alexyangli / absa_keras Goto Github PK

Keras implementation (tensorflow backend) of aspect based sentiment analysis

• Cabasc, WWW 2018
  Liu et al. "Content Attention Model for Aspect Based Sentiment Analysis"

• RAM, EMNLP 2017
  Chen et al. "Recurrent Attention Network on Memory for Aspect Sentiment Analysis"

• IAN, IJCAI 2017
  Ma ei al. "Interactive Attention Networks for Aspect-Level Sentiment Classification"

• MemNet, EMNLP 2016
  Tang et al. "Aspect Level Sentiment Classification with Deep Memory Network"

• ATAE-LSTM(AE-LSTM, AT-LSTM), EMNLP 2016
  Wang et al. "Attention-based LSTM for Aspect-level Sentiment Classification"

• TD-LSTM(TC-LSTM), COLING 2016
  Tang et al. "Effective LSTMs for Target-Dependent Sentiment Classification"

1. download glove embedding glove.42B.300d.zip, unzip and put it in raw_data

wget https://nlp.stanford.edu/data/wordvecs/glove.42B.300d.zip
unzip glove.42B.300d.zip
mv glove.42B.300d.txt ./raw_data

2. pre-processing

python3 process_raw.py
python3 preprocess.py

python3 train.py

• python==3.6.4
• keras==2.2.4
• nltk==3.2.5
• tensorflow=1.6.0

see data_analysis.md

Note: results in the () are the performances of models with word embeddings fixed but aspect embeddings fine tuned; results in the [] are the performances of models with both word embeddings and aspect embeddings fixed

• Accuracy

• Macro-F1

• Combining with ELMo Embedding

There are 2 ways to use ELMo in this project. Before training model, change the configuration as belows:

1. In config.py, set self.use_elmo to True and self.elmo_hub_url to None or https://tfhub.dev/google/elmo/2. By running train.py the tensorflow hub will download the elmo module and cached in the local temporay directory. But the module will be deleted after a machine reboot.
2. So, I recommend you load the elmo module locally. First, download the elmo using url: https://tfhub.dev/google/elmo/2?tf-hub-format=compressed. Then change its name to tfhub_elmo_2 (whatever you want), untar the file, and put it in the raw_data directory. It will be like:
   
   And don't forget to set self.use_elmo in config.py to True and self.elmo_hub_url to raw_data/tfhub_elmo_2 (the name you just renamed).

Here's the results of combining ELMo Embedding:

• Personal conclusion

1. Surprisingly, I failed to achieved similar performance as stated in the parper of Memnet. Or maybe there are bugs in the code?
2. TD-LSTM performs unexpectedly well.
3. Models with fixed embeddings are generally better than those with fine-tuned embeddings, which is consistent with the paper of RAM.
4. L2 regularization didn't help.
5. I tried elmo embedding, but it didn't bring a performance gain.