeng=6 usePrev=true useNext=true useSequences=true usePrevSequences=true maxLeft=1 useTypeSeqs=true useTypeSeqs2=true useTypeySequences=true wordShape=chris2useLC useDisjunctive=true
Oh, okay. Here's an example. Suppose we want to build an NER system for Jane Austen novels. We might train it on chapter 1 of Emma. Download that file. You can convert it to one token per line with our tokenizer (included in the box) with the following command:
java -cp stanford-ner.jar edu.stanford.nlp.process.PTBTokenizer jane-austen-emma-ch1.txt > jane-austen-emma-ch1.tok
We then need to make training data where we label the entities.
There are various annotation tools available, or you could do this by
hand in a text editor. One way is to default to making everything an
other (for which the default label is "O" in our software, though you
can specify it via the
backgroundSymbol property) and then to
hand-label the real entities in a text editor. The first step can be
done with Perl using this command:
perl -ne 'chomp; print "$_\tO\n"' jane-austen-emma-ch1.tok > jane-austen-emma-ch1.tsv
and if you don't want to do the second, you can skip to downloading our input file. We have marked only one entity type, PERS for person name, but you could easily add a second entity type such as LOC for location, to this data.
You will then also want some test data to see how well the system is doing. You can download the text of chapter 2 of Emma and the gold standard annotated version of chapter 2.
Stanford NER CRF allows all properties to be specified on the command line, but it is easier to use a properties file. Here is a simple properties file (pretty much like the one above!), but explanations for each line are in comments, specified by "#":
# location of the training file trainFile = jane-austen-emma-ch1.tsv # location where you would like to save (serialize) your # classifier; adding .gz at the end automatically gzips the file, # making it smaller, and faster to load serializeTo = ner-model.ser.gz # structure of your training file; this tells the classifier that # the word is in column 0 and the correct answer is in column 1 map = word=0,answer=1 # This specifies the order of the CRF: order 1 means that features # apply at most to a class pair of previous class and current class # or current class and next class. maxLeft=1 # these are the features we'd like to train with # some are discussed below, the rest can be # understood by looking at NERFeatureFactory useClassFeature=true useWord=true # word character ngrams will be included up to length 6 as prefixes # and suffixes only useNGrams=true noMidNGrams=true maxNGramLeng=6 usePrev=true useNext=true useDisjunctive=true useSequences=true usePrevSequences=true # the last 4 properties deal with word shape features useTypeSeqs=true useTypeSeqs2=true useTypeySequences=true wordShape=chris2useLC
Here is that properties file as a downloadable link:
Once you have such a properties file, you can train a classifier with the command:
java -cp stanford-ner.jar edu.stanford.nlp.ie.crf.CRFClassifier -prop austen.prop
An NER model will then be serialized to the location specified in the
properties file (
once the program has completed. To check how well it works, you can run the test command:
java -cp stanford-ner.jar edu.stanford.nlp.ie.crf.CRFClassifier -loadClassifier ner-model.ser.gz -testFile jane-austen-emma-ch2.tsv
In the output, the first column is the input tokens, the second column is the correct (gold) answers, and the third column is the answer guessed by the classifier. By looking at the output, you can see that the classifier finds most of the person named entities but not all, mainly due to the very small size of the training data (but also this is a fairly basic feature set). The code then evaluates the performance of the classifier for entity level precision, recall, and F1. It gets 80.95% F1. (A commonly used script for NER evaluation is the Perl script conlleval, but either it needs adapted or else the raw IO input format used here needs to be mapped to IOB encoding for it to work correctly and give the same answer.)
So how do you apply this to make your own non-example NER model? You need 1) a training data source, 2) a properties file specifying the features you want to use, and (optional, but nice) 3) a test file to see how you're doing. For the training data source, you need each word to be on a separate line and annotated with the correct answer; all columns must be tab-separated. If you want to explicitly specify more features for the word, you can add these in the file in a new column and then put the appropriate structure of your file in the map line in the properties file. For example, if you added a third column to your data with a new feature, you might write "map= word=0, answer=1, mySpecialFeature=2".
Right now, most arbitrarily named features (like mySpecialFeature)
will not work without making modifications to the source code.
To see which features can already be attached to a CoreLabel, look
edu.stanford.nlp.ling.AnnotationLookup. There is a
table which creates a mapping between key and annotation type. For
example, if you search in this file for
will see that
lemma produces a
LemmaAnnotation. If you have added a new annotation, you
can add its type to this table, or you can use one of the known names
that already work, like
tag, lemma, chunk, web.
If you modify AnnotationLookup, you need to read the data from the column, translate it to the desired object type, and attach it to the CoreLabel using a CoreAnnotation. Quite a few CoreAnnotations are provided in the class appropriately called CoreAnnotations. If the particular one you are looking for is not present, you can add a new subclass by using one of the existing CoreAnnotations as an example.
