README.md

Argument classification and graph creation

Run argument classifier to label as claim, premise, or non-argumentative unit

The classifier was trained on data found in AMPERSAND: Argument Mining for PERSuAsive oNline Discussions and Parsing Argumentation Structures in Persuasive Essays, which can be found on S3: s3://convosumm/data/claim-premise-data/.

Download the model and run on a dummy sentence:

aws s3 cp --recursive s3://convosumm/checkpoints/CLPR/ $CLPR_PATH
python scripts/arg_classifier_test.py $CLPR_PATH "This is a test sentence."

Run claim/premise prediction for filtering without graph creation

python scripts/arg_classifier.py
python scripts/process_arg_classifier_results.py

Run claim/premise prediction for each comment and then join claims into a graph, potentially across comments

See the _process() function for details about the expected input format.

python Argument-Graph-Mining-code/recap_am/app.join_separate_graphs.py INPUT_FILENAME OUTPUT_DIR 

Run claim/premise prediction for each commment, create subject node for each comment and connect all subject nodes to conversation node (no connections between comments)

See the _process() function for details about the expected input format.

python Argument-Graph-Mining-code/recap_am/app.separate_graphs.py INPUT_FILENAME OUTPUT_DIR

Load the graph from .json format and save source as {train,val,test}.graph

python scripts/load_graph2txt.py $OUTPUT_DIR

BART

Run fairseq preprocessing (for vanilla BART using 2048 tokens)

./scripts/prep.sh

Run BART training in fairseq (num of visible devices * UPDATE_FREQ = 32)

./scripts/finetune.sh $BART_PATH $DATA_DIR $CHECKPOINT_DIR $TENSORBOARD_DIR $CUDA_VISIBLE_DEVICES $UPDATE_FREQ  3e-5 20 200 -1

Run BART inference

python scripts/inference.py $MODEL_DIR checkpoint_best.pt $DATA-BIN $TEST_SOURCE_FILE $OUTPUT_FILE 4 1 80 120 $BATCH_SIZE 2048 ./misc/encoder.json ./misc/vocab.bpe 

Longformer

See ./longformer-code/run.sh for an example of running the Longformer.