/CDSF-Transferability

Primary LanguagePython

CDSF-Transferability

The source code of our ICASSP 2022 paper: Exploring Transferability Measures and Domain Selection in Cross-Domain Slot Filling.

  • Personal Homepage
  • Basic Introduction
  • Code Files
  • Running Tips
  • Citation

Personal Homepage

Basic Introduction

  • During Cross Domain Slot Filling (CDSF), when does a source model help the target task?
  • If a source task contains slot types that the target one does not own (non-targeted slots) or the target has new slot types to identify (zero-shot slots), does the transfer process become better or worse?
  • We implement CDSF on the Snips [1] benchmark which contains 7 domains, and investigate the domain transferability and negative transfer in CDSF.

Code Files

  • paths.py Data file, log file and other file paths
  • snips_data.py Process and load Snips data, and SnipsDataset (torch.utils.data.Dataset)
  • plot_data.py Plot data information
  • model.py Slotfilling network architectures (torch.nn.Module), including both Coarse and Coach SF models as introduced in our paper
  • slot_filling.py Coarse SF class, including train and test functions
  • slot_filling_coach.py Coach SF class, including train and test functions
  • crf.py Implementation of CRF
  • text.py & utils.py & tools.py Some helper functions
  • train_dyn.py Train Coarse CDSF via dynamic transfer, i.e, sequentially adding source domains sorted by the shared slot numbers
  • train_dyn_coach.py Train Coach CDSF via dynamic transfer, i.e, sequentially adding source domains sorted by the shared slot numbers
  • analyze_dyn.py Analyze and plot the logged results

Running Tips

  • Snips Data: the utilized data is the same as and downloaded from Coach
  • Word Embeddings: we use both word-level [2] and character-level [3] embeddings to obtain 400d vectors for tokens and slot descriptions, the detailed implementations could be found in text.py (we use the provided embeddings in torchtext)
  • Package Versions: python==3.7.3, nltk=3.6.2, torch==1.9.0, torchtext==0.10.0
  • Running: directly run train_dyn.py or train_dyn_coach.py, and then run analyze_dyn.py to show the results

Citation

  • Xin-Chun Li, Yan-Jia Wang, Le Gan, De-Chuan Zhan. Exploring Transferability Measures and Domain Selection in Cross-Domain