Major updates to the original repo:
- Tweak the model to improve the accuracy by 8%. Current accuracy (for both datasets) is close to that of the original paper. Major changes(improvement):
- Use spacy for tokenization (very important)
- Correct position encoding
- Combine source and target word index
- Accept multi-target words as input
- Remove conflict training examples
- Set Embeddings to be not trainable (to generalize for other datasets)
- Provide training logs
- Upgrade to tensorflow 1.0 and fix some minor bugs
TensorFlow implementation of Tang et al.'s EMNLP 2016 work. (improvements on ganeshjawahar's implementation)
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.
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.
Download the 300-dimensional pre-trained word vectors from Glove and save it in the 'data' folder as 'data/glove.6B.300d.txt'.
Install the requirement by pip install -r requirements.txt
You also need to download spacy model via python -m spacy download en
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.
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]
| Model | In Paper | This Code |
|---|---|---|
| MemNet (1) | 67.66 | |
| MemNet (2) | 71.14 | 69.0 |
| MemNet (3) | 71.74 | |
| MemNet (4) | 72.21 | |
| MemNet (5) | 71.89 | |
| MemNet (6) | 72.21 | |
| MemNet (7) | 72.37 | |
| MemNet (8) | 72.05 | |
| MemNet (9) | 72.21 |
| Model | In Paper | This Code |
|---|---|---|
| MemNet (1) | 76.10 | |
| MemNet (2) | 78.61 | 78.0 |
| MemNet (3) | 79.06 | |
| MemNet (4) | 79.87 | |
| MemNet (5) | 80.14 | |
| MemNet (6) | 80.05 | |
| MemNet (7) | 80.32 | |
| MemNet (8) | 80.14 | |
| MemNet (9) | 80.95 |
- More than 90% of the code is borrowed from ganeshjawahar.
- 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.
MIT