The sentiment analysis software is written in Python, and supports Python 3.7 and above. In addition, folders data and models must contain some files. Check README files in respective folders for further information.

To install requried dependencies, you can run pip install -r requirements.txt.

This project is also available as a docker container. It is recommended to use Docker and docker-compose for running this project. Commands to run phases within Docker containers are given in chapter Usage along with the regular commands for running the project locally. We use the same docker-compose volumes for all steps. Therefore, as long as you do not delete these volumes (by e.g. runnnig docker-compose down), the intermediate data should be persistent.

The docker container contains all data, models and code required for running this project successfully. Its size is around 3.8GB.

Before any sentiment prediction can be done on the SentiCoref corpus we have to preprocess the data. This can be done by running

python -m nlp_code.preprocessing

# Docker
docker-compose run preprocessing

from the root folder of the repository. This will perform lemmatisation and POS tagging on the corpus using the Stanza library. Results at this step will be saved to data/cache.

Next, feature extraction is performed, using FeaturePipeline with defined feature extracotrs. The latter are defined in module nlp_code.features. Features extracted from preprocessed text are saved in data/features. These TSV files contain one row of extracted features for each word in a coreference chain.

It is recommended that feature extractors are only added to the FeaturePipeline, since removing them might break previously implemented models that depend on some removed features.

After preprocessing is completed, model evaluation can be started by running

python -m nlp_code.models

# Docker
docker-compose run models

from the root folder of the repository. Models can use any of the features stored in data/features.

To run the CustomSentiCorefModel, you can run

python -m nlp_code.models_neural

# Docker
docker-compose run models_neural

from the root folder of the repository. This will perform the required preprocessing, training and evaluation of this model. Make sure you downloaded the required BERT model, described here.

To run the BertEmbeddingsSentiCoref (takes quite some time), you can run

python -m nlp_code.bert_embeddings

# Docker
docker-compose run bert_embeddings

For running the pretrained model with balanced training set run the Docker command (see instructions from CustomSentiCorefModel) or download the trained model from here, unzip it into models/ and run

python -m nlp_code.pretrained_bert_embeddings_balanced

# Docker
docker-compose run pretrained_bert_embeddings_balanced