The edu.arizona.sista.processors package aims to be a one-stop place for natural language (NL) processors.
We currently provide two APIs: one for
Stanford's CoreNLP, and one for a faster processor (FastNLPProcessor)
that cherry picks fast components from multiple sources (Stanford and MaltParser).
Both CoreNLPProcessor and FastNLPProcessor now include a full-fledged Rhetorical Structure Theory (RST) discourse parser. The version in CoreNLPProcessor relies on constituent syntax, whereas the one in FastNLPProcessor uses dependency syntax. The latter is marginally worse (~2 F1 points lower for the complete task) but it is two orders of magnitude faster.
Furthermore, this code contains a machine learning (ML) package (edu.arizona.sista.learning), which includes implementations for common ML algorithms (e.g., Perceptron, Logistic Regression, Support Vector Machines, Random Forests) for both classification and ranking.
This software requires Java 1.6, Scala 2.9 or higher, and CoreNLP 1.3.4 or higher.
This code is licensed under Apache License Version 2.0. However, some of the libraries used here, most notably CoreNLP, are GPL.
(c) Mihai Surdeanu, 2013 -
Contributors: Peter Jansen, Daniel Fried
- 3.2 - Added a discourse parser to
FastNLPProcessor. This performs marginally worse than the one inCoreNLPProcessor, but it is much faster for end-to-end processing, due to the shift-reduce syntactic parser. - 3.1 - Minimal functionality added to the learning package. Changed to CoreNLP 3.3.1.
- 3.0 - Added a RST discourse parser to
CoreNLPProcessor. Added theedu.arizona.sista.learningpackage. Utils classes are now underedu.arizona.sista.utilsrather thanedu.arizona.sista.processors.utils. - 2.2 - Various bug fixes. Added support for basic dependencies to
CoreNLPProcessor. - 2.1 - Bug fix in FastNLPProcessor: better root detection algorithm, robust to malt inconsistencies.
- 2.0 - We now support two processors:
CoreNLPProcessorandFastNLPProcessor. Changed the package name frome.a.s.processortoe.a.s.processors. Added Java usage example to README. Updated to CoreNLP 3.3.0. Added better unit tests to check for thread safetiness. - 1.5 - Bug fixing. Made the string interning process (see
Processor.in) local to each thread to avoid concurrency issues in multi-threaded programs. Added new unit tests. Added minor functionality to Lexicon. - 1.4 - Code cleanup. Added some minor new functionality such as finding base NPs in the Trees class.
- 1.3 - Reverted back to the
1.xversion numbers, since we will add other software here not just CoreNLP. Added correct mvn dependencies for the CoreNLP jars. Removed theinstall*.shscripts, which are no longer needed. - 3.2.0 - Updated to Scala 2.10.1 and CoreNLP 3.2.0. Changed versioning system to be identical to CoreNLP's, so it's clear which CoreNLP version is used.
- 1.0 - Initial release
This software is available on maven as well. Add the following dependencies to your pom.xml to use it:
<dependency>
<groupId>edu.arizona.sista</groupId>
<artifactId>processors</artifactId>
<version>3.2</version>
</dependency>
<dependency>
<groupId>edu.arizona.sista</groupId>
<artifactId>processors</artifactId>
<version>3.2</version>
<classifier>models</classifier>
</dependency>
- Simple API - the APIs provided are, at least in my opinion, simpler than those provided for the original code. For example, when using CoreNLP you won't have to deal with hash maps that take class objects as keys. Instead, we use mostly arrays of integers or strings, which are self explanatory.
- Memory efficient - arrays are more memory efficient than hash maps. Furthermore, we used our own implementation to intern strings (i.e., avoiding to store duplicated strings multiple times). Due to these changes, I measured up to 99% decrease in memory for the annotations corresponding to a typical natural language text, when compared to the original CoreNLP code. (Note: this reduction takes effect only after CoreNLP finishes its work.)
- Faster access - again, because we use arrays instead of hash maps, access to the NL annotations (once constructed) is considerably faster than in the original Stanford code.
