Resolve coreference using Stanford CoreNLP

By | February 22, 2015

Coreference resolution is the task of finding all expressions that refer to the same entity in a text. Stanford CoreNLP coreference resolution system is the state-of-the-art system to resolve coreference in the text. To use the system, we usually create a pipeline, which requires tokenization, sentence splitting, part-of-speech tagging, lemmarization, named entity recoginition, and parsing. However sometimes, we use others tools for preprocessing, particulaly when we are working on a specific domain. In these cases, we need a stand-alone coreference resolution system. This post demenstrates how to create such a system using Stanford CoreNLP.

Load properties

In general, we can just create an empty Properties, because the Stanford CoreNLP tool can automatically load the default one in the model jar file, which is under edu.stanford.nlp.pipeline.

In other cases, we would like to use specific properties. The following code shows one example of loading the property file from the working directory.

private static final String PROPS_SUFFIX = ".properties";

  private Properties loadProperties(String name) {
    return loadProperties(name, 
  private Properties loadProperties(String name, ClassLoader loader) {
    if (name.endsWith(PROPS_SUFFIX))
      name = name.substring(0, name.length() - PROPS_SUFFIX.length());
    name = name.replace('.', '/');
    name += PROPS_SUFFIX;
    Properties result = null;

    // Returns null on lookup failures
    System.err.println("Searching for resource: " + name);
    InputStream in = loader.getResourceAsStream(name);
    try {
      if (in != null) {
        InputStreamReader reader = new InputStreamReader(in, "utf-8");
        result = new Properties();
        result.load(reader); // Can throw IOException
    } catch (IOException e) {
      result = null;
    } finally {

    return result;

Initialize the system

After getting the properties, we can initialize the coreference resovling system. For example,

try {
      corefSystem = new SieveCoreferenceSystem(new Properties());
      mentionExtractor = new MentionExtractor(
    } catch (Exception e) {
        "ERROR: cannot create DeterministicCorefAnnotator!");
      throw new RuntimeException(e);


To feed the resolving system, we first need to understand the structure of annotation, which represents a span of text in a document. This is the most tricky part in this post, because to my knowledge there is no document to explain it in details. The Annotation class itself is just an implementation of Map.

Basically, an annotation contains a sequence of sentences (which is another map). For each sentence, we need to provide a seqence of tokens (a list of CoreLabel), the parsing tree (Tree), and the dependency graph (SemanticGraph).

  CoreAnnotations.SentencesAnnotation -> sentences
    CoreAnnotations.TokensAnnotation -> tokens
    TreeCoreAnnotations.TreeAnnotation -> Tree
    SemanticGraphCoreAnnotations.CollapsedDependenciesAnnotation -> SemanticGraph


The sequence of tokens represents the text of one sentence. Each token is an instance of CoreLabel, which stores word, tag (part-of-speech), lemma, named entity, normailzied named entity, etc.

List<CoreLabel> tokens = new ArrayList<>();
for(int i=0; i<n; i++) {
  // create a token
  CoreLabel token = new CoreLabel();
ann.set(TokensAnnotation.class, tokens);

Parse tree

A parse tree is an instance of Tree. If you use the Penn treebank style, Stanford corenlp tool provide an easy to parse the format.

Tree tree = Tree.valueOf(getText());
ann.set(TreeAnnotation.class, tree);

Semantic graph

Semantic graph can be created using typed dependencis from the tree by rules. However the code is not that straightforward.

GrammaticalStructureFactory grammaticalStructureFactory =
    new EnglishGrammaticalStructureFactory();
GrammaticalStructure gs = grammaticalStructureFactory
SemanticGraph semanticGraph = 
    new SemanticGraph(gs.typedDependenciesCollapsed());

Please note that Stanford Corenlp provide different types of dependencies. Among others, coreference system needs “collapsed-dependencies”, so to set the annotation, you may write

          new SemanticGraph(gs.typedDependenciesCollapsed()));

Resolve coreference

At last, you can feed the system with the annotation. The following code is one example. It is a bit long but easy to understand.

private void annotate(Annotation annotation) {
    try {
      List<Tree> trees = new ArrayList<Tree>();
      List<List<CoreLabel>> sentences = 
        new ArrayList<List<CoreLabel>>();

      // extract trees and sentence words
      // we are only supporting the new annotation standard for this Annotator!
      if (annotation.containsKey(CoreAnnotations.SentencesAnnotation.class)) {
        // int sentNum = 0;
        for (CoreMap sentence : annotation
            .get(CoreAnnotations.SentencesAnnotation.class)) {
          List<CoreLabel> tokens = sentence
          Tree tree = sentence.get(TreeCoreAnnotations.TreeAnnotation.class);
          MentionExtractor.mergeLabels(tree, tokens);
      } else {
          "ERROR: this coreference resolution system requires SentencesAnnotation!");

      // extract all possible mentions
      // this is created for each new annotation because it is not threadsafe
      RuleBasedCorefMentionFinder finder = new RuleBasedCorefMentionFinder();
      List<List<Mention>> allUnprocessedMentions = finder
          .extractPredictedMentions(annotation, 0, corefSystem.dictionaries());

