Sentiment Analysis with Hierarchical Neural Networks - Pytorch flavor

This repository holds Pytorch (near)-implementations of the following papers:

• Hierarchical Attention Networks for Document Classification paper
• Neural Sentiment Classification with User & Product Attention paper

Implementations are single-gpu batched.

Requirements:

• Pytorch 0.4

• TQDM

• Gensim (Only for BuildW2VEmb.py to build word embeddings with word2vec)

• Spacy 1.9 (Only for prepare_data.py to tokenize reviews and split in sentences)

Quickstart:

1. chmod +x minimal_ex_cuda.sh
2. ./minimal_ex_cuda.sh

Main Scripts:

• han.py -> Hierarchical Attention Networks for Document Classification model training script.
• nscupa.py -> Neural Sentiment Classification with User & Product Attention model training script.

Input formats:

Data input (filename argument)

han.py and nscupa.py expect a pickled dictionnary with the following keys as input:

input = {"data":data,"splits":splits,"rows":("user_id","item_id","review","rating")}

Where:

• The "data" field holds is a list of ("user_id","item_id","review","rating") tuples where id's and ratings are str and review is tokenized and sentence-splitted text list(list(str)).
• The "splits" key is a list integers list(int). It's each data point split.
• The "rows" is used as helper for tuple indexing in the former scripts.

=> A helper script prepare_data.py is provided to create such input pickle file.

Pre-trained embeddings (emb optional argument)

han.py and nscupa.py can use pre-trained embedding. It expects a .txt where each line is a word followed by its vector. The first line of this file provides the number of words and the size of each vectors. BuildW2VEmb.py is provided to build word embeddings from data using the word2vec-skipgram algorithm.

If no pre-trained embeddings are provided han.py and nscupa.py build embedding dictionnaries and vectors on the fly (---max-feat arg).

Helper scripts

• prepare_data.py transforms gzip files as found on Julian McAuley Amazon product data page to a list of (user,item,review,rating) tuples.
• minimal_ex(_cuda).sh Does everything and start learning (just chmod +x them).
• fmtl.py holds data managing objects.
• Nets.py holds neural network models.
• beer2json.py is an helper script to convert ratebeer/beeradvocate datasets.
• BuildW2VEmb.py can help you build word embeddings from data.

Note

This code has been written for python-3.6. If you are only using python-2.7, switch.