standford nlp 教程






Annotator dependencies

tokenize TokenizerAnnotator None
cleanxml CleanXmlAnnotator tokenize
ssplit WordsToSentenceAnnotator tokenize
pos POSTaggerAnnotator tokenize, ssplit
lemma MorphaAnnotator tokenize, ssplit, pos
ner NERClassifierCombiner tokenize, ssplit, pos, lemma
regexner RegexNERAnnotator ?
sentiment SentimentAnnotator ?
parse ParserAnnotator tokenize, ssplit
depparse DependencyParseAnnotator tokenize, ssplit, pos
dcoref DeterministicCorefAnnotator tokenize, ssplit, pos, lemma, ner, parse
relation RelationExtractorAnnotator tokenize, ssplit, pos, lemma, ner, depparse
natlog NaturalLogicAnnotator tokenize, ssplit, pos, lemma, depparse (Can also use parse)
quote QuoteAnnotator None


注意:3.6,3.7 版本的要求jdk1.8或以上









Stanford CoreNLP can be downloaded via the link below. This will download a large (536 MB) zip file containing (1) the CoreNLP code jar, (2) the CoreNLP models jar (required in your classpath for most tasks) (3) the libraries required to run CoreNLP, and (4) documentation / source code for the project. This is everything for getting going on English! Unzip this file, open the folder that results and you’re ready to use it.

Other languages: For working with another (human) language, you need additional model files. We have model files for several other languages. And we have more model files for English, including for dealing with uncased English (that is, English which is not conventionally capitalized, whether texting or telegrams). You can find the latest models in the table below. Versions for earlier releases are available on the release history page.

Arabic download 3.7.0
Chinese download 3.7.0
English download 3.7.0
English (KBP) download 3.7.0
French download 3.7.0
German download 3.7.0
Spanish download 3.7.0

If you want to change the source code and recompile the files, see these instructions. Previous releases can be found on the release history page.

Java: Stanford CoreNLP now requires Java 8. If you do not have this installed you should first of all install Java 8. Probably the JDK, but the JRE will do if you are only going to be a user.

GitHub: Here is the Stanford CoreNLP GitHub site.

Maven: You can find Stanford CoreNLP on Maven Central. The crucial thing to know is that CoreNLP needs its models to run (most parts beyond the tokenizer) and so you need to specify both the code jar and the models jar in your pom.xml, as follows: (Note: Maven releases are made several days after the release on the website.)


If you want to get a language models jar off of Maven for Arabic, Chinese, German, or Spanish, add this to your pom.xml:


Replace “models-chinese” with one or more of “models-english”, “models-english-kbp”, “models-arabic”, “models-french”, “models-german” or “models-spanish” for resources for other languages!


import edu.stanford.nlp.simple.*;

public class SimpleCoreNLPDemo {
    public static void main(String[] args) { 
        // Create a document. No computation is done yet.
        Document doc = new Document("add your text here! It can contain multiple sentences.");
        for (Sentence sent : doc.sentences()) {  // Will iterate over two sentences
            // We're only asking for words -- no need to load any models yet
            System.out.println("The second word of the sentence '" + sent + "' is " + sent.word(1));
            // When we ask for the lemma, it will load and run the part of speech tagger
            System.out.println("The third lemma of the sentence '" + sent + "' is " + sent.lemma(2));
            // When we ask for the parse, it will load and run the parser
            System.out.println("The parse of the sentence '" + sent + "' is " + sent.parse());
            // ...

package lhy;

import java.util.Properties;

import edu.stanford.nlp.pipeline.Annotation;
import edu.stanford.nlp.pipeline.StanfordCoreNLP;
import edu.stanford.nlp.util.PropertiesUtils;

public class test2 {

	public static void main(String[] args) throws FileNotFoundException, IOException {
		String text = "克林顿说,华盛顿将逐步落实对韩国的经济援助。"
		        + "金大中对克林顿的讲话报以掌声:克林顿总统在会谈中重申,他坚定地支持韩国摆脱经济危机。";
		Annotation document = new Annotation(text);

		Properties props = PropertiesUtils.asProperties("annotators","tokenize","segment.model","edu/stanford/nlp/models/segmenter/chinese/ctb.gz","tokenize.language","zh"
//		 Properties props = new Properties();
//		 props.load(new FileInputStream("/home/lhy/Workspaces/MyEclipse_data/lhy/src/main/resources/"));
		StanfordCoreNLP corenlp = new StanfordCoreNLP(props);



CoreNLP includes a simple web API server for servicing your human language understanding needs (starting with version 3.6.0). This page describes how to set it up. CoreNLP server provides both a convenient graphical way to interface with your installation of CoreNLP and an API with which to call CoreNLP using any programming language. If you’re writing a new wrapper of CoreNLP for using it in another language, you’re advised to do it using the CoreNLP Server.

