This repository contains the source code and the dataset for vaccine attitude detection.

The annotations are given in the form of ID,stance,aspect_span_start:aspect_span_end,opinion_span_start:opinion_span_end,aspect_catetegory
in the Datasets_Raw folder.

To obtain tweet text,

1. cd twitter_get_text_by_id_twitter4j
2. Open ./settings/crawler.properties and setup your consumerKey, consumerSecret, access token and access token secret.
   1. For the acquisition of consumerKey, consumerSecret, access token and access token secret, please refer to https://developer.twitter.com/en/docs/developer-portal/overview. The Standard v1.1 is sufficient.
3. run twitter_get_text_by_id_twitter4j by either java -jar twitter_vac_opi_cwl_by_id.jar ./settings/crawler.properties or javac -cp "./*" ./src/main/org/backingdata/twitter/crawler/rest/TwitterRESTTweetIDlistCrawler.java The tweets are stored in ./saves in json format.

cd VADMlmFineTuning
VADtransformer is firstly trained unsupervised. The model will be saved to ../datasets/mlm-vad.

To perform unsupervised training,

1. Replace tweetIDs in UnannotatedTwitterID_training.csv and UnannotatedTwitterID_testing.csv with obtained tweet text.
2. Put the tweet text file in ../datasets. The format is the same as vad_train_finetune.txt.
3. cd src and run train_vad_albert_vae.py

cd VADStanceAndTextspanPrediction

In the previous step we obtain the unsupervised pre-trained VAD, scilicet the TopicDrivenMaskedLM. At this stage we wrap the model with classifiers and constrains, and train the model.

To perform supervised training,

1. Move the saved model (i.e., the pytorch_model.bin file) from the ../datasets/mlm-vad of VAD unsupervised training to the ./datasets/albertconfigs/vadlm-albert-large-v2/vad-cache folder. For your convenience a saved TopicDrivenMaskedLM is ready-to-use in the ./datasets/albertconfigs/vadlm-albert-large-v2/vad-cache folder.
2. Move the saved config of the model (i.e., the config.json file) from the ../datasets/mlm-vad of VAD unsupervised training to the ./datasets/albertconfigs/vadlm-albert-large-v2/vadlm-albert-large-v2 folder. For your convenience a saved config.json is ready-to-use in the ./datasets/albertconfigs/vadlm-albert-large-v2/vadlm-albert-large-v2 folder.
3. cd src and run vadtrain_eval_predict.py for training and testing.
   1. Training: Uncomment line 1559-1578 of vadtrain_eval_predict.py and run the file. Checkpoints will be saved in ./datasets/vadcheckpoints/5-fold-211103/vadlm-albert-large-v2/
   2. Testing: Uncomment line 1580-1608 of vadtrain_eval_predict.py and run the file. The prediction will be output in same directory. A saved model can be downloaded via this link. You can place the save model in ./datasets/vadcheckpoints/5-fold-211103/vadlm-albert-large-v2/ for a quick start.