Semi-supervised_learning_for_part_of_speech_tagging

NLP Capstone Project with Columbia University and J.P. Morgan

Please check dev branch for the latest version.

Preparations

Dependencies Preparations

0. (Ignore this step) Check the cuda version on GCP as

sudo nvidia-smi

1. No need to configure virtual environment for this machine. Just use the base environment and install the dependencies as

pip install -r requirements.txt

Model/Data Preparations

1. Download model/base_model.pt from Google Drive into your model/ directory.
2. Download data/gweb_sancl/ from Google Drive into your data/ directory.

Directory Preparation

Make sure you have the following directories in the root before running the script

.
├── Analysis_int_res.ipynb
├── Analysis_output_Online_fixed_self_learning.ipynb
├── Analysis_output_Online_nonfixed_self_learning.ipynb
├── LICENSE
├── Online_fixed_self_learning_v5.ipynb
├── Online_nonfixed_self_learning_v5.ipynb
├── Online_token_self_learning_v5.ipynb
├── README.md
├── Scratch_fixed_self_learning_v5.ipynb
├── Scratch_nonfixed_self_learning_v5.ipynb
├── Scratch_token_self_learning_v5.ipynb
├── analysis.py
├── build_model.py
├── create_pseudo_data.py
├── create_pseudo_data_by_tokens.py
├── data
│   └── gweb_sancl
│       ├── pos_fine
│       │   ├── answers
│       │   ├── emails
│       │   ├── newsgroups
│       │   ├── reviews
│       │   ├── weblogs
│       │   └── wsj
│       └── unlabeled
│           └── gweb-answers.unlabeled.txt
├── docs
├── intermediate_result
├── metrics
├── model
├── plots_tags
├── requirements.txt
├── result
├── scripts
├── setup.sh
└── utils.py

Directory Structure

Here follows the brief introduction about the specific details for each directory:

1. metrics: store the metrics at each loop after self training including precision, f1 and recall
2. plots: store the plots for metrics at different parameter settings
3. model: store the model settings to save time
4. data: store the data we are gonna use
5. docs: store the meeting records for the project
6. pickles: store the serialized python object after self-training for future usages.

Results