Getting Started

Stanford CoreNLP ships with a built-in server, which requires only the CoreNLP dependencies. To run this server, simply run:

# Run the server using all jars in the current directory (e.g., the CoreNLP home directory)
java -mx4g -cp "*" edu.stanford.nlp.pipeline.StanfordCoreNLPServer -port 9000 -timeout 15000

Note the the timeout is in milliseconds.

If you want to process non-English languages, use this command with the appropriate language properties:

# Run a server using Chinese properties
java -Xmx4g -cp "*" edu.stanford.nlp.pipeline.StanfordCoreNLPServer -serverProperties -port 9000 -timeout 15000

Each language has a models jar which must also be on the CLASSPATH. The most recently models jars for each language can be found here.

If no value for port is provided, port 9000 will be used by default. You can then test your server by visiting


You should see a website similar to, with an input box for text and a list of annotators you can run. From this interface, you can test out each of the annotators by adding/removing them from this list. (Note: The first use will be slow to respond while models are loaded – it might take 30 seconds or so, but after that the server should run quite quickly.) You can test out the API by sending aPOST request to the server with the appropriate properties. An easy way to do this is with wget. The following will annotate the sentence “the quick brown fox jumped over the lazy dog” with part of speech tags:

wget --post-data 'The quick brown fox jumped over the lazy dog.' 'localhost:9000/?properties={"annotators":"tokenize,ssplit,pos","outputFormat":"json"}' -O -

Or if you only have or prefer curl:

curl --data 'The quick brown fox jumped over the lazy dog.' 'http://localhost:9000/?properties={%22annotators%22%3A%22tokenize%2Cssplit%2Cpos%22%2C%22outputFormat%22%3A%22json%22}' -o -

The rest of this document: describes the API in more detail, describes a Java client to the API as a drop-in replacement for theStanfordCoreNLP annotator pipeline, and talks about administering the server.

API Documentation

The greatest strength of the server is the ability to make API calls against it.

NOTE: Please do not make API calls against It is not set up to handle a large volume of requests. Instructions for setting up your own server can be found in the Dedicated Server section.

There are three endpoints provided by the server, which we’ll describe in more detail below. Each of them takes as input a series of url parameters, as well as POST data consisting of the serialized document or raw text to be annotated. The endpoints are:

  • / Provides an interface to annotate documents with CoreNLP.
  • /tokensregex Provides an interface for querying text for TokensRegex patterns, once it has been annotated with CoreNLP (using the enpoint above).
  • /semgrex Similar to /tokensregex above, this endpoint matches text against semgrex patterns.

Annotate with CoreNLP: /

This endpoint takes as input a JSON-formatted properties string under the key properties=<properties>, and as POSTdata text to annotate. The properties should mirror the properties file passed into the CoreNLP command line, except formatted as a JSON object. For example, the following will tokenize the input text, run part of speech tagging, and output it as JSON to standard out:

wget --post-data 'the quick brown fox jumped over the lazy dog' 'localhost:9000/?properties={"annotators": "tokenize,ssplit,pos", "outputFormat": "json"}' -O -

A common property to set is the output format of the API. The server supports all output formats provided by CoreNLP. These are listed below, along with their relevant properties:

  • JSON: Print the annotations in JSON format. This corresponds to the property: {"outputFormat": "json"}.
  • XML: Print the annotations in XML format. This corresponds to the property: {"outputFormat": "xml"}.
  • Text: Print the annotations in a human-readable text format. This is the default format for the CoreNLP command-line interface. This corresponds to the property: {"outputFormat": "text"}.
  • Serialized: Print the annotations in a losslessly serialized format. This is the recommended option when calling the API programmatically from a language that supports one of the serialized formats. In addition to setting the output format flag, you must also provide a valid serializer class. For example, for protocol buffers, this would be:
    {"outputFormat": "serialized", 
     "serializer": "edu.stanford.nlp.pipeline.ProtobufAnnotationSerializer"}

