Resolve coreference using Stanford CoreNLP
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, Thread.currentThread().getContextClassLoader()); } 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 { IOUtils.closeIgnoringExceptions(in); } 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(corefSystem.dictionaries(), corefSystem.semantics()); } catch (Exception e) { System.err.println("ERROR: cannot create DeterministicCorefAnnotator!"); e.printStackTrace(); throw new RuntimeException(e); }
Annotation
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).
Annotation CoreAnnotations.SentencesAnnotation -> sentences CoreAnnotations.TokensAnnotation -> tokens TreeCoreAnnotations.TreeAnnotation -> Tree SemanticGraphCoreAnnotations.CollapsedDependenciesAnnotation -> SemanticGraph
Tokens
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(); token.setWord(word); token.setTag(tag); token.setNer(ner); ... tokens.add(token); } 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 .newGrammaticalStructure(tree); 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
ann.set( CollapsedDependenciesAnnotation.class, 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 .get(CoreAnnotations.TokensAnnotation.class); sentences.add(tokens); Tree tree = sentence.get(TreeCoreAnnotations.TreeAnnotation.class); trees.add(tree); MentionExtractor.mergeLabels(tree, tokens); MentionExtractor.initializeUtterance(tokens); } } else { System.err .println("ERROR: this coreference resolution system requires SentencesAnnotation!"); return; } // 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( annotation, sentences, trees, allUnprocessedMentions); 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 .getLinks(result); if (VERBOSE) { System.err.printf("Found %d coreference links:\n", links.size()); for (Pair<IntTuple, IntTuple> link : links) { System.err.printf( "LINK (%d, %d) -> (%d, %d)\n", link.first.get(0), link.first.get(1), link.second.get(0), link.second.get(1)); } } // // 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) continue; Set<CoreLabel> coreferentTokens = Generics.newHashSet(); for (CorefMention mention : corefChain.getMentionsInTextualOrder()) { CoreMap sentence = annotation.get( CoreAnnotations.SentencesAnnotation.class).get( mention.sentNum - 1); CoreLabel token = sentence.get( CoreAnnotations.TokensAnnotation.class).get( mention.headIndex - 1); coreferentTokens.add(token); } for (CoreLabel token : coreferentTokens) { token.set( CorefCoreAnnotations.CorefClusterAnnotation.class, coreferentTokens); } } } } catch (RuntimeException e) { throw e; } catch (Exception e) { throw new RuntimeException(e); } }
Reference: | Resolve coreference using Stanford CoreNLP from our JCG partner Yifan Peng at the PGuru blog. |
Doesn’t work