Implementation in Tensorflow of A Structured Self-attentive Sentence Embedding with the sentiment analysis task.

The code organization is similar to NeuroNER.

SelfSent relies on Python 3.5 and TensorFlow 1.0+. For the package stanford_corenlp_pywrapper, just install it from https://github.com/mpagli/stanford_corenlp_pywrapper.

You need to create a data folder next to the src folder. Then for each dataset, you have to create a separate folder. In this folder, you just need to put a file all.json where each line correspond to a json sample with its attributes.

Here a sample of Yelp dataset:

{"review_id":"IYE_M_cRsk-AhVYeYvnADg","user_id":"r-zUIQPaHzvIyL93wQaoiQ","business_id":"HE23DlZWAO_JF1VIHA60TQ",**"stars":3**,"date":"2012-10-09",**"text":"This is the Capitol Square branch."**,"useful":0,"funny":0,"cool":0,"type":"review"}

In the case of review star prediction, the needed attributes are text and stars.

{"review_id":"IYE_M_cRsk-AhVYeYvnADg","stars":3,"date":"2012-10-09","text":"This is the Capitol Square branch.","user_id":"r-zUIQPaHzvIyL93wQaoiQ","business_id":"HE23DlZWAO_JF1VIHA60TQ""useful":0,"funny":0,"cool":0,"type":"review"}

The parameters do_split (force to split even if the pickle files exist), training, valid, test in parameters.ini will split the dataset accordingly.

It also needs some word embeddings, which should be downloaded from http://neuroner.com/data/word_vectors/glove.6B.100d.zip, unzipped and placed in /data/word_vectors. This can be done on Ubuntu and Mac OS X with:

# Download some word embeddings
mkdir -p SelfSent-master/data/word_vectors
cd SelfSent-master/data/word_vectors
wget http://neuroner.com/data/word_vectors/glove.6B.100d.zip
unzip glove.6B.100d.zip

Be sure that use_pretrained_model = false and have at least all.json in the data folder.

You need to have a pretrained model, which is composed of:

• dataset.pickle
• model.ckpt.data-00000-of-00001
• model.ckpt.index
• model.ckpt.meta
• parameters.ini

Don't forget to put use_pretrained_model = true and the path to the pretrained model folder.

Don't hesitate to contact for any feedback or create issues/pull requests.