projekt-opal/classification

Classification of DCAT themes using decision trees and TF-IDF (D3.4)

OPAL Classification

This component predicts objects of the dcat:theme predicate of dcat:Dataset subjects, based on the dc:description property. The implementation was done using WEKA https://github.com/Waikato/weka-3.8. You can run the application with the default values with mvn clean install and mvn exec:java -Dexec.mainClass="tools.Main" -Dexec.args="-c j48 -ngrams 1" -Dexec.cleanupDaemonThreads=false
The result of the evaluation of the cross-validation of the training data and the evaluation of the test data is printed to console.

The following arguments can be provided:
-c {naive, j48}, default it j48
-ngrams {1,...,n}, default is 1
-query, sparql query default is SELECT * WHERE { ?s a <http://www.w3.org/ns/dcat#Dataset> ; <http://www.w3.org/ns/dcat#theme> ?o ; <http://purl.org/dc/terms/description> ?d FILTER ( lang(?d) = "en" ) } LIMIT 300
-endpoint, sparql endpoint, default is: https://www.europeandataportal.eu/sparql

Pre-processing

The following steps were taken:

1. Removed punctuation
   
2. Converted all text to lower case
   
3. Tokenization and Lemmatization
   
4. Removed the standard english stop words if either the lemma or the original word coincides
   

Word vectorization

The standard TF-IFD word vectors were computed.

Results

In the interest of time and since the approach is slow, the classifier was trained with 160 instances. That number might be too small to be representative.
The following accuracy was obtained for the cross-validation method with 4 folds:

Classifier	1-gram	2-gram	3-gram	4-gram
J48	75,625%	59,375%	59,375%	59,375%
NaiveBayes	47,5%	31,875%	36,875%	35%

The following accuracy was obtained for the evaluation of the test data.

Classifier	1-gram	2-gram	3-gram	4-gram
J48	62,07%	50%	59,09%	55,32%
NaiveBayes	28,09%	29,35%	27,59%	28,05%

Note

This component was developed by Ana Alexandra Morim da Silva during the OPAL hackathon. The component was on of the winners of the hackathon.

Credits

Data Science Group (DICE) at Paderborn University

This work has been supported by the German Federal Ministry of Transport and Digital Infrastructure (BMVI) in the project Open Data Portal Germany (OPAL) (funding code 19F2028A).