Skip to content

Latest commit

 

History

History
95 lines (63 loc) · 3.74 KB

README.md

File metadata and controls

95 lines (63 loc) · 3.74 KB
Raw

CUED_SPEECH @ TREC2020 Podcast Summarisation Track

Overview

  • Report: CUED_SPEECH AT TREC 2020 PODCAST SUMMARISATION TRACK
  • It is a two-stage approach as the BART model can only take up to 1024 input words:
    • Stage1: Truncation or Filtering (by Hierarchical Model)
    • Stage2: Running BART
  • Scripts have absolute paths (to model/data), typically variables in CAPITAL => change them to your own paths.
  • Feel free to email [email protected] if you think this repository might be useful for your work, but it's not working.

Requirements

  • python 3.7
  • torch 1.2.0
  • transformers (HuggingFace) 2.11.0

(I expect it to work with newer versions too, but not guranteed & not tested)

Data Preparation

Spotify Podcast: Download & Pre-processing

  • Download link: https://podcastsdataset.byspotify.com/
  • data/processor.py: split the data into chunks such that each chuck contains 10k instance, e.g. id0-id9999 in podcast_set0 (we could use your own data processing pipeline!!). Note that to use our trained weights, we use BART tokenizer to process the data.
  • data/processor_testset.py: for pre-processing test data
  • data/loader.py contains functions (mostly hard coded) for data loading, batching, etc.

Train & Fine-tune BART

*Currently, all the configurations must be set in the training script!!

Standard Fine-tuning:

python train_bartvanilla.py

L_rl training:

python train_bartvanilla_rl.py

Decode (Inference) BART

*Again, all configutations must be set in the script (note that the current settings are the one used in TREC2020)

decoding:

python decode_testset.py start_id end_id
  • The test set consists of 1,027 samples. See data/processor_testset.py about how to prepare the test data. If used in a single machine, you can unable ID randomization.
  • To use sentence filtering at test time, you need to decode the test data using the hierarchical model first as it is a two-stage process. Refer to Decode Hierarchical Model section.

ensemble decoding ("token-level combination / product-of-expectations"):

python ensemble_decode_testset.py start_id end_id

Train Hierarchical Model

For performing sentence filtering, i.e. content selection

python train_hiermodel.py

Configurations are set inside this script.

Decode Hierarchical Model section

This is the first stage before BART decoding unless you do truncation at test time:

python hier_filtering_testset.py start_id end_id

If you want to train BART with filtered data:

python hier_filtering_trainset.py start_id end_id

Trained Weights

Download links (~8GB) - same file, any option that works for you (if none of the links work, please email me [email protected]):

[Google Drive] [Cambridge OneDrive]

BART

  • cued_speech_bart_baseline.pt (run_id: cued_speechUniv3)
  • cued_speech_bart_filtered.pt (run_id: cued_speechUniv4)
  • cued_speech_ensemble3_x.pt (run_id: cued_speechUniv2)
    • cued_speech_ensemble3_1.pt
    • cued_speech_ensemble3_2.pt
    • cued_speech_ensemble3_3.pt

Hierarchical Model

  • HIERMODEL_640_50_step30000.pt (trained using max num sentences = 640, max num word in sent = 50 for 30k steps)