If the feature you attached to the CoreLabel is not already used as a feature in NERFeatureFactory, you will need to add code that extracts the feature from the CoreLabel and adds it to the feature set. Bear in mind that features must have unique names, or they will conflict with existing features, which is why we add markers such as "-GENIA", "-PGENIA", and "-NGENIA" to our features. As long as you choose a unique marker, the feature itself can be any string followed by its marker and will not conflict with any existing features. Processing is done using a bag of features model, with all of the features mixed together, which is why it is important to not have any name conflicts.
Once you've annotated your data, you make a properties file with the
features you want. You can use the example properties file, and refer
to the NERFeatureFactory for more possible features. Finally, you can
test on your annotated test data as shown above or annotate more text
-textFile command rather than
Here are some tips on memory usage for CRFClassifier:
qnSize. The default is 25. Using 10 is perfectly adequate. If you're short of memory, things will still work with much smaller values, even just a value of 2.
saveFeatureIndexToDisk = true. The feature names aren't actually needed while the core model estimation (optimization) code is run. This option saves them to a file before the optimizer runs, enabling the memory they use to be freed, and then loads the feature index from disk after optimization is finished.
maxLeft=1and no features that refer to the
answerclass more than one away - it's okay to refer to word features any distance away). While the code supports arbitrary order CRFs, building second, third, or fourth order CRFs will greatly increase memory usage and normally isn't necessary. Remember:
maxLeftrefers to the size of the class contexts that your features use (that is, it is one smaller than the clique size). A first order CRF can still look arbitrarily far to the left or right to get information about the observed data context.
true. CRFClassifier will then write (potentially huge) files in the current directory listing the features generated for each token position. Options that generate huge numbers of features include
maxNGramLengis a large number.
flag useObservedSequencesOnly=true. This makes it so that you can only label adjacent words with label sequences that were seen next to each other in the training data. For some kinds of data this actually gives better accuracy, for other kinds it is worse. But unless the label sequence patterns are dense, it will reduce your memory usage.
featureDiffThresh, for example
featureDiffThresh=0.05. In training, CRFClassifier will train one model, drop all the features with weight (absolute value) beneath the given threshold, and then train a second model. Training thus takes longer, but the resulting model is smaller and faster at runtime, and usually has very similar performance for a reasonable threshold such as 0.05.
Instead of setting the
trainFile property or flag, set
the trainFileList property or flag. Use a comma
separated list of files.
Typically you would load a classifier from disk with the
CRFClassifier.getClassifier() method and then use it to
classify some text. See the example
NERDemo file. The
two most flexible classification methods to call are called
classify(). These return a
List<CoreLabel>, or a list of those, and take the
same type or a String, respectively. A
everything you could need: the original token, its (Americanized, Penn
Treebank) normalized form used in the system, its begin and end character offsets, a
record of the whitespace around it, and the class assigned to the
token. Print some out and have a look (
print some for a sample sentence, if run with no arguments). There are also a number of
other classification methods that take a String of text as input, and
provide various forms of user-friendly output. The method
classifyToCharacterOffsets returns a list of triples of an
entity name and its begin and end character offsets. The method
classifyToString(String, String, boolean) will return you a
String with NER-classified text in one of several formats (plain text or
or without token normalization and the preservation of spacing versus
tokenized. One of the versions of it may well do what you would like to see.
for examples of the use of several (but not all) of these methods.
Yes! But you'll need to make your own custom jar file. If you insert
into the jar file an NER model with name myModel and you put it
inside the jar file under the
/classifiers/ path as
/classifiers/myModel, then you can load it when running
from a jar file with a command like:
java -mx500m -cp stanford-ner.jar edu.stanford.nlp.ie.crf.CRFClassifier -loadClassifier classifiers/myModel -textFile sample.txt
Yes, you can. You might look at
edu.stanford.nlp.ie.NERServer as an example of having the
CRFClassifier run on a socket and wait for text to annotate and
then returning the results. Here's a complete Unix/Linux/Mac OS X
example, run from inside the folder of the distribution:
$ cp stanford-ner.jar stanford-ner-with-classifier.jar
$ jar -uf stanford-ner-with-classifier.jar classifiers/ner-eng-ie.crf-3-all2008.ser.gz
$ java -mx500m -cp stanford-ner-with-classifier.jar edu.stanford.nlp.ie.NERServer -port 9191 -loadJarClassifier ner-eng-ie.crf-3-all2008.ser.gz
$ java -cp stanford-ner-with-classifier.jar edu.stanford.nlp.ie.NERServer -port 9191 -client
Input some text and press RETURN to NER tag it, or just RETURN to finish.
