Skip to content

Latest commit

 

History

History
549 lines (417 loc) · 21.8 KB

README.md

File metadata and controls

549 lines (417 loc) · 21.8 KB
Raw

CogComp Core-Utilities

cogcomp-core-utilities is a Java library that is designed to help programming NLP applications by providing a uniform representation of various NLP annotations of text (like parse trees, parts of speech, semantic roles, coreference, etc.)

Overview

This library provides basic useful functionality in Java. The goals of this library are:

  1. Augment the Java library with useful data structures and algorithms that can be used across many NLP projects.
  2. Add support for recurring experiment-related tasks like cross-validation and inter-annotator agreement.
  3. Provide other utility classes for reading files, interface to the shell, etc.

Functionality

  • Data structures
    • TextAnnotation. The library also provides an easy way to access the Curator.
    • Record (internal datastructure for Curator)
    • LBJava data structures
    • Pair and Triple classes
    • Trees, where the nodes can be arbitrary objects and a utility class to read trees from the bracket notation (like the Penn Treebank notation.)
    • Queryable list to support SQL like operations on the elements of the list
    • Bounded priority queue, to help with beam search
  • Experiment utilities
    • P/R/F1 reporting (see EvaluationRecord)
    • Statistical significance testing
    • Cross-validation helper
    • Android notification sender
  • Algorithms
    • Matching arbitrary lists with patterns
    • Levenstein distance
    • Longest common subsequence
    • Searching for patterns in trees
    • Replacing parts of trees that match a pattern
    • Graph search algorithms -- breadth first, depth first, uniform cost and beam.
  • IO
    • Corpus readers (CoNLL, PTB, Ontonotes, etc.)
    • Reading a file, one line at a time
    • Utility functions like mkdir, ls, etc
  • Transformers
    • A transformer defines a general purpose interface that transforms one object into another. This is used extensively in the project Edison. For example, any annotation that is performed on text can be thought of as the result of a transformer.
    • A special transformer is a Predicate, which transforms an object into a Boolean.
  • Search
    • Beam search
    • Breadth/Depth first search
    • Graph search
  • Miscellaneous utilities
    • ArgMax
    • Counter
    • A command line interface that uses Java reflection to expose static functions of a pre-defined class to the shell
    • And much more...

Examples and Clarification

This set of examples goes over the basics of the data structures. Recollect that different annotations over text are called Views, each of which is a graph of Constituents and Relations. The object that manages views corresponding to a single piece of text is called a TextAnnotation.

Creating a TextAnnotation

A TextAnnotation can be thought of as a container that stores different layers of annotations over some text.

  1. Using the LBJ sentence splitter and tokenizer:

    The simplest way to define a TextAnnotation is to just give the text to the constructor. Note that in the following example, text1 consists of three sentences. The corresponding ta1 will use the sentence slitter defined in the Learning Based Java (LBJava) library to split the text into sentences and further apply the LBJ tokenizer to tokenize the sentence.

    String text1 = "Good afternoon, gentlemen. I am a HAL-9000 "
  + "computer. I was born in Urbana, Il. in 1992";

String corpus = "2001_ODYSSEY";
String textId = "001";

// Create a TextAnnotation using the LBJ sentence splitter 
// and tokenizers. 
TextAnnotationBuilder tab = new TokenizerTextAnnotationBuilder(new IllinoisTokenizer());

TextAnnotation ta1 = tab.createTextAnnotation(corpus, textId, text1);

  2. Using pre-tokenized text:

    Quite often, our data source could specify the tokenization for text. We can use this to create a TextAnnotation by specifying the sentences and tokens manually. In this case, the input to the constructor consists of the corpus, text identifier and a List of strings. Each element in the list will be treated as a sentence. This constructor assumes that the sentences in the list are white-space tokenized.

    String corpus = "2001_ODYSSEY";
String textId2 = "002";

String[] sentence1 = {"Good",  "afternoon", ",", "gentlemen", "."};
String[] sentence2 = {"I", "am", "a", "HAL-9000", "computer", "."};

List<String[]> tokenizedSentences = Arrays.asList(sentence1, sentence2);
TextAnnotation ta2 = BasicTextAnnotationBuilder.createTextAnnotationFromTokens(
		corpus, textId2, tokenizedSentences);

Views

The library stores all information about a specific annotation over text in an object called View. A View is a graph, where the nodes are Constituents and the edges are Relations. In its most general sense, a View is a graph whose nodes are labeled spans of text. The edges between the nodes represent the relationships between them. A TextAnnotation can be thought of as a container of views, indexed by their names.

