Aspect Level Sentiment Classification with Deep Memory Network

TensorFlow implementation of Tang et al.'s EMNLP 2016 work.

Problem Statement

Given a sentence and an aspect occurring in the sentence, this task aims at inferring the sentiment polarity (e.g. positive, negative, neutral) of the aspect.

Example

For example, in sentence ''great food but the service was dreadful!'', the sentiment polarity of aspect ''food'' is positive while the polarity of aspect ''service'' is negative.

Quick Start

Download the 300-dimensional pre-trained word vectors from Glove and save it in the 'data' folder as 'data/glove.6B.300d.txt'.

Train a model with 7 hops on the Laptop dataset.

python main.py --show True

Note this code requires TensorFlow, Future and Progress packages to be installed. As of now, the model might not replicate the performance shown in the original paper as the authors have not yet confirmed the optimal hyper-parameters for training the memory network.

Training options

• edim: internal state dimension [300]
• lindim: linear part of the state [75]
• nhop: number of hops [7]
• batch_size: batch size to use during training [128]
• nepoch: number of epoch to use during training [100]
• init_lr: initial learning rate [0.01]
• init_hid: initial internal state value [0.1]
• init_std: weight initialization std [0.05]
• max_grad_norm: clip gradients to this norm [50]
• pretrain_file: pre-trained glove vectors file path [../data/glove.6B.300d.txt]
• train_data: train gold data set path [./data/Laptop_Train_v2.xml] or [./data/Restaurants_Train_v2.xml]
• test_data: test gold data set path [./data/Laptops_Test_Gold.xml] or [./data/Restaurants_Test_Gold.xml]
• show: print progress [False]

Performance - Laptop Dataset (todo)

Performance - Restaurant Dataset (todo)

Acknowledgements

• More than 80% of the code is borrowed from carpedm20.
• Using this code means you have read and accepted the copyrights set by the dataset providers.

Author

Ganesh J

Licence

MIT