President Barack Obama met Fidel Castro at the United Nations in New York.
President/O Barack/PERSON Obama/PERSON met/O Fidel/PERSON Castro/PERSON at/O the/O United/ORGANIZATION Nations/ORGANIZATION in/O New/LOCATION York/LOCATION ./O
With a bit of work, we're sure you can adapt that example to work in a REST, SOAP, AJAX, or whatever system. If not, pay us a lot of money, and we'll work it out for you. Or if you don't want to do that, maybe look at what Hat Doang did. An adaptation of that code now runs our online demo.
In recent versions of our NER code, we use the typesafe heterogeneous container pattern that Josh Bloch has talked about in various places such as this talk. It's neat but somewhat stresses the implementation of generic types in Java. The code is correct and should compile okay. It does compile okay with current versions of Sun javac v1.5 or v1.6 and with the current version of the Eclipse compiler. If it doesn't compile for you, you should upgrade your Java compiler or complain to the person who makes it.
NERDemo.javaincluded in the 2008-05-07 version of the software?
In this release (only) we made a booboo, and didn't update the
NERDemo.java file to correspond to changes in the main
code, and the supplied code doesn't compile.
Here is a downloadable version of
which will work with that version. The commands you need to use it are
(These are the commands for Linux or the Mac OS X command-line; for Windows, replace the colons above with semicolons and the slash with a backslash.)
javac -cp "stanford-ner.jar:." NERDemo.java
java -cp "stanford-ner.jar:." NERDemo
java -mx400m -cp "stanford-ner.jar:." NERDemo classifiers/ner-eng-ie.crf-3-all2008-distsim.ser.gz myFile.txt
Several options are available from the command-line for determining the output format of the classifier.
You can choose an
slashTags (the default). See the example of each below (these are
shell command lines, the last bit of which suppresses message printing, so you can see just the output).
Even more power is available if you are using the API.
classifier.classifyToString(String text, String outputFormat, boolean preserveSpaces)
method supports 6 output styles (of which 3 are available with the
the XML output options preserve spaces, but the slash tags one doesn't). Even more flexibility can be
obtained by using other of the
classify.* methods in the API. These will return
classified versions of the input, which you can print out however your heart desires!
There are also methods like
classifyToCharacterOffsets(String) which returns just the
See the examples in
$ cat PatrickYe.txt I complained to Microsoft about Bill Gates. They told me to see the mayor of New York. $ $ java -mx500m -cp stanford-ner.jar edu.stanford.nlp.ie.crf.CRFClassifier -loadClassifier classifiers/ner-eng-ie.crf-3-all2008-distsim.ser.gz -textFile PatrickYe.txt -outputFormat slashTags 2> /dev/null I/O complained/O to/O Microsoft/ORGANIZATION about/O Bill/PERSON Gates/PERSON ./O They/O told/O me/O to/O see/O the/O mayor/O of/O New/LOCATION York/LOCATION ./O $ $ java -mx500m -cp stanford-ner.jar edu.stanford.nlp.ie.crf.CRFClassifier -loadClassifier classifiers/ner-eng-ie.crf-3-all2008-distsim.ser.gz -textFile PatrickYe.txt -outputFormat inlineXML I complained to <ORGANIZATION>Microsoft</ORGANIZATION> about <PERSON>Bill Gates</PERSON>. They told me to see the mayor of <LOCATION>New York</LOCATION>. $ $ java -mx500m -cp stanford-ner.jar edu.stanford.nlp.ie.crf.CRFClassifier -loadClassifier classifiers/ner-eng-ie.crf-3-all2008-distsim.ser.gz -textFile PatrickYe.txt -outputFormat xml <wi num="0" entity="O">I</wi> <wi num="1" entity="O">complained</wi> <wi num="2" entity="O">to</wi> <wi num="3" entity="ORGANIZATION">Microsoft</wi> <wi num="4" entity="O">about</wi> <wi num="5" entity="PERSON">Bill</wi> <wi num="6" entity="PERSON">Gates</wi><wi num="7" entity="O">.</wi> <wi num="0" entity="O">They</wi> <wi num="1" entity="O">told</wi> <wi num="2" entity="O">me</wi> <wi num="3" entity="O">to</wi> <wi num="4" entity="O">see</wi> <wi num="5" entity="O">the</wi> <wi num="6" entity="O">mayor</wi> <wi num="7" entity="O">of</wi> <wi num="8" entity="LOCATION">New</wi> <wi num="9" entity="LOCATION">York</wi><wi num="10" entity="O">.</wi>
The classifier used for the CoNLL 2003 shared task isn't the same as the one we distribute. The former was a 3rd order CMM (MEMM) classifier, while the latter is a 1st order CRF. So you shouldn't have any detailed assumptions about being able to replicate the results of that paper. (But the features used in that paper are fairly carefully documented, and an earlier version of NERFeatureFactory was used in that system, so between the two you should be able to work out pretty exactly the features that paper used.)
