Sentiment Analysis

2020 Spring Machine Learning project haoshibo@pku.edu.cn

Introduction

An implement of sentiment analysis on IMDB dataset with bi-LSTM and word attention.

Requirements

• python 3.6
• pytorch 1.4
• torchtext 0.5
• spacy 2.2.4

Preprocess

Download aclImdb_v1.tar.gz and unzip it in root directory. Run preprocess.py to get jsonlines files of train.jsonl, test.jsonl. You should also download `glove.6B.100d' before training.

Training

Run train.py, before that you can set the variable 'modelname' to specify the version of model you are training. You can choose Three types of model:

1. two layers of Bi-LSTM with h_n as the encoding of the text.
2. two layers of Bi-LSTM with element-wise average h_t as the encoding of the text.
3. two layers of Bi-LSTM with word attention

For the first and second model, you should set in train.py

model = RNN(INPUT_DIM, EMBEDDING_DIM, HIDDEN_DIM, OUTPUT_DIM, N_LAYERS,
            BIDIRECTIONAL, DROPOUT, PAD_IDX)

and choose the comment line in class RNN in model.py to decide which one you are training.

To use word attention model, just set in train.py

model = attentionRNN(INPUT_DIM, EMBEDDING_DIM, HIDDEN_DIM, OUTPUT_DIM, ATT_DIM, N_LAYERS,
                    BIDIRECTIONAL, DROPOUT, PAD_IDX)

Load Model and Test on specific sentence

run test.py after set 'modelname' in the file

Experiment Result

Word Attention: Epoch: 01 | Epoch Time: 3m 25s Train Loss: 0.454 | Train Acc: 78.39% Val. Loss: 0.288 | Val. Acc: 88.66% Epoch: 02 | Epoch Time: 3m 33s Train Loss: 0.273 | Train Acc: 88.99% Val. Loss: 0.258 | Val. Acc: 89.99% Epoch: 03 | Epoch Time: 3m 35s Train Loss: 0.211 | Train Acc: 91.84% Val. Loss: 0.241 | Val. Acc: 90.72% Epoch: 04 | Epoch Time: 3m 30s Train Loss: 0.167 | Train Acc: 93.81% Val. Loss: 0.268 | Val. Acc: 90.46% Epoch: 05 | Epoch Time: 3m 33s Train Loss: 0.133 | Train Acc: 95.28% Val. Loss: 0.321 | Val. Acc: 90.30% Test Loss: 0.250 | Test Acc: 90.01%