- Tool-independent API - we support multiple NL and machine learning tools. If you use this code, you will have to change your code only minimally (i.e., only the constructor for the
Processorobject). - Discourse parsing - we include a complete RST parser; simply instantiate
CoreNLPProcessorwithwithDiscourse = true.
This is a standard Maven project, so use the mvn package command to build the jar file, which will be stored in the target/ directory,e.g., target/processors-3.0.jar. Run the scripts/mk_model_jar script to build the models jar file, which will stored under the same target/ directory, e.g., target/processors-3.0-models.jar.
.
Add these generated jar files to your $CLASSPATH, along with the other necessary dependency jars. Take a look at scripts/run to see which dependencies are necessary at runtime.
Most of the examples here use Scala. However, this software can be used as is from Java as well! Scroll down towards the end of this document to see a Java usage example.
// create the processor
val proc:Processor = new CoreNLPProcessor(withDiscourse = true)
// the actual work is done here
val doc = proc.annotate("John Smith went to China. He visited Beijing, on January 10th, 2013.")
// you are basically done. the rest of this code simply prints out the annotations
// let's print the sentence-level annotations
var sentenceCount = 0
for (sentence <- doc.sentences) {
println("Sentence #" + sentenceCount + ":")
println("Tokens: " + sentence.words.mkString(" "))
println("Start character offsets: " + sentence.startOffsets.mkString(" "))
println("End character offsets: " + sentence.endOffsets.mkString(" "))
// these annotations are optional, so they are stored using Option objects, hence the foreach statement
sentence.lemmas.foreach(lemmas => println("Lemmas: " + lemmas.mkString(" ")))
sentence.tags.foreach(tags => println("POS tags: " + tags.mkString(" ")))
sentence.entities.foreach(entities => println("Named entities: " + entities.mkString(" ")))
sentence.norms.foreach(norms => println("Normalized entities: " + norms.mkString(" ")))
sentence.dependencies.foreach(dependencies => {
println("Syntactic dependencies:")
val iterator = new DirectedGraphEdgeIterator[String](dependencies)
while(iterator.hasNext) {
val dep = iterator.next
// note that we use offsets starting at 0 (unlike CoreNLP, which uses offsets starting at 1)
println(" head:" + dep._1 + " modifier:" + dep._2 + " label:" + dep._3)
}
})
sentence.syntacticTree.foreach(tree => {
println("Constituent tree: " + tree)
// see the edu.arizona.sista.utils.Tree class for more information
// on syntactic trees, including access to head phrases/words
})
sentenceCount += 1
println("\n")
}
// let's print the coreference chains
doc.coreferenceChains.foreach(chains => {
for (chain <- chains.getChains) {
println("Found one coreference chain containing the following mentions:")
for (mention <- chain) {
// note that all these offsets start at 0 too
println("\tsentenceIndex:" + mention.sentenceIndex +
" headIndex:" + mention.headIndex +
" startTokenOffset:" + mention.startOffset +
" endTokenOffset:" + mention.endOffset +
" text: " + doc.sentences(mention.sentenceIndex).words.slice(mention.startOffset, mention.endOffset).mkString("[", " ", "]"))
}
}
})
// let's print the discourse tree
doc.discourseTree.foreach(dt => {
println("Document-wide discourse tree:")
println(dt.toString())
})
The above code generates the following output:
Sentence #0:
Tokens: John Smith went to China .
Start character offsets: 0 5 11 16 19 24
End character offsets: 4 10 15 18 24 25
Lemmas: John Smith go to China .
POS tags: NNP NNP VBD TO NNP .
Named entities: PERSON PERSON O O LOCATION O
Normalized entities: O O O O O O
Syntactic dependencies:
head:1 modifier:0 label:nn
head:2 modifier:1 label:nsubj
head:2 modifier:4 label:prep_to
Constituent tree: (ROOT (S (NP (NNP John) (NNP Smith)) (VP (VBD went) (PP (TO to) (NP (NNP China)))) (. .)))
