Word and Constituent Boundaries in Hierarchical Multiscale Recurrent Neural Networks

Branches under development

• Parser benchmark on Penn Treebank: ptb
• Question Answering: train_qa

References

• hierarchical-rnn package
• Hierarchical Multiscale Recurrent Neural Networks

Corpus

• Text8
• Penn Treebank partially available from NLTK

>>> import nltk
>>> nltk.download()
...
Identifier> treebank

• Generate groundtruth boundary labels from Penn Treebank under treebank/: python convert_boundary.py --path TARGET_PATH --threshold MIN_TOKENS

Usages

Parser Benchmark

• End-to-end training, testing, and evaluation on NYU HPC clusters:

sbatch ptb_pipe.sbt

• Tuning configurations: modify hierarchical-rnn/config.yml
• Relax, wait, and collect pickled output(s)

Updated Progress

1. F1 score of HM-RNN boundary detection:

   1. (finished) Convert parsing in PTB to 1s/0s boundary indicators, and use that as ground truth boundaries
   2. (finished) Test trained HM-LSTM models on PTB, and store layer-wise indicators
   3. (finished) calculate F1 scores of HM-LSTM for some layer’s boundary indicators (TODO: plot fancy figures)
   4. (finished) Calculate BPC (LM evaluation metric) by these HM-LSTM on PTB
   5. Train more models; compare the correlation/trending of F1 and BPC

2. Statistically analyze with PCFG from PTB:

   1. (finished) Compute PCFGs from PTB
   2. Pick the model with best syntactic meanings of HM-LSTM boundary indicators / highest F1 score
   3. Find out if/what constituencies detected by HM-LSTM boundary coincide with PCFGs

3. QA on children book dataset

   1. (finished) Setup data preprocessing, pipeline to hm-lstm model
   2. (finished) Tune to improve test precision
   3. Replace self embedding nets with GloVe pre-trained word embeddings
   4. Beat the baseline performance of vanilla LSTM