/ISCR-DRL

Interactive Spoken Content Retrieval by Deep Reinforcement Learning

Primary LanguageLex

Interactive Spoken Content Retrieval

Installation

  1. Lasagne, Theano
  2. Progressbar, tqdm
  3. tsne, cprint

Work flow

  1. Built language models from PTV transcripts * Transcript directory should be a directory with T0001,T0002,...T5047 transcription files * Specify transcript directory in src/transcript2docmodel.py * Takes approximately 6 hours mainly due to 100k (keyterm) * cmd: python src/transcript2docmodel.py
  2. Train agent * run.py: Specify data, fold, feature, experiment_prefix(directory to save results), result_directory with argparser * Other argument can be adjusted/added/altered, see for yourself * cmd: python src/run.py
  3. View Results * Use merge_csvs.py to merge result/*.log * cmd: python result/parse_log_to_csv.py $dir

Change feature

  1. Change feature type: src/IR/statemachine.py, run_training.py - if/else condition in constructor, featureExtraction & argparser

Change cost

  1. Change cost table: src/IR/actionmanager.py, possibly add another option in run_training.py, argparse

Visualize

  1. specify network pickle ,feature file, number of features, save_path with src/run_visualize.py
  2. use jupyter notebook to open result/plot_feature_action.ipynb & previous save h5 file

Other Notes

  • Don't ask me about the code and the data storage format, it's just as it is
  • I believe there are bugs in Wen's data, naming a few
    • Some keyterms/requests do not exist, can reproduce if I can access Wen's recognition transcripts
  • Other cutting methods: snownlp