Introduction

• A classification of Sentiment Analysis which is implemented by pytorch.

• There are many data in data, *.txt of that are came from sent-conv-torch, *.conll.out of that are generated from our lab's parser.
  • TREC
  • SUBJ
  • MR
  • CR
  • MPQA

• My processed data by preprocessed.sh will be saved in processed_data.

• There are many models in models.
  • Pooling
  • CNN
  • Multi_Channel_CNN
  • Multi_Layer_CNN
  • CharCNN
  • GRU
  • LSTM
  • LSTM_CNN
  • TreeLSTM
  • biTreeLSTM
  • TreeLSTM_rel
  • biTreeLSTM_rel
  • CNN_TreeLSTM(ready to refresh)
  • LSTM_TreeLSTM(ready to refresh)
  • Transformer(TODO)

• My log will be saved in log.

• There are many scripts in utils.
  • Alphabet.py which is used to build dictionary.
  • Common.py which is saved unk-key and pad-key.
  • Embedding.py which is used to load pre_train embedding by Yang Song.
  • Evaluate.py which is used to calculate the F1.
  • Feature.py which is implemented a sentence's features, including word, word_id, label, root and so on.
  • build-batch.py which is used to build the data's mini batch.
  • log.py which is used to save the log.
  • opts.py which is implemented the argparses.
  • trainer.py which is used to train the data.
  • tree.py which is implemented the tree's methods.

Requirement

    python : 3.5+
    pytorch : 0.4.1
    cuda : 8.0 (support GPU, you can choose)

Usage

• first step

    sh preprocess.sh
  

• second step

    sh run.sh
  

• third step (also called decoder step which will output a file whose predictions were wrong. If necessary)

    sh decoder.sh
  

Result

Data/Model(acc)	TREC	SUBJ	MR	CR	MPQA
Pooling	76.12	92.10	75.92	79.03	85.97
CNN	91.40	93.20	77.05	83.60	88.34
Char_CNN	92.20	93.30	78.66	83.60	88.25
Multi_Channel_CNN	89.20	93.40	78.56	81.45	88.06
Multi_Layer_CNN	91.00	93.70	78.28	83.06	88.44
LSTM	89.20	92.50	78.94	81.99	89.57
LSTM_CNN	90.08	93.40	79.51	82.80	88.82
GRU	89.40	92.80	78.28	82.26	89.48
TreeLSTM	89.60	92.60	79.79	84.41	88.91
biTreeLSTM	90.40	92.70	79.13	83.87	88.91
TreeLSTM_rel	91.29	92.20	80.36	82.26	89.06
biTreeLSTM_rel	91.20	92.80	80.26	83.60	89.10
CNN_TreeLSTM	-	-	-	-	-
LSTM_TreeLSTM	-	-	-	-	-

In addition:

Data

• TREC: TREC question dataset - task involves classifying a question into 6 question types (whether the question is about abbreviation, entity, description, human, location and numeric value) (Li and Roth, 2002).
• SUBJ: Subjectivity dataset where the task is to classify a sentence as being subjective or objective (Pang and Lee, 2004).
• MR: Movie reviews with one sentence per review. Classification involves detecting positive/negative reviews (Pang and Lee, 2005).
• CR: Customer reviews of various products (cameras, MP3s etc.). Task is to predict positive/negative reviews (Hu and Liu, 2004).
• MPQA: The MPQA Opinion Corpus contains news articles from a wide variety of news sources manually annotated for opinions and other private states (i.e., beliefs, emotions, sentiments, speculations, etc.).

Emphasize

• pre_trained_embed which is using glove.6B.100d.txt.
• TreeLSTM which is using ChildSum method.

Future Work

Other methods about TreeLSTM will be updated in the near future.

Question

Glad to receive your report by [email protected], If you have any questions about this code !