Norm Violation Detection

This is the repo for norm violation detection on Reddit.

Data (Skip this if you're just doing inference)

Decompress the data file

tar -xzf data.tar.gz

Checkpoint (Skip this if you're training the model from scratch)

The BERT-LSTM model checkpoint is accessible from this Google Drive link. Download the model to ckps/clf/bert-base-uncased/1/seed=2022.

The T5 model checkpoint is accessible from this Google Drive link. Download the model to ckps/prompt/t5-base/1/seed=2022.

Using the API

Instantiate the API:

Using BERT

python api-inference.py --task=clf --model_name=bert-base-uncased

Using T5

python api-inference.py --task=prompt --mdoel_name=t5-base

If GPUs are available, specify the GPU(s) to use:

python api-inference.py --task=prompt --model_name=t5-base --gpu=0

Then prepare the query data as in api-test-data.json.

Calling the API:

curl -X POST -H 'Content-Type: application/json' -d '@api-test-data.json' http://localhost:5000/api

The following is an example of the API. In the input file api-test-data.json there are three conversations, along with there subreddits and specified rules. The model outputs the confidence scores that the last comments from the conversations violate their corresponding rules.