CoreNLP stops with a FileNotFoundException, RuntimeIOException, or after failing to find a class or data file. How do I fix this?

The most likely cause of these errors is that one or more of the important jar files is missing. If it occurs when loading the models, make sure the current models file is in the classpath. The basic models file has a name like stanford-corenlp-V.V.V-models.jar, depending on the version. For other language models, you may also need additional models jars, which will have the language name in them. If you encounter this exception when trying to produce XML output, make sure xom.jar is included. Finally, if it seems to occur when loading SUTime, be sure to include joda-time.jar, etc.

Basically, you want to include all of the jar files in the download directory unless you are sure a particular jar is not needed.

How do I use the API?

A brief demo program included with the download will demonstrate how to load the tool and start processing text. When using this demo program, be sure to include all of the appropriate jar files in the classpath.

Once you have tried this, there is quite a bit of information on the CoreNLP home page describing what Annotators are available, what annotations they add to the text, and what options they support.

What character encoding does Stanford CoreNLP use?

By default, it uses Unicode’s UTF-8. You can change the encoding used when reading files by either setting the Java encoding property or more simply by supplying the program with the command line flag -encoding FOO (or including the corresponding property in a properties file that you are using).

Can you say more about adding a custom annotator?

Here are the steps:

  • Extend the class edu.stanford.nlp.pipeline.Annotator I assume you’re writing your own code to do the processing. Whatever code you write, you want to call it from a class that is a subclass of Annotator. Look at any of the existing Annotator classes, such as POSTaggerAnnotator, and try to emulate what it does.
  • Have a constructor with the signature (String, Properties) If your new annotator is FilterAnnotator, for example, it must have a constructor FilterAnnotator(String name, Properties props) in order to work.
  • Add the property customAnnotatorClass.FOO=BAR Using the same example, suppose your full class name is, and you want the new annotator to have the name “filter”. When creating the CoreNLP properties, you need to add the flag You can then add “filter” to the list of annotators in the annotators property. When you do that, the constructor FilterAnnotator(String, Properties) will be called with the name “filter” and the properties files you run CoreNLP with. This lets you define any property flag you want. For example, you could name a flag filter.verbose and then extract that flag from the properties to determine the verbosity of your new annotator.

What is the format of the XML output for coref?

Here is a sample block of coref xml output:

    <mention representative="true">

The entire coref section is demarked by a <coreference> section. Each individual chain is then demarked by another <coreference>. (This is perhaps an unfortunate naming, but at this point there are no plans to change it.) Inside the <coreference> section for each chain is a block describing each of the mentions. One mention will be labeled the representative mention. There are fields for sentence, indexed from 1 the range of words, from start (inclusive) to end (not inclusive), also indexed from 1, and head, the index in the sentence of the head word of this mention.

CoreNLP runs out of memory?

Either give CoreNLP more memory, use fewer annotators, or give CoreNLP smaller documents. Nearly all our annotators load large model files which use lots of memory. Running the full CoreNLP pipeline requires the sum of all these memory requirements. Typically, this means that CoreNLP needs about 2GB to run the entire pipeline. Additionally, the coreference module operates over an entire document. Unless things are size-limited, as either sentence length or document size increases, processing time and space increase without bound.

Running from the command line, you need to supply a flag like -Xmx2g.
If running CoreNLP from within Eclipse, follow these instructions to increase the memory given to a program being run from inside Eclipse. Increasing the amount of memory given to Eclipse itself won’t help.

What does SET mean in the NER output?

This is part of SUTime. It applies to repeating events such as “every other week” or “every two weeks”. SET is not the best name for such an event, but it matches the TIMEX3 standard (see section 2.3 of the linked document)

How do I run CoreNLP on other languages?

