Stanford Core NLP - understanding coreference resolution

Question

I\'m having some trouble understanding the changes made to the coref resolver in the last version of the Stanford NLP tools. As an example, below is a sentence and the corre

Answer 1

The first number is a cluster id (representing tokens, which stand for the same entity), see source code of SieveCoreferenceSystem#coref(Document). The pair numbers are outout of CorefChain#toString():

public String toString(){
    return position.toString();
}

where position is a set of postion pairs of entity mentioning (to get them use CorefChain.getCorefMentions()). Here is an example of a complete code (in groovy), which shows how to get from positions to tokens:

class Example {
    public static void main(String[] args) {
        Properties props = new Properties();
        props.put("annotators", "tokenize, ssplit, pos, lemma, ner, parse, dcoref");
        props.put("dcoref.score", true);
        pipeline = new StanfordCoreNLP(props);
        Annotation document = new Annotation("The atom is a basic unit of matter, it   consists of a dense central nucleus surrounded by a cloud of negatively charged electrons.");

        pipeline.annotate(document);
        Map<Integer, CorefChain> graph = document.get(CorefChainAnnotation.class);

        println aText

        for(Map.Entry<Integer, CorefChain> entry : graph) {
          CorefChain c =   entry.getValue();                
          println "ClusterId: " + entry.getKey();
          CorefMention cm = c.getRepresentativeMention();
          println "Representative Mention: " + aText.subSequence(cm.startIndex, cm.endIndex);

          List<CorefMention> cms = c.getCorefMentions();
          println  "Mentions:  ";
          cms.each { it -> 
              print aText.subSequence(it.startIndex, it.endIndex) + "|"; 
          }         
        }
    }
}

Output (I do not understand where 's' comes from):

The atom is a basic unit of matter, it consists of a dense central nucleus surrounded by a cloud of negatively charged electrons.
ClusterId: 1
Representative Mention: he
Mentions: he|atom |s|
ClusterId: 6
Representative Mention:  basic unit 
Mentions:  basic unit |
ClusterId: 8
Representative Mention:  unit 
Mentions:  unit |
ClusterId: 10
Representative Mention: it 
Mentions: it |

Answer 2

These are the recent results from the annotator.

1. [1, 1] 1 The atom
2. [1, 2] 1 a basic unit of matter
3. [1, 3] 1 it
4. [1, 6] 6 negatively charged electrons
5. [1, 5] 5 a cloud of negatively charged electrons

The markings are as follows :

[Sentence number,'id']  Cluster_no  Text_Associated

The text belonging to the same cluster refers to the same context.

Answer 3

I've been working with the coreference dependency graph and I started by using the other answer to this question. After a while though I realized that this algorithm above is not exactly correct. The output it produced is not even close to the modified version I have.

For anyone else who uses this article, here is the algorithm I ended up with which also filters out self references because every representativeMention also mentions itself and a lot of mentions only reference themselves.

Map<Integer, CorefChain> coref = document.get(CorefChainAnnotation.class);

for(Map.Entry<Integer, CorefChain> entry : coref.entrySet()) {
    CorefChain c = entry.getValue();

    //this is because it prints out a lot of self references which aren't that useful
    if(c.getCorefMentions().size() <= 1)
        continue;

    CorefMention cm = c.getRepresentativeMention();
    String clust = "";
    List<CoreLabel> tks = document.get(SentencesAnnotation.class).get(cm.sentNum-1).get(TokensAnnotation.class);
    for(int i = cm.startIndex-1; i < cm.endIndex-1; i++)
        clust += tks.get(i).get(TextAnnotation.class) + " ";
    clust = clust.trim();
    System.out.println("representative mention: \"" + clust + "\" is mentioned by:");

    for(CorefMention m : c.getCorefMentions()){
        String clust2 = "";
        tks = document.get(SentencesAnnotation.class).get(m.sentNum-1).get(TokensAnnotation.class);
        for(int i = m.startIndex-1; i < m.endIndex-1; i++)
            clust2 += tks.get(i).get(TextAnnotation.class) + " ";
        clust2 = clust2.trim();
        //don't need the self mention
        if(clust.equals(clust2))
            continue;

        System.out.println("\t" + clust2);
    }
}

And the final output for your example sentence is the following:

representative mention: "a basic unit of matter" is mentioned by:
The atom
it

Usually "the atom" ends up being the representative mention but in the case it doesn't surprisingly. Another example with a slightly more accurate output is for the following sentence:

The Revolutionary War occurred during the 1700s and it was the first war in the United States.

produces the following output:

representative mention: "The Revolutionary War" is mentioned by:
it
the first war in the United States