Sentiment_Analysis

Chinese financial short text sentiment analysis

Requirements

Unix/Linux operating System
python2.7
python package: keras, gensim, sklearn, jieba, pypinyin, h5py, numpy, pandas, Theano. Version details see requirements text.

Data preprocessing

lexical analysis

First, run:

coreNLP_lexical_analysis.py(Need to download stanford-corenlp-full-2018-01-3 file in advance) and jieba_lexical_analysis.py

Second, combined result of coreNLP lexical analysis and jieba lexical analysis to binarization through binarization.py.

phrase structure analysis

run phrase_structure_preprocess.py

Train and Evaluation

run ./main_code/PYGS_CNN-BiLSTM.py

80% data use to be training and 20% data use to be test. (Corpus will be available after paper published.)

Example output

Epoch 1/30 14525/14525 [==============================] - 58s 4ms/step - loss: 0.9198 - mean_absolute_error: 0.3688 - acc: 0.5497 - val_loss: 0.7305 - val_mean_absolute_error: 0.2787 - val_acc: 0.6820

Epoch 2/30 14525/14525 [==============================] - 58s 4ms/step - loss: 0.7009 - mean_absolute_error: 0.2770 - acc: 0.6956 - val_loss: 0.6097 - val_mean_absolute_error: 0.2319 - val_acc: 0.7525

Epoch 3/30 14525/14525 [==============================] - 58s 4ms/step - loss: 0.5910 - mean_absolute_error: 0.2305 - acc: 0.7541 - val_loss: 0.5633 - val_mean_absolute_error: 0.2028 - val_acc: 0.7781

Model file

lstm.h5 is network training weight
lstm.yml is training network structure
Word2vec_model.pkl is Word2vec model
events.out.tfevents.1532231423.dl-B85M-D3H is tensorboard file

Cite our paper

Rao, D., Huang, S., Jiang, Z. et al. A dual deep neural network with phrase structure and attention mechanism for sentiment analysis. Neural Comput & Applic (2021). https://doi.org/10.1007/s00521-020-05652-6

fip-lab/Sentiment_Analysis