Skip to content

Latest commit

 

History

History
45 lines (32 loc) · 3.37 KB

README.md

File metadata and controls

45 lines (32 loc) · 3.37 KB
Raw

CogComp-DatalessClassifier

Given a label ontology, and textual descriptions of those labels, Dataless-Classifier is capable of classifying arbitrary text into that ontology.

It is particularly useful in those scenarios where it is difficult/expensive to gather enough training data to train a supervised text classifier. Dataless-Classifier utilizes the semantic meaning of the labels to bypass the need for explicit supervision. For more information, please visit our main project page.

Some key points:

  • The Main classes for the Dataless Annotators are:
    • ESADatalessAnnotator for the ESA-based Dataless Annotator
    • W2VDatalessAnnotator for the Word2Vec-based Dataless Annotator
  • Dataless Annotators add the DATALESS_ESA and DATALESS_W2V views to the input TextAnnotation respectively, and it requires the presence of a TOKENS view with the end-user's desired Tokenization.
  • Since Labels/Topics are inferred at the Document-Level, all topic annotations span the entire document.
  • Sample invocation has been provided in the main functions of each annotator.
  • Both annotators load up embeddings in memory, and thus can easily consume upto 10GB RAM.

Label Hierarchy

Dataless Classification requires the end-user to specifcy a Label hierarchy (with label descriptions), which it classifies into. The Label hierarchy needs to be provided using a very specific format:

  • labelNamePath: Specify your label id to label name mapping here in the labelID \t labelName format (label id can be any ID specific to your system, however we use the label name itself as ID in our sample hierachy for readibility)
  • labelHierarchyPath: The first line of this file should contain tab-separated list of Top-Level nodes in the hierarchy (i.e. the ones directly connected to the root). Then, every following line should specify the connections in the hierachy in the parentLabelID \t childLabelID1 \t childLabelID2 \t ... format.
  • labelDescPath: Dataless' performance hinges on good label descriptions, which you specify in this file in the labelID \t labelDescription format.

We provide a sample 20newsgroups hierarchy with label descriptions inside data/hierarchy/20newsgroups, where:

  • idToLabelNameMap.txt should be used as labelNamePath
  • parentChildIdMap.txt should be used as labelHierarchyPath
  • labelDesc_Kws_simple.txt should be used as labelDescPath

We also provide improved 20newsgroups label descriptions in labelDesc_Kws_embellished.txt which corresponds to the label descriptions used in [2], whereas the labelDesc_Kws_simple.txt corresponds to the label descriptions used in [1].

Embeddings

ESA and Word2Vec Embeddings are fetched from the DataStore on demand.

Config

A sample config file with the default values has been provided in the config folder .. config/project.properties

To check whether you are properly set to use the project or not, run:

  • mvn -Dtest=ESADatalessTest#testPredictions test to test the ESADatalessAnnotator.
  • mvn -Dtest=W2VDatalessTest#testPredictions test to test the W2VDatalessAnnotator.

If you use this software for research, please cite the following papers:

[1] Chang, Ming-Wei, et al. "Importance of Semantic Representation: Dataless Classification." AAAI. Vol. 2. 2008.

[2] Song, Yangqiu, and Dan Roth. "On Dataless Hierarchical Text Classification." AAAI. Vol. 7. 2014.