Tensor Recursive Model

We provide an extension neural and probabilistic models of recursive models for tree-structure data based on tensor theory.

List of publication based on this repository:

• "Generalising Recursive Neural Models by Tensor Decomposition",
  Daniele Castellana, Davide Bacciu,
  International Joint Conference on Neural Networks (IJCNN), Glasgow, UK, 2020

• "Tensor Decompositions in Recursive Neural Networks for Tree-Structured Data",
  Daniele Castellana, Davide Bacciu,
  European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning, Bruges, Belgium, 2020

• "Learning from Non-Binary Constituency Trees via Tensor Decomposition",
  Daniele Castellana, Davide Bacciu,
  Proceedings of the 28th International Conference on Computational Linguistics, Barcelona, Spain, 2020

Code Structure

The code is structured as follows:

• data contains the BoolSent dataset
• models contains the implementation of the recursive models (both probabilistic and neural);
• preprocessing contains the code to preprocess the datasets;
• tasks contains all the files to execute the experiments; there is a readme files for each task;
• tree_utils contains utils code.

Also, two other repositories are used:

• exputils provides the utils to run an experiment, train a model, parse configuration, etc..
• thlogprob provides a minimal library to handle distribution using pytorch.

See the readme for more information.

How to run an experiment

1. download the raw data (see next section);

2. in NLP tasks, sentences should be parsed running the command:

   python tasks/task_name/parse_raw_data raw_data_folder output_folder

3. Run the preprocessor using the command:

   python preprocess.py --config-file preproc_config_file,
   where preproc_config_file can be found in the folder tasks/task_name/config_files

4. Run the experiment using the command:

   python run.py --config-file run_config_file,
   where run_config_file can be found in the folder tasks/task_name/config_files.

Where to download the dataset

• ListOps
• LRT
• SICK
• SST
• TREC