Sentence #1:
Tokens: He visited Beijing , on January 10th , 2013 .
Start character offsets: 26 29 37 44 46 49 57 61 63 67
End character offsets: 28 36 44 45 48 56 61 62 67 68
Lemmas: he visit Beijing , on January 10th , 2013 .
POS tags: PRP VBD NNP , IN NNP JJ , CD .
Named entities: O O LOCATION O O DATE DATE DATE DATE O
Normalized entities: O O O O O 2013-01-10 2013-01-10 2013-01-10 2013-01-10 O
Syntactic dependencies:
head:1 modifier:0 label:nsubj
head:1 modifier:2 label:dobj
head:1 modifier:8 label:tmod
head:2 modifier:5 label:prep_on
head:5 modifier:6 label:amod
Constituent tree: (ROOT (S (NP (PRP He)) (VP (VBD visited) (NP (NP (NNP Beijing)) (, ,) (PP (IN on) (NP (NNP January) (JJ 10th))) (, ,)) (NP-TMP (CD 2013))) (. .)))
Found one coreference chain containing the following mentions:
sentenceIndex:1 headIndex:2 startTokenOffset:2 endTokenOffset:3 text: [Beijing]
Found one coreference chain containing the following mentions:
sentenceIndex:1 headIndex:0 startTokenOffset:0 endTokenOffset:1 text: [He]
sentenceIndex:0 headIndex:1 startTokenOffset:0 endTokenOffset:2 text: [John Smith]
Found one coreference chain containing the following mentions:
sentenceIndex:1 headIndex:5 startTokenOffset:5 endTokenOffset:9 text: [January 10th , 2013]
Found one coreference chain containing the following mentions:
sentenceIndex:0 headIndex:4 startTokenOffset:4 endTokenOffset:5 text: [China]
Document-wide discourse tree:
elaboration (LeftToRight)
TEXT:John Smith went to China .
TEXT:He visited Beijing , on January 10th , 2013 .
For more details about the annotation data structures, please see the edu/arizona/sista/processor/Document.scala file.
Changing processors is trivial: just replace the first line in the above example with:
val proc:Processor = new FastNLPProcessor()
// everything else stays the same
FastNLPProcessor uses the Stanford tokenizer, POS tagger, and NER, but replaces its parser with maltparser, trained to generated Stanford "basic" (rather than "collapsed") dependencies. This means that this annotator does not produce constituent trees and coreference chains. However, because of the faster parser, you should see a speedup increase of at least one order of magnitude. The output of the above code with FastNLPProcessor is:
Sentence #0:
Tokens: John Smith went to China .
Start character offsets: 0 5 11 16 19 24
End character offsets: 4 10 15 18 24 25
Lemmas: John Smith go to China .
POS tags: NNP NNP VBD TO NNP .
Named entities: PERSON PERSON O O LOCATION O
Normalized entities: O O O O O O
Syntactic dependencies:
head:1 modifier:0 label:nn
head:2 modifier:1 label:nsubj
head:2 modifier:3 label:prep
head:2 modifier:5 label:punct
head:3 modifier:4 label:pobj
Sentence #1:
Tokens: He visited Beijing , on January 10th , 2013 .
Start character offsets: 26 29 37 44 46 49 57 61 63 67
End character offsets: 28 36 44 45 48 56 61 62 67 68
Lemmas: he visit Beijing , on January 10th , 2013 .
POS tags: PRP VBD NNP , IN NNP JJ , CD .
Named entities: O O LOCATION O O DATE DATE DATE DATE O
Normalized entities: O O O O O 2013-01-10 2013-01-10 2013-01-10 2013-01-10 O
Syntactic dependencies:
head:1 modifier:0 label:nsubj
head:1 modifier:2 label:dobj
head:1 modifier:3 label:punct
head:1 modifier:4 label:prep
head:1 modifier:9 label:punct
head:4 modifier:5 label:pobj
head:5 modifier:6 label:amod
head:5 modifier:7 label:punct
head:5 modifier:8 label:num
val doc = proc.annotateFromSentences(List("John Smith went to China.", "He visited Beijing."))
