This is the repository for our ACL 2019 paper Unsupervised PCFG Induction with Normalizing Flow.

Dependencies:

Required packages:

• Python 3.6+
• pytorch 1.1+
• gensim
• nltk
• numpy
• scipy
• bidict

Newest versions of the packages should also work.

---

How to use:

1. You will need a linetrees file, which is a one-sentence-per-line file with bracketed trees without the ROOT node. The make_linetoks.py and make_ints_file.py in utils folder will use this file to first generate a linetoks file, one-sentence-per-line with just the terminals, and dict and ints file where the terminals are replaced with indices.

2. embed_with_multilingual_elmo.py in utils requires ElmoForManyLang. You can also get pretrained Elmo models from there. Use the script to generate the Elmo embeddings for the dataset.

3. A config file is needed. One sample config file is provided in the config folder along with the necessary text files in uyghur_data. The options are explained in the config file.

4. The running command is python dimi-trainer.py config/yourconfig.ini. If the Elmo embeddings are generated for the provided Uyghur file, issuing python dimi-trainer.py config/uyghur.ini should start the model immediately. A GPU is required for running the system.

5. Results will be dumped out into the provided output folder. The main diagnostic file is running_status.txt, which includes a whole array of different measurements for grammar quality. The *.vittrees.gz files are gzipped files of Viterbi trees of the dataset.