Opinatt: Mining fine-grained opinions on closed captions of YouTube videos with an attention-RNN

Code and dataset for the paper "Mining fine-grained opinions on closed captions of YouTube videos with an attention-RNN". If you use this code or dataset, please cite the paper.

1. Clone this repository cd ~; git clone https://github.com/epochx/opinatt

2. Install required libraries

   • Install Theano and Tensorflow: pip install Theano tensorflow-gpu
   • Download and install Senna, http://ronan.collobert.com/senna/
   • Download and install CoreNLP 3.6, http://stanfordnlp.github.io/CoreNLP/history.html

3. Go to the opinatt diretory cd ~/opinatt and download the required data there

   • Download GoogleNews embeddings from https://code.google.com/archive/p/word2vec/
   • Download WikiDeps embeddings from https://levyomer.wordpress.com/2014/04/25/dependency-based-word-embeddings/
   • Download and unzip the SemEval2014 V2 Train data from http://alt.qcri.org/semeval2014/task4/index.php?id=data-and-tools

4. Create the working environment

   • Create data folder structure, prepare_data.sh path_where_to_create_data_folder
   • Modify opinatt/enlp/settings.py accordingly

5. Preprocess corpora by running python process_corpus.py

6. Generate training data in JSON format:

   • For aspect extraction run python generate_json.py --json_path path/to/json/file --train LaptopsTrain --test LaptopsTest --embeddings SennaEmbeddings.
   • For aspect extraction and sentiment classification using collapsed tags, run python generate_json.py --json_path path/to/json/file --train LaptopsTrain --embeddings SennaEmbeddings --sentiment.
   • For joint aspect extraction and sentiment classification run python generate_json.py --json_path path/to/json/file --train LaptopsTrain --embeddings SennaEmbeddings --joint
   • When generating the training data for the Youtubean dataset, or when doing so including sentiment labels for the SemEval corpora, make sure not to pass the --test flag, as there are no test sets for those settings. The script will generate the corresponding splits based on the training portion for each case. Add the --strict label to only use sentences with a single sentiment.

7. Train the models:

   • For the baseline: python run_baseline.py --json_path path/to/json/file --results_path path/to/results/folder
   • For the attention-RNN: python run.py --json_path path/to/json/file --results_path path/to/results/folder

For more details on how to use the provided scripts use the --help option.

Contact

Feel free to email emarrese@weblab.t.u-tokyo.ac.jp for any pertinent questions/bugs regarding the code.