The tokens are not stored in a View. The TextAnnotation knows the tokens of the text and each Constituent of every view is defined in terms of the tokens. A constituent can represent zero tokens or spans.

Sentences are stored as a view. In the terminology above, the Constituents will correspond to the sentences. There are no Relations between them. (The ordering between the sentences is not explicitly represented because this can be inferred from the Constituents which refer to the tokens.) So the graph that this View represents is a degenerate graph, with only nodes and no edges.

This example shows how to use the View datastructure to create an arbitrary view.

String corpus = "2001_ODYSSEY";
String textId = "001";
String text1 = "Good afternoon, gentlemen. I am a HAL-9000 computer.";
	
TextAnnotation ta1 = new TextAnnotation(corpus, textId, text1);

View myView = new View("MyViewName", "MyViewGenerator", ta1, 0.121);
ta1.addView("MyViewName", myView);

Constituent m1 = new Constituent("M1", "MyViewName", ta1, 5, 6);
myView.addConstituent(m1);

Constituent m2 = new Constituent("M2", "MyViewName", ta1, 7, 10);
myView.addConstituent(m2);

Constituent m3 = new Constituent("M1", "MyViewName", ta1, 8, 9);
myView.addConstituent(m3);

Constituent m4 = new Constituent("M2", "MyViewName", ta1, 9, 10);
myView.addConstituent(m4);

Relation r1 = new Relation("Subject-Object", m1, m2, 0.001);
myView.addRelation(r1);

Relation r2 = new Relation("NameOf", m3, m4, 0.12);
myView.addRelation(r2);

System.out.println(myView.getConstituents());
System.out.println(myView.getRelations());

System.out.println(r1.getSource());
System.out.println(r2.getTarget());

Accessing the text and tokens

Edison keeps track of the raw text along with the tokens it contains. So, we can get the original text using the function getText() and also the tokenized text using the function getTokenizedText(). The function getToken(int tokenId) gives us the tokens in the text.

// Print the text. This prints the raw text that was used to
// create the TextAnnotation object. In the case where the
// second constructor is used, the text is printed whitespace
// tokenized.
System.out.println(ta1.getText());
System.out.println(ta2.getText());
 
// Print the tokenized text. The tokenization scheme is
// specified by the constructor, which in the first example
// defaults to the LBJ tokenizer, and in the second one is
// specified manually.
System.out.println(ta1.getTokenizedText());
System.out.println(ta2.getTokenizedText());
 
// Print the tokens
for (int i = 0; i < ta.size(); i++) {
    System.out.print(i + ":" + ta.getToken(i) + "\t");
}
System.out.println();

Accessing sentences

Each TextAnnotation knows the views it contains. To get these, use the function getAvailableViews(), which returns a set of strings representing the names of the views it contains.

The following code prints all the available views in the TextAnnotation ta1 defined above. It then goes over each sentence and prints them.

System.out.println(ta1.getAvailableViews());
 
// Print the sentences. The Sentence class has many of the same
// methods as a TextAnnotation.
List<Sentence> sentences = ta1.sentences();
 
System.out.println(sentences.size() + " sentences found.");
 
for (int i = 0; i < sentences.size(); i++) {
    Sentence sentence = sentences.get(i);
    System.out.println(sentence);
}

Accessing Constituents

This example gets all the shallow parse constituents. In the shallow parse constituent, each chunk will have one constituent. There are no relations between the chunks.

SpanLabelView shallowParseView = (SpanLabelView) ta
                .getView(ViewNames.SHALLOW_PARSE);
                
List<Constituent> shallowParseConstituents = shallowParseView
                .getConstituents();
for (Constituent c : shallowParseConstituents) {
    System.out.println(c.getStartSpan() + "-" + c.getEndSpan() + ":"
            + c.getLabel() + " " + c.getSurfaceString());
}

Creating complex features

One can combine the simple datastructures introduced so far and create relatively complex features. Here we create features by combination of edge labels for dependency parsing and named-entity recognition.

SpanLabelView ne = (SpanLabelView) ta.getView(ViewNames.NER);
TreeView dependency = (TreeView) ta.getView(ViewNames.DEPENDENCY);

System.out.println(dependency);
System.out.println(ne);

for (Constituent neConstituent : ne.getConstituents()) {
    List<Constituent> depConstituents = (List<Constituent>) dependency
            .where(Queries.containedInConstituent(neConstituent));

    for (Constituent depConstituent : depConstituents) {
        System.out.println("Outgoing relations");

        for (Relation depRel : depConstituent.getOutgoingRelations()) {
            System.out.println("\t" + neConstituent + "--"
                    + depRel.getRelationName() + "--> "
                    + depRel.getTarget());
        }