    The serializers currently supported are:

    • edu.stanford.nlp.pipeline.ProtobufAnnotationSerializer Writes the output to a protocol buffer, as defined in the definition file edu.stanford.nlp.pipeline.CoreNLP.proto.
    • edu.stanford.nlp.pipeline.GenericAnnotationSerializer Writes the output to a Java serialized object. This is only suitable for transferring data between Java programs. This also produces relatively large serialized objects.
    • edu.stanford.nlp.pipeline.CustomAnnotationSerializer Writes the output to a (lossy!) textual representation, which is much smaller than the GenericAnnotationSerializer but does not include all the relevant information.

The server also accepts input in a variety of formats. By default, it takes input as raw text sent as POST data to the server. However, it can also be configured to read the POST data using one of the CoreNLP serializers. This can be set up by setting the propertiesinputFormat and inputSerializer. For example, to read the data as a protocol buffer (useful if, e.g., it is already partially annotated), simply include the following in your url parameter properties={...}:

{"inputFormat": "serialized",
 "inputSerializer": "edu.stanford.nlp.pipeline.ProtobufAnnotationSerializer"}

A complete call to the server, taking as input a protobuf serialized document at path /path/to/file.proto and returning as a response a protobuf for the document annotated for part of speech and named entity tags (to the file/path/to/annotated_file.proto could be:

wget --post-file /path/to/file.proto 'localhost:9000/?properties={"inputFormat": "serialized", "inputSerializer", "edu.stanford.nlp.pipeline.ProtobufAnnotationSerializer", "annotators": "tokenize,ssplit,pos,lemma,ner", "outputFormat": "serialized", "serializer", "edu.stanford.nlp.pipeline.ProtobufAnnotationSerializer"}' -O /path/to/annotated_file.proto

Query TokensRegex: /tokensregex

Similar to the CoreNLP target, /tokensregex takes a block of data (e.g., text) as POST data, and a series of url parameters. Currently, only plain-text POST data is supported. The two relevant url parameters are:

  • pattern: The TokensRegex pattern to annotate.
  • filter: If true, entire sentences must match the pattern, rather than the API finding matching sections.

The response is always in JSON, formatted as follows:

{"sentences": {
	"0": {
	  "text": "the matched text",
	  "begin": 2,
	  "end": 5,
	  "$captureGroupKey": {
		  "text": "the matched text",
		  "begin": 2,
		  "end": 5,

Query Semgrex: /semgrex

Similar to the CoreNLP target, and nearly identical to TokensRegex, /semgrex takes a block of data (e.g., text) as POST data, and a series of url parameters. Currently, only plain-text POST data is supported. The two relevant url parameters are:

  • pattern: The Semgrex pattern to annotate.
  • filter: If true, entire sentences must match the pattern, rather than the API finding matching sections.

The response is always in JSON, formatted identically to the tokensregex output, with the exception that all spans are single words (only the root of the match is returned):

{"sentences": {
	"0": {
	  "text": "text",
	  "begin": 4,
	  "end": 5,
	  "$captureGroupKey": {
		  "text": "text",
		  "begin": 4,
		  "end": 5,

Java Client

CoreNLP includes a Java client to the server – StanfordCoreNLPClient – which mirrors the interface of the annotation pipeline ( as closely as possible. The primary motivating use cases for using this class and not the local pipeline are:

  • The models are not re-loaded every time your program runs. This is useful when debugging a block of code which runs CoreNLP annotations, as the CoreNLP models often take on the order of minutes to load from disk.
  • The machine running the server has more compute and more memory than your local machine. Never again must Chrome and CoreNLP compete for the same memory.

The constructors to StanfordCoreNLPClient take the following 3 required arguments, and a fourth optional argument:

  1. Properties props: Mirroring the local pipeline exactly, these are the properties to use when annotating text with the pipeline. Minimally, this specifies the annotators to run.
  2. String host: The hostname of the server.
  3. int port: The port that the server is running on.
  4. int threads: Optionally, the number of threads to hit the server with. If, for example, the server is running on an 8 core machine, you can specify this to be 8, and the client will allow you to make 8 simultaneous requests to the server. Note that there is nothing that ensures that you have these threads reserved on the server: two clients can both hit the server with 8 threads, and the server will just respond half as fast.