For our distributed CRF (2009 version), the CoNLL classifier that we distribute wasn't very well tuned. Here are two better models:
These haven't been tested on the CoNLL English testb data, but based on their performance on the testa data (91.3 and 93.0 F1), we'd predict that they'd get about 85.3 and 87.5 F1 respectively on the testb data. So, their performance is within range of previous CoNLL 2003 results. The first is trained only on the training data, while the second makes use of distributional similarity features trained on a larger data set. Also, these models are IOB2 classifiers unlike the IO classifiers which we distribute with the software.
This is almost always a character encoding issue. You need to know
how your German text files are encoded (normally either one of the
traditional 8-bit encodings ISO-8859-1 or ISO-8859-15, or Unicode
utf-8). Then you either need to know the default platform character
encoding of your computer or to override it (these days the default
encoding is normally utf-8, unless you're
on Mac OS X, where Mac OS X Java still uses the 8-bit MacRoman
encoding as its default, despite most of the rest of the OS using utf-8).
You can override the
encoding of read and written files with the
The asymptotic memory growth in number of states
depends on the order of the CRF. If the CRF is order
then the memory growth will be
Unfortunately, no. The licenses that come with that data do not allow for redistribution. However, you can see the models section of the NER page for a brief description of what the data sets were.
Yes, the underlying CRF is deterministic. If you apply the NER to the same sentence more than once, though, it is possible to get different answers the second time. The reason for this is the NER remembers whether it has seen a word in lowercase form before.
The exact way this is used as a feature is in the word shape feature, which treats words such as "Brown" differently if it has or has not seen "brown" as a lowercase word before. If it has, the word shape will be "Initial upper, have seen all lowercase", and if it has not, the word shape will be "Initial upper, have not seen all lowercase".
This feature can be turned off in recent versions with the flag
Use the following options:
-tokenizerFactory edu.stanford.nlp.process.WhitespaceTokenizer -tokenizerOptions "tokenizeNLs=true"
None of our current models use pos tags by default. This is largely because the features used by the Stanford POS tagger are very similar to those used in the NER system, so there is very little benefit to using POS tags.
However, it certainly is possible to train new models which do use POS tags. The training data would need to have an extra column with the tag information, and you would then add
tag=X to the
None of the models we release were trained with gazette features turned on.
However, it is not difficult to train a new model which does use the gazette features. Set the
useGazettes parameter to true and set
gazette to the file you want to use. You need to supply the gazette at training time, and then the NER will learn features based on words in that gazette. The gazette will be included in the model, so does not need to be redistributed or given at test time. Any gazettes supplied at test time will be treated as additional information to be included in the gazette.
There are two different ways for gazette features to match. The first is exact match, which means that if
John Bauer is a gazette entry, the feature will only fire when
John Bauer is encountered in the text. This is turned on when the
cleanGazette option is set to true. The other is partial match, which is turned on with the
sloppyGazette option. In that case, features fire whenever any of the words match, so
John by itself would also trigger a feature in this case.
The gazette files should be of the format
CLASS1 this is an example
CLASS2 this is another example
CLASS2, etc represent the classes the model is trained to recognize, that is not strictly necessary. For example, you could have a model trained to recognize PERSON, and have two different gazette classes which represent lists of people; one could be first names and one could be last names, even if you intend them all to be features the NER generally uses to recognize names.
If a gazette is used, this does not guarantee that words in the gazette are always used as a member of the intended class, and it does not guarantee that words outside the gazette will not be chosen. It simply provides another feature for the CRF to train against. If the CRF has higher weights for other features, the gazette features may be overwhelmed.
If you want something that will recognize text as a member of a class if and only if it is in a list of words, you might prefer either the regexner or the tokensregex tools included in Stanford CoreNLP. The CRF NER is not guaranteed to accept all words in the gazette as part of the expected class, and it may also accept words outside the gazette as part of the class.
An example of how to use the gazette is presented in
Note that both entries in the first and second chapters show up in
austen.gaz.txt; otherwise, the gazette would not help recognize names
that had not shown up anywhere in the training data. If you train the
NER model with the provided gazette, you will note that
Captain Weston is correctly identified all three times in
the new model, whereas the original model does not label Captain
correctly in all cases.
You can discuss other topics with Stanford NER developers and users by
java-nlp-user mailing list
(via a webpage). Or you can send other questions and feedback to
Site design by Bill MacCartney