        System.out.println("Incoming relations");

        for (Relation depRel : depConstituent.getIncomingRelations()) {
            System.out
                    .println("\t" + depRel.getSource() + "--"
                            + depRel.getRelationName() + "--> "
                            + neConstituent);
        }
    }
}

Creating AnnotatorService

AnnotatorService is our super-wrapper that provides access to different annotations and free caching. Currently we have two classes implementing AnnotatorService:

  1. cogcomp-curator
  2. nlp-pipeline

The image below describes the different ways of creating TextAnnotation objects from either tokenized or raw text.

schema 001

Below is an example of how to use PipelineFactory to create new annotations.

using edu.illinois.cs.cogcomp.pipeline.main.PipelineFactory;

AnnotatorService annotator = PipelineFactory.buildPipeline();

// Or alternatively to use the curator:
// using edu.illinois.cs.cogcomp.curator.CuratorFactory;
// AnnotatorService annotator = CuratorFactory.buildCuratorClient();

and then create a TextAnnotation component and add the Views you need:

TextAnnotation ta = annotator.createBasicTextAnnotation(corpusID, taID, "Some text that I want to process.");

Of course the real fun begins now! Using AnnotatorService you can add different annotation Views using their canonical name:

annotator.addView(ta, ViewNames.POS);
annotator.addView(ta, ViewNames.NER_CONLL);

These Views as well as the TextAnnotation object are now locally cached for faster future access.

You can later print the existing views:

System.out.println(ta1.getAvailableViews());

or access the views them directly:

TokenLabelView posView = (TokenLabelView) ta.getView(ViewNames.POS);

for (int i = 0; i < ta.size(); i++) {
    System.out.println(i + ":" + posView.getLabel(i));
}

Creating Annotators

The AnnotatorService class is based on stringing together classes that extend the Annotator abstract class. This class is used within the project to wrap the Illinois POS, Chunker, Lemmatizer, and NER.

In your extension of Annotator you need to define the following methods:

addView(TextAnnotation ta)

This method is invoked by Annotator.getView(TextAnnotation ta), and does the actual work of creating a View object, populating it with Contituent and Relation objects as appropriate, and adding it to the TextAnnotation argument. This method should first check that the argument contains all the necessary prerequisite views (the names of these views are also required by the Annotator constructor) and throw an exception if they are not.

initialize(ResourceManager rm)

This method should read configuration parameters (including paths from which to load resources) from the ResourceManager argument. Your class constructor will provide a ResourceManager object as an argument to the Annotator constructor. This is used for lazy initialization, if active, in which case the first call to getView() will call initialize() with this configuration.

One key configuration default, AnnotatorConfigurator.IS_SENTENCE_LEVEL indicates whether an annotator requires context beyond a single sentence. For example, when the POS tagger predicts a label for a word, it uses only information from the same sentence. This means that a document can be split up into sentences, which can then be processed one at a time by the POS annotator, without degrading accuracy. NER, on the other hand, has features that depend on a context that may extend into previous sentences: if the sentence-by-sentence approach is used for NER, some degradation is to be expected.

Configurators

For ease of use of your own NLP software, especially classes that extend Annotator or AnnotatorService, you are encouraged to create a Configurator class that specifies the relevant configuration flags and their default values.

The POS, Chunker, Lemmatizer and NER modules all have their own extension of the Configurator class for this purpose. This makes it easy to specify a default constructor for your annotator, and to specify only non-default configuration options when you instantiate one or more classes that use ResourceManager and Configurator to manage configuration options.

A configurator must inherit from edu.illinois.cs.cogcomp.core.utilities.configuration.Configurator.java and override the getDefaultConfig() method. Specify any configuration parameters as public static edu.illinois.cs.cogcomp.core.utilities.configuration.Property fields in the Configurator subclass you create. This specifies configuration parameter keys and default values, and makes the names available to clients.

The getDefaultConfig() method should simply populate a ResourceManager object with all default values -- there are convenience methods in Configurator to make this easier.

When implementing your Annotator's simplest constructor, instantiate your configurator subclass and call its getDefaultConfig() method. Implement a second Constructor that takes a ResourceManager as its argument, instantiate your configurator, and call its getConfig(ResourceManager) method.

Here's the constructor code from ChunkAnnotator:

    public ChunkerAnnotator(boolean lazilyInitialize) {
        this(lazilyInitialize, new ChunkerConfigurator().getDefaultConfig());
    }

    public ChunkerAnnotator(boolean lazilyInitialize, ResourceManager rm) {
        super(ViewNames.SHALLOW_PARSE, new String[] {ViewNames.POS}, lazilyInitialize, new ChunkerConfigurator().getConfig(rm));
    }

This allows you to only specify configuration parameters you wish to modify from their default values, using the keys specified in the relevant Configurator's Property fields. You can either directly populate a java Properties object with these key/value pairs, then instantiate a ResourceManager and call your Configurator's getConfig() method with that; or you can write these non-default values into a text file with each key/value pair on a new line, with the key and value separated by a tab character.

