Sentence Embedding Matrix with Self-Attention

This repository contains a implementation for the paper A Structured Self-Attentive Sentence Embedding, which was published in ICLR 2071.

Requirements

Here is the environment we need:

• Python(3.8)
• Pytorch(1.12.1+cuda11.3)
• torchtext(0.6.0) (very important)
• spacy(3.4.1)
• en-core-web-sm(3.4.0)

Dataset

We train the model on two open datasets: PAN15 Author Profiling for the task of Author Profiling and Yelp for the task of Sentiment Analysis, but only achieved close to the performance as the paper said on the Yelp dataset.

Models	Yelp	Age
BiLSTM + Max Pooling + MLP	61.75%	60.35%
CNN + Max Pooling + MLP	58.13%	60.69%
Our Model	63.00%	61.90%

Training

For author profiling task:

python train.py --model_type SelfAttention --dataset Age --Age_train_path 'your trainset path' --Age_test_path 'your testset path'

And for sentiment analysis task:

python train.py --model_type SelfAttention --dataset Yelp --Yelp_train_path 'your trainset path' --Yelp_test_path 'your testset path'

Here are some useful options for the train script.

Options	Optional values	Meanings
--model_type	SelfAttention | BiLSTM | CNN	Which model we use for training
--dataset	Age | Yelp	Which task(dataset) we use for training
--seed		The random seed.
--LSTM_hidden		The out dimension of LSTM for model SelfAttention or BiLSTM
--MLP_hidden		The hidden units of the classifier for the downstream application.
--aspects		The hyperparameter $r$ in the paper
...	...	...

Pretrained Weigths

Here is a pretrained weight(code: 8smh) for the Yelp dataset.

Visualization:

For one single sentence's visualization, execute the following command:

python visualization.py --sentence 'your sentence' --muti_sentence False

And for a text file including multiple sentences as input, execute:

python visualization.py --muti_sentence True --path 'your text file path'

Some of the visualization results are as follows. The reason why the visualization is not as significant as in the paper might be that the algorithms are different. At least the visualization algorithm used in this repository is not likely to have so many dark red areas. (Do not take this seriously:-D