An example programmatic usage of the client, hitting a server at localhost:9000 with up to 2 threads, is as follows. Note that this exactly mirrors the usage of the conventional pipeline.

// creates a StanfordCoreNLP object with POS tagging, lemmatization, NER, parsing, and coreference resolution
Properties props = new Properties();
props.setProperty("annotators", "tokenize, ssplit, pos, lemma, ner, parse, dcoref");
StanfordCoreNLPClient pipeline = new StanfordCoreNLPClient(props, "localhost", 9000, 2);
// read some text in the text variable
String text = ... // Add your text here!
// create an empty Annotation just with the given text
Annotation document = new Annotation(text);
// run all Annotators on this text

You can also run the client from the command line, and get an interface similar to the command line usage for the local CoreNLP program. The following will annotate a file input.txt with part-of-speech, lemmas, named entities, constituency parses, and coreference:

java -cp "*" -Xmx1g edu.stanford.nlp.pipeline.StanfordCoreNLPClient -annotators tokenize,ssplit,pos,lemma,ner,parse,dcoref -file input.txt

NOTE: Again, please do not make API calls against It is not set up to handle a large volume of requests. Instructions for setting up your own server can be found in the Dedicated Server section.

Once you have your own server(s) set up, you can run against them with a command like this:

java edu.stanford.nlp.pipeline.StanfordCoreNLPClient -cp "*" -annotators tokenize,ssplit,pos,lemma,ner,parse,dcoref -file input.txt  -backends localhost:9000

You specify one or more back-end servers in a comma-separated list as the arguments of the -backends option. Each is specified as host:port.

Providing that the server has foreign language models available on its classpath, you can ask for it to work with texts in other languages. If you have the French properties file and a file called french.txt in your current directory, then you should be able to successfully give a command like this:

java -cp "*" edu.stanford.nlp.pipeline.StanfordCoreNLPClient -props -annotators tokenize,ssplit,pos,depparse ile french.txt -outputFormat conllu -backends localhost:9000

Usage via other programming languages

There are now modules for several programming languages, including Python and JavaScript, which work by talking to a Stanford CoreNLP server instance. Indeed, this is now normally the best way to implement an interface to CoreNLP in other languages.

Check out what is available on the Other programming languages and packages page.

Server Administration

This section describes how to administer the server, including starting and stopping the server, as well as setting it up as a startup task

Starting the Server

The server is started directly though calling it with java. For example, the following will start the server in the background on port 1337, assuming your classpath is set properly:

nohup java -mx4g edu.stanford.nlp.pipeline.StanfordCoreNLPServer 1337 &

The classpath must include all of the CoreNLP dependencies. The memory requirements of the server are the same as that of CoreNLP, though it will grow as you load more models (e.g., memory increases if you load both the PCFG and Shift-Reduce constituency parser models). A safe minimum is 4gb; 8gb is recommended if you can spare it.


If running the server under docker, the container’s port 9000 has to be published to the host. Give a command like:docker run -p 9000:9000 --name coreNLP --rm -i -t motiz88/corenlp. If, when going to localhost:9000/, you see the error This site can’t be reached. localhost refused to connect, then this is what you failed to do!

Stopping the Server

The server can be stopped programmatically by making a call to the /shutdown endpoint with an appropriate shutdown key. This key is saved to the file /tmp/corenlp.shutdown when the server starts. An example command to shut down the server would be:

wget "localhost:9000/shutdown?key=`cat /tmp/corenlp.shutdown`" -O -

Dedicated Server

This section describes how to set up a dedicated CoreNLP server on a fresh Linux install. As always, make sure you understand the commands being run below, as they largely require root permissions:

  1. Place all of the CoreNLP jars (code, models, and library dependencies) in a directory /opt/corenlp. The code will be in a jar named stanford-corenlp-<version>.jar. The models will be in a jar namedstanford-corenlp-<version>-models.jar; other language, caseless or shift-reduce models can also be added here. The minimal library dependencies, included in the CoreNLP release, are:
    • joda-time.jar
    • jollyday-<version>.jar
    • protobuf.jar
    • xom-<version>.jar
  2. Install authbind. On Ubuntu, this is as easy as sudo apt-get install authbind.