Other than English, we currently provide trained CoreNLP models for Chinese. To run CoreNLP on Chinese text, you first have to download the models, which can be found in our release history. Include this .jar in your classpath, and use the file it contains to process Chinese. For example, if you put the .jar in your distribution directory, you could run (adjusting the .jar version file extensions to your current release): java -cp stanford-corenlp-VV.jar:stanford-chinese-corenlp-VV-models.jar -Xmx3g edu.stanford.nlp.pipeline.StanfordCoreNLP -props -file your-chinese-file.txt

How do I fix CoreNLP giving a NoSuchMethodError or NoSuchFieldError?

If you see an Exception stacktrace message like:

Exception in thread "main" java.lang.NoSuchFieldError: featureFactoryArgs


Exception in thread "main" java.lang.NoSuchMethodError: edu.stanford.nlp.tagger.maxent.TaggerConfig.getTaggerDataInputStream(Ljava/lang/String;)Ljava/io/DataInputStream;


Caused by: java.lang.NoSuchMethodError: edu.stanford.nlp.util.Generics.newHashMap()Ljava/util/Map;
    at edu.stanford.nlp.pipeline.AnnotatorPool.<init>(
    at edu.stanford.nlp.pipeline.StanfordCoreNLP.getDefaultAnnotatorPool(

then this isn’t caused by the shiny new Stanford NLP tools that you’ve just downloaded. It is because you also have old versions of one or more Stanford NLP tools on your classpath.

The straightforward case is if you have an older version of a Stanford NLP tool. For example, you may still have a version of Stanford NER on your classpath that was released in 2009. In this case, you should upgrade, or at least use matching versions. For any releases from 2011 on, just use tools released at the same time – such as the most recent version of everything :) – and they will all be compatible and play nicely together.

The tricky case of this is when people distribute jar files that hide other people’s classes inside them. People think this will make it easy for users, since they can distribute one jar that has everything you need, but, in practice, as soon as people are building applications using multiple components, this results in a particular bad form of jar hell. People just shouldn’t do this. The only way to check that other jar files do not contain conflicting versions of Stanford tools is to look at what is inside them (for example, with the jar -tf command).

In practice, if you’re having problems, the most common cause (in 2013-2014) is that you have ark-tweet-nlp on your classpath. The jar file in their github download hides old versions of many other people’s jar files, including Apache commons-codec (v1.4), commons-lang, commons-math, commons-io, Lucene; Twitter commons; Google Guava (v10); Jackson; Berkeley NLP code; Percy Liang’s fig; GNU trove; and an outdated version of the Stanford POS tagger (from 2011). You should complain to them for creating you and us grief. But you can then fix the problem by using their jar file from Maven Central. It doesn’t have all those other libraries stuffed inside.

How do I fix CoreNLP giving an “IllegalArgumentException: Unknown option: -retainTmpSubcategories”, when trying to use it with a language other than English?

You need to add the flag -parse.flags "" (or the corresponding property parse.flags: &nbsp; ). It’s sort of a misfeature/bug that the default properties of CoreNLP turn this option on by default, because it is useful for English, but it isn’t defined for other languages, and so you get an error.)

How do I add constraints to the parser in CoreNLP?

The parser can be instructed to keep certain sets of tokens together as a single constituent. If you do this, it will try to make a parse which contains a subtree where the exact set of tokens in that subtree are the ones specified in the constraint.

For any sentence where you want to add constraints, attach the ParserAnnotations.ConstraintAnnotation to that sentence. This annotation is a List<ParserConstraint>, where ParserConstraint specifies the start (inclusive) and end (exclusive) of the range and a pattern which the enclosing constituent must match. However, there is a bug in the way patterns are handled in the parser, so it is strongly recommended to use .* for the matching pattern.

How can I get original Stanford Dependencies instead of Universal Dependencies?

If you want CoreNLP to output the original Stanford Dependencies instead of the new Universal Dependencies, simply add the option -parse.originalDependencies or the property ("parse.originalDependencies", true) to your command or code, respectively.

Note, however, that some annotators that use dependencies such as natlog might not function properly if you use this option. In case you are using the Neural Network Dependency Parser, use the following model to get Stanford Dependencies:

-depparse.model "edu/stanford/nlp/models/parser/nndep/english_SD.gz"