// everything else stays the same
val doc = annotateFromTokens(List(
List("John", "Smith", "went", "to", "China", "."),
List("There", ",", "he", "visited", "Beijing", ".")))
// everything else stays the same
You can of course use only some of the annotators provided by CoreNLP by calling them individually. To illustrate,
the Processor.annotate() method is implemented as follows:
def annotate(doc:Document): Document = {
tagPartsOfSpeech(doc)
lemmatize(doc)
recognizeNamedEntities(doc)
parse(doc)
chunking(doc)
labelSemanticRoles(doc)
resolveCoreference(doc)
discourse(doc)
doc.clear()
doc
}
(Note that CoreNLP currently does not support chunking and semantic role labeling.) You can use just a few of these annotators. For example, if you need just POS tags, lemmas and named entities, you could use the following code:
val doc = proc.mkDocument("John Smith went to China. He visited Beijing, on January 10th, 2013.")
proc.tagPartsOfSpeech(doc)
proc.lemmatize(doc)
proc.recognizeNamedEntities(doc)
doc.clear()
Note that the last method called (doc.clear()) clears the internal structures created by the actual CoreNLP annotators.
This saves a lot of memory, so, although it is not strictly necessary, I recommend you call it.
This package also offers serialization code for the generated annotations. The two scenarios currently supported are:
// saving to a PrintWriter, pw
val someAnnotation = proc.annotate(someText)
val serializer = new DocumentSerializer
serializer.save(someAnnotation, pw)
// loading from an InputStream, is
val someAnnotation = serializer.load(is)
// saving to a String object, savedString
val someAnnotation = proc.annotate(someText)
val serializer = new DocumentSerializer
val savedString = serializer.save(someAnnotation)
// loading from a String object, fromString
val someAnnotation = serializer.load(fromString)
Note that space required for these serialized annotations is considerably smaller (8 to 10 times) than the corresponding serialized Java objects. This is because we store only the information required to recreate these annotations (e.g., words, lemmas, etc.) without storing any of the Java/Scala objects and classes.
Classes that implement the Processor trait intern String objects to avoid allocating memory repeatedly for the same string.
This is implemented by maintaining an internal dictionary of strings previously seen. This dictionary is unlikely to
use a lot of memory due to the Zipfian distribution of language. But, if memory usage is a big concern, it can be cleaned by
calling:
Processor.in.clear()
I recommend you do this only after you annotated all the documents you plan to keep in memory.
Although I see no good reason for doing this, you can disable the interning of strings completely by setting the internStrings = false in the
CoreNLProcessor constructor, as such:
val processor = new CoreNLPProcessor(internStrings = false)
Scala is (largely) compatible with Java, so this library can be directly used from Java. Below is Java code that translates most of the functionality from the first Scala example in this document to Java:
public class ProcessorJavaExample {
public static void main(String [] args) throws Exception {
// create the processor
Processor proc = new CoreNLPProcessor(true);
// for much faster processing, use FastNLPProcessor
// Processor proc = new FastNLPProcessor(true);
// the actual work is done here
Document doc = proc.annotate("John Smith went to China. He visited Beijing, on January 10th, 2013.");
// you are basically done. the rest of this code simply prints out the annotations
// let's print the sentence-level annotations
int sentenceCount = 0;
for (Sentence sentence: doc.sentences()) {
System.out.println("Sentence #" + sentenceCount + ":");
System.out.println("Tokens: " + mkString(sentence.words(), " "));
System.out.println("Start character offsets: " + mkString(sentence.startOffsets(), " "));
System.out.println("End character offsets: " + mkString(sentence.endOffsets(), " "));
// these annotations are optional, so they are stored using Option objects, hence the isDefined checks
if(sentence.lemmas().isDefined()){
System.out.println("Lemmas: " + mkString(sentence.lemmas().get(), " "));
}
if(sentence.tags().isDefined()){
System.out.println("POS tags: " + mkString(sentence.tags().get(), " "));
}
if(sentence.entities().isDefined()){
System.out.println("Named entities: " + mkString(sentence.entities().get(), " "));
}
if(sentence.norms().isDefined()){
System.out.println("Normalized entities: " + mkString(sentence.norms().get(), " "));
}
if(sentence.dependencies().isDefined()) {
System.out.println("Syntactic dependencies:");
DirectedGraphEdgeIterator<String> iterator = new
DirectedGraphEdgeIterator<String>(sentence.dependencies().get());
while(iterator.hasNext()) {
// this is a scala Tuple3
Tuple3<Object, Object, String> dep = iterator.next();
// note that we use offsets starting at 0 (unlike CoreNLP, which uses offsets starting at 1)
System.out.println(" head:" + dep._1() + " modifier:" + dep._2() + " label:" + dep._3());
}
}
if(sentence.syntacticTree().isDefined()) {
System.out.println("Constituent tree: " + sentence.syntacticTree().get());
// see the edu.arizona.sista.struct.Tree class for more information
// on syntactic trees, including access to head phrases/words
}
sentenceCount += 1;
System.out.println("\n");
}
// let's print the coreference chains
if(doc.coreferenceChains().isDefined()) {
// these are scala.collection Iterator and Iterable (not Java!)
Iterator<Iterable<CorefMention>> chains = doc.coreferenceChains().get().getChains().iterator();
while(chains.hasNext()) {
Iterator<CorefMention> chain = chains.next().iterator();
System.out.println("Found one coreference chain containing the following mentions:");
while(chain.hasNext()) {
CorefMention mention = chain.next();
// note that all these offsets start at 0 too
System.out.println("\tsentenceIndex:" + mention.sentenceIndex() +
" headIndex:" + mention.headIndex() +
" startTokenOffset:" + mention.startOffset() +
" endTokenOffset:" + mention.endOffset());
}
}
}
}
public static String mkString(String [] sa, String sep) {
StringBuilder os = new StringBuilder();
for(int i = 0; i < sa.length; i ++) {
if(i > 0) os.append(sep);
os.append(sa[i]);
}
return os.toString();
}
public static String mkString(int [] sa, String sep) {
StringBuilder os = new StringBuilder();
for(int i = 0; i < sa.length; i ++) {
if(i > 0) os.append(sep);
os.append(Integer.toString(sa[i]));
}
return os.toString();
}
}
The output of this code is:
Sentence #0:
Tokens: John Smith went to China .
Start character offsets: 0 5 11 16 19 24
End character offsets: 4 10 15 18 24 25
Lemmas: John Smith go to China .
POS tags: NNP NNP VBD TO NNP .
Named entities: PERSON PERSON O O LOCATION O
Normalized entities: O O O O O O
Syntactic dependencies:
head:1 modifier:0 label:nn
head:2 modifier:1 label:nsubj
head:2 modifier:4 label:prep_to
Constituent tree:
(ROOT
(S
(NP
(NNP John)
(NNP Smith)
)
(VP
(VBD went)
(PP
(TO to)
(NP
(NNP China)
)
)
)
(. .)
)
)
Sentence #1:
Tokens: He visited Beijing , on January 10th , 2013 .
Start character offsets: 26 29 37 44 46 49 57 61 63 67
End character offsets: 28 36 44 45 48 56 61 62 67 68
Lemmas: he visit Beijing , on January 10th , 2013 .
POS tags: PRP VBD NNP , IN NNP JJ , CD .
Named entities: O O LOCATION O O DATE DATE DATE DATE O
Normalized entities: O O O O O 2013-01-10 2013-01-10 2013-01-10 2013-01-10 O
Syntactic dependencies:
head:1 modifier:0 label:nsubj
head:1 modifier:2 label:dobj
head:1 modifier:8 label:tmod
head:2 modifier:5 label:prep_on
head:5 modifier:6 label:amod
Constituent tree:
(ROOT
(S
(NP
(PRP He)
)
(VP
(VBD visited)
(NP
(NP
(NNP Beijing)
)
(, ,)
(PP
(IN on)
(NP
(NNP January)
(JJ 10th)
)
)
(, ,)
)
(NP-TMP
(CD 2013)
)
)
(. .)
)
)
Found one coreference chain containing the following mentions:
sentenceIndex:1 headIndex:2 startTokenOffset:2 endTokenOffset:7
Found one coreference chain containing the following mentions:
sentenceIndex:0 headIndex:1 startTokenOffset:0 endTokenOffset:2
sentenceIndex:1 headIndex:0 startTokenOffset:0 endTokenOffset:1
Found one coreference chain containing the following mentions:
sentenceIndex:1 headIndex:5 startTokenOffset:5 endTokenOffset:7
Found one coreference chain containing the following mentions:
sentenceIndex:0 headIndex:4 startTokenOffset:4 endTokenOffset:5
Found one coreference chain containing the following mentions:
sentenceIndex:1 headIndex:8 startTokenOffset:8 endTokenOffset:9
The discourse parser in processors is inspired by the parser of Feng and Hirst and the HILDA parser of Hernault et al., but with a different feature set.
It is transparently integrated in both CoreNLPProcessor and FastNLPProcessor: just instantiate it as CoreNLPProcessor(withDiscourse = true) or FastNLPProcessor(withDiscourse = true). If discourse is enabled, Document.discourseTree stores the discourse tree for the entire document as an instance of the DiscourseTree class.
Following the conventions from other modern discourse parsing work, the discourse tree:
- Is represented as a binary tree, containing hypotactic relations (containing one nucleus and one satellite node) or paratactic relations (both nodes have equal importance).
- Stores the relation labels in the parent node (in
DiscourseTree.relationLabel) rather than the satellite nodes (like the RST corpus). We use the same 18 labels as Feng and Hirst. - Stores the relation direction in
DiscourseTree.relationDirection. The direction can beLeftToRight(meaning the nucleus is the left child),RightToLeft(the right node is the nucleus), orNone(for paratactic relations).
If you end up using this discourse parser, I would appreciate it if you cited this work:
Peter Jansen, Mihai Surdeanu, and Peter Clark. Discourse Complements Lexical Semantics for Non-factoid Answer Reranking. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (ACL), 2014. [bib]
processors now contains a machine learning (ML) package (edu.arizona.sista.learning), which includes implementations for common ML algorithms (e.g., Perceptron, Logistic Regression, Support Vector Machines, Random Forests) for both classification and ranking.
The structure of this package is heavily inspired by Stanford's CoreNLP. Similar to CoreNLP, we use a Datum trait to store a single data point, which is implemented by BVFDatum to store binary-valued-feature datums, or by RVFDatum to store real-valued-feature datums. A collection of data points is stored as a Dataset, which is similarly implemented by BVFDataset or RVFDataset. All classifiers implement the Classifier trait, which has three main methods: train, which trains a model a given dataset, classOf, which returns the most likely prediction for a given datum, and scoresOf, which returns the scores for all known labels for a given datum. We currently support the following classifiers: large-margin Perceptron (PerceptronClassifier), linear SVMs and logistic regression from liblinear (LibLinearClassifier), dual SVMs from libsvm (LibSVMClassifier), and random forests from fast-random-forest (RandomForestClassifier).
A similar structure exists for ranking problems, with RankingDataset used to store a corpus of ranking examples, and RankingClassifier as the API to be implemented by all ranking classifiers. We currently support the following classifiers: ranking Perceptron (PerceptronRankingClassifier), ranking SVMs from svm-rank (SVMRankingClassifier), and boosted decision trees from jforests (JForestsRankingClassifier).
For usage examples, including how to create datums and datasets from scratch or import them from the svm-light format, please take a look at the examples under src/test/scala/edu/arizona/sista/learning.