  3. Create a user nlp with permissions to read the directory /opt/corenlp. Allow the user to bind to port 80:
sudo mkdir -p /etc/authbind/byport/
sudo touch /etc/authbind/byport/80
sudo chown nlp:nlp /etc/authbind/byport/80
sudo chmod 600 /etc/authbind/byport/80
  1. Copy the startup script from the source jar at path edu/stanford/nlp/pipeline/demo/corenlp to/etc/init.d/corenlp. An easy way to get this is:
sudo wget -O /etc/init.d/corenlp
  1. Give executable permissions to startup script: sudo chmod a+x /etc/init.d/corenlp

  2. Link the script to /etc/rc.d/ln -s /etc/init.d/corenlp /etc/rc.d/rc2.d/S75corenlp

The CoreNLP server will now start on startup, running on port 80 under the user nlp. To manually start/stop/restart the server, you can use:

sudo service corenlp [start|stop|restart]

Quirks and Subtleties

This section documents some of the subtle quirks of the server, and the motivations behind them.

Character Encoding

The official HTTP 1.1 specification recommends ISO-8859-1 as the encoding of a request, unless a different encoding is explicitly set by using the Content-Type header. However, for most NLP applications this is an unintuitive default, and so the server instead defaults to UTF-8. To enable the ISO-8859-1 default, pass in the -strict flag to the server at startup.

Default Properties

The server has different default properties than the regular CoreNLP pipeline. These are:

  • The default output format is json rather than text (-outputFormat json). This is more natural for most cases when you would be making API calls against a server.
  • By default, the server will not pretty print the output, opting instead for a minified output. This is the same as setting the property-prettyPrint false.
  • The default annotators do not include the parse annotator. This is primarily for efficiency. The annotators enabled by default are: -annotators tokenize, ssplit, pos, lemma, ner, depparse, coref, natlog, openie.
  • As a necessary consequence of not having the parse annotator, the default coref mention detector is changed to use dependency parsers: dep.

Undocumented Features

Well, I guess they’re documented now:

  • Hitting Shift+Enter on any input field in the web demo (e.g., the main text input) is equivalent to clicking the Submit (orMatch) button. Furthermore, if the input is empty, it will fill itself with a default input. Useful if – to take a purely hypothetical example – you’re developing the web server and don’t want to re-type the same sentence everytime you re-load the website.

Using the Server in a Shell Script

When booting up an instance of the server for a shell script, make sure you wait for the server to be available before interacting with it. An example using the netcat tool on linux:

java -mx4g edu.stanford.nlp.pipeline.StanfordCoreNLPServer &
# Wait until server starts
while ! nc -z localhost 9000; do
    sleep 0.1 # wait for 1/10 of the second before check again
# Rest of script
# ...

其他语言使用standford nlp:

Below are interfaces and packages for running Stanford CoreNLP from other languages or within other packages. They have been written by many other people (thanks!). In general you should contact these people directly if you have problems with these packages.




Okay, Docker isn’t a language, but you know what we mean…. Note on running the CoreNLP server under docker: The container’s port 9000 has to be published to the host. Give a command like:docker run -p 9000:9000 --name coreNLP --rm -i -t motiz88/corenlp. If, when going to localhost:9000/, you see the error This site can’t be reached. localhost refused to connect, then this is what you failed to do!


JavaScript (node.js)








Thrift server

ZeroMQ/ØMQ servers

  • stanford-0mq by Diane Napolitano. An implementation of a server for Stanford’s CoreNLP suite using Ømq and a basic client/server/JSON requests configuration. Last commit: Oct 2015.
  • stanford-corenlp-zeromq by URXtech. Basic JSON wrapper around CoreNLP.
  • corenlp-zmq by Thom Neale. A Dockerfile and Ansible provisioning script to build and run a Stanford CoreNLP server process with a single ZMQ broker font-end that proxies incoming requests to one or more back-end Scala workers. Last commit: 2015.
  • corenlp-server by Eric Kow. Simple Java server communicating with clients via XML through ZeroMQ. Example Python client included. Last commit: 2014.

个人分类: NLP
上一篇jvm 调优参数
下一篇centos 安装 nvidia的两种方式
想对作者说点什么? 我来说一句