      // add the relevant info to mentions and order them for coref
      Document document = mentionExtractor.arrange(
      List<List<Mention>> orderedMentions = document.getOrderedMentions();
      if (VERBOSE) {
        for (int i = 0; i < orderedMentions.size(); i++) {
          System.err.printf("Mentions in sentence #%d:\n", i);
          for (int j = 0; j < orderedMentions.get(i).size(); j++) {
            System.err.println("\tMention #"
                + j
                + ": "
                + orderedMentions.get(i).get(j).spanToString());

      Map<Integer, CorefChain> result = corefSystem.coref(document);
      annotation.set(CorefCoreAnnotations.CorefChainAnnotation.class, result);

      // for backward compatibility
      if (OLD_FORMAT) {
        List<Pair<IntTuple, IntTuple>> links = SieveCoreferenceSystem

        if (VERBOSE) {
          System.err.printf("Found %d coreference links:\n", links.size());
          for (Pair<IntTuple, IntTuple> link : links) {
                "LINK (%d, %d) -> (%d, %d)\n",

        // save the coref output as CorefGraphAnnotation

        // cdm 2013: this block didn't seem to be doing anything needed....
        // List<List<CoreLabel>> sents = new ArrayList<List<CoreLabel>>();
        // for (CoreMap sentence:
        // annotation.get(CoreAnnotations.SentencesAnnotation.class)) {
        // List<CoreLabel> tokens =
        // sentence.get(CoreAnnotations.TokensAnnotation.class);
        // sents.add(tokens);
        // }

        // this graph is stored in CorefGraphAnnotation -- the raw links found
        // by the coref system
        List<Pair<IntTuple, IntTuple>> graph = new ArrayList<Pair<IntTuple, IntTuple>>();

        for (Pair<IntTuple, IntTuple> link : links) {
          // Note: all offsets in the graph start at 1 (not at 0!)
          // we do this for consistency reasons, as indices for syntactic
          // dependencies start at 1
          int srcSent = link.first.get(0);
          int srcTok = orderedMentions.get(srcSent - 1).get(
              link.first.get(1) - 1).headIndex + 1;
          int dstSent = link.second.get(0);
          int dstTok = orderedMentions.get(dstSent - 1).get(
              link.second.get(1) - 1).headIndex + 1;
          IntTuple dst = new IntTuple(2);
          dst.set(0, dstSent);
          dst.set(1, dstTok);
          IntTuple src = new IntTuple(2);
          src.set(0, srcSent);
          src.set(1, srcTok);
          graph.add(new Pair<IntTuple, IntTuple>(src, dst));
        annotation.set(CorefCoreAnnotations.CorefGraphAnnotation.class, graph);

        for (CorefChain corefChain : result.values()) {
          if (corefChain.getMentionsInTextualOrder().size() < 2)
          Set<CoreLabel> coreferentTokens = Generics.newHashSet();
          for (CorefMention mention : corefChain.getMentionsInTextualOrder()) {
            CoreMap sentence = annotation.get(
                mention.sentNum - 1);
            CoreLabel token = sentence.get(
                mention.headIndex - 1);
          for (CoreLabel token : coreferentTokens) {
    } catch (RuntimeException e) {
      throw e;
    } catch (Exception e) {
      throw new RuntimeException(e);

6 thoughts on “Resolve coreference using Stanford CoreNLP

  1. Clement Yeo

    Dear Yifan, Will I be able to get a copy of your full source code on the execution of co reference resolution in Java. I am currently working on a project which requires Coreference resolution but need some help with the coding. Regards, Clement Yeo

    1. Yifan Peng Post author

      My source code is embedded in my research project. I am afraid it is very hard to send you a copy of full codes. On the other hand, you can post your “not-working” code here, and I am glad to provide some advice.

  2. Anita


    thank you for your instructions so far 🙂 Do you mind to explain what you mean with “VERBOSE”, “OLD_FORMAT” and “IntTuple”? Also, I did not fully understand where I should include the different code parts you posted.

    Thank you for your help and kind regards!

  3. Rene

    The code is not working. ‘corefSystem’ in ‘private void annotate(Annotation annotation) ‘ is not known in that scope and there is no class declaration.

    Also what is ‘Document’ there are more than one Document implementations in Stanford, but both are not working.

    Too bad that this blog post is more or less useless.

    1. Yifan Peng Post author

      I am sorry for the inconvenience. I believe the code changed greatly after I wrote this post. I didn’t keep track the changes.

  4. zafari

    Dear Yifan, Will I be able to use Stanford Coreference system to get anaphoricity status of a mention please?
    the following code gets most of the mentions’ properties except the one I want.
    StanfordCoreNLP pipeline = new StanfordCoreNLP(props);
    List sentences = document.get(SentencesAnnotation.class);
    corefSystem = new SieveCoreferenceSystem(props);
    mentionExtractor = new MentionExtractor(corefSystem.dictionaries(), corefSystem.semantics());
    RuleBasedCorefMentionFinder finder = new RuleBasedCorefMentionFinder();
    List<List> allUnprocessedMentions = finder.extractPredictedMentions(document, 0, corefSystem.dictionaries());
    would you help me please???


Leave a Reply

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.