compuframes

Required libraries/external files in lib folder

• commons-cli-1.2.jar
• gson-2.2.4.jar
• segan.jar
• commons-compress-1.4.1.jar
• mulan.jar
• stopwords.txt
• commons-math3-3.1-SNAPSHOT.jar
• opennlp-maxent-3.0.1-incubating.jar
• weka-3.7.6.jar
• en-sent.bin
• opennlp-tools-1.5.1-incubating.jar
• en-token.bin
• opennlp-uima-1.5.1-incubating.jar

Compile

• To compile: ant compile
• To build jar file: ant jar
• To make a clean build: ant clean-build

Process data

To perform preprocessing of the text data on the whole dataset

java -Xmx10000M -Xms10000M -cp 'dist/compuframes.jar:lib/*' experiment.MultilabelExperiment --json-file <json-file> -v -d --data-folder <processed-data-folder> --format-folder <format-folder> --run-mode preprocess -s -l --u 5 --b 10 --bs 5 --V 10000 --doc-type <document-type>

• <json-file>: path to the JSON file
• <processed-data-folder>: path to the folder containing processed data
• <format-folder>: subfolder of <processed-data-folder> to store one specific instance of processed data (e.g., with a specific set of preprocessing parameter of the same dataset)
• <document-type>: the type of documents to consider. This can be either "tobacco" or "immigration". If this is not set, all documents are considered.
• -s: whether stopwords are removed (stopwords are stored in lib/stopwords.txt)
• -l: whether stemming is performed
• --u: the minimum raw count of unigrams
• --b: the minimum raw count of bigrams
• --bs: the minimum chi-square score for a bigram to be considered
• --V: the maximum vocab size For other arguments, use -help.

Working example:

java -Xmx10000M -Xms10000M -cp 'dist/compuframes.jar:lib/*' experiment.MultilabelExperiment --json-file data/rounds_4-9_singly_coded.json -v -d --data-folder data --format-folder format --run-mode preprocess -s -l --u 5 --b 10 --bs 5 -s -l --V 10000

To create cross-validated data

java -Xmx10000M -Xms10000M -cp 'dist/compuframes.jar:lib/*' experiment.MultilabelExperiment --json-file <json-file> -v -d --run-mode create-cv -s -l --u 5 --b 10 --bs 5 --V 10000 --num-folds <number-of-folds> --tr2dev-ratio <train-to-dev-ratio> --cv-folder <cross-validation-folder>

• <number-of-folds>: number of cross validation folds
• <train-to-dev-ratio>: the ratio between the number of training instances and the number of development instances
• <cross-validation-folder>: folder to store the cross validated data

This will creates <number-of-folds> sets of train/dev/test. It will perform preprocessing (like in section 1 above) on the training set and use the vocabulary extracted from the training data to perform preprocessing on the development and test sets.

Working example:

java -Xmx10000M -Xms10000M -cp 'dist/compuframes.jar:lib/*' experiment.MultilabelExperiment --json-file data/rounds_4-9_singly_coded.json -v -d --run-mode create-cv -s -l --u 5 --b 10 --bs 5 -s -l --V 10000 --num-folds 5 --tr2dev-ratio 0.8 --cv-folder data/cv

Run document-level multi-label model

java -Xmx10000M -Xms10000M -cp 'dist/compuframes.jar:lib/*:../segan/dist/lib/*:../segan/dist/*' experiment.MultilabelExperiment -v -d --dataset compuframes --run-mode run --num-folds <number-of-folds> --cv-folder <cross-validation-folder> --model <model-name> --expt-folder <experiment-folder> --fold <fold-number>

• <model-name>: the name of a supported model
• <experiment-folder>: folder to store the learned model and prediction results
• <fold-number>: the fold to be run. If this is not specified, all folds in the cross-validation folder will be run.

Working example: running random baseline

java -Xmx10000M -Xms10000M -cp 'dist/compuframes.jar:lib/*:../segan/dist/lib/*:../segan/dist/*' experiment.MultilabelExperiment -v -d --dataset compuframes --run-mode run --num-folds 5 --cv-folder data/cv --model random --expt-folder experiments --fold 0