For the Chunker example, suppose you want to change the model path. In ChunkerConfigurator this is specified as:

    public static final Property MODEL_PATH = new Property("modelPath", MODEL_DIR_PATH.value
            + MODEL_NAME.value + ".lc");

To directly populate a Properties object:

    import edu.illinois.cs.cogcomp.core.resources.ResourceManager;
    import java.util.Properties;
    import edu.illinois.cs.cogcomp.chunker.main.ChunkerAnnotator;

    Properties props = new Properties;
    props.setProperty(ChunkerConfigurator.MODEL_PATH.key, "/some/other/path");
    ChunkerAnnotator ca = new ChunkerAnnotator(new ResourceManager(props));

To use a text file instead, create a text file (for this example, "config/altChunker.config") with the single line:

modelPath	/some/other/path

...and use it to instantiate a ResourceManager object:

    ChunkerAnnotator ca = new ChunkerAnnotator(new ResourceManager("config/altChunker.config"));

Serialization and Deserialization

To store TextAnnotation objects on disk, serialization/deserialization is supported in the following formats:

  • Binary Serialization: Binary serialization using Apache Common's SerializationUtils to serialize the TextAnnotation class.
import edu.illinois.cs.cogcomp.core.utilities.SerializationHelper;

// Serialize and save to file
SerializationHelper.serializeTextAnnotationToFile(ta, "text_annotation.bin", true);

// Read file from disk and deserialize
TextAnnotation ta = SerializationHelper.deserializeTextAnnotationFromFile("text_annotation.bin")
  • JSON: Lightweight human-readable data-interchange format.
import edu.illinois.cs.cogcomp.core.utilities.SerializationHelper;

String jsonString = SerializationHelper.serializeToJson(ta);

TextAnnotation ta = SerializationHelper.deserializeFromJson(jsonString);
import edu.illinois.cs.cogcomp.core.utilities.SerializationHelper;

// Serialize and save to file
SerializationHelper.serializeTextAnnotationToProtobuf(ta, "text_annotation.buf", true);

// Read file from disk and deserialize
TextAnnotation ta = SerializationHelper.deserializeFromProtobuf("text_annotation.buf");

More usage information in the SerializationHelper class.

Generating Protocol Buffer Java Code

Note: If you make any change to TextAnnotation class which involves adding/removing data items, make sure to update the protocol buffer schema and the corresponding serialization code accordingly.

Install the Protocol Buffer compiler locally.

On macOS, you can install the compiler using Homebrew: brew install protobuf.

Run the following commands from the repository root:

protoc --java_out=core-utilities/src/main/java core-utilities/src/main/proto/TextAnnotation.proto

mvn license:format

StringTransformation

The StringTransformation class helps users processing noisy or marked-up text to track changes in character offsets, and to map any annotations developed on the cleaned up text back to the original source. StringTransformation keeps a copy of an original text, a current text, and a set of pending edits. It accepts edits specified as a span of the current text and a String to replace that span (which may be the empty string in the case of deletion). Pending edits can be applied on demand, and are automatially applied when the current text is retrieved. Clients can specify offsets in the original String and find their counterparts in the modified text, or vice versa. This allows users to clean up text, process it with NLP compoments, then map annotations back to the original text. Such functionality is useful in NLP tasks where evaluation scripts require annotations to be specified in terms of the original source text. Corpus reader components such as XmlDocumentProcessor use StringTransformation to track changes made when xml elements are stripped; StringTransformationCleanup applies common text cleanup (removing non-standard characters, punctuation errors, etc.) but allows users to map any annotations acquired from the cleaned-up text to the original text. See StringTransformationText for some examples.

TextAnnotationUtilities

The TextAnnotationUtilities class provides helper functions for copying constituents and views between TextAnnotations. In particular, it supports excerpting annotations for sentences or other subspans of some TextAnnotation's underlying text into a shorter, self-contained TextAnnotation, or mapping from a sentence TextAnnotation to one for a longer span containing that sentence.

There is also support for StringTransformation, mapping a TextAnnotation associated with a transformed text to an equivalent TextAnnotation over the original text.

Citation

If you use this code in your research, please provide the URL for this github repository in the relevant publications.
If you use any of the NLP modules, please check their README files to see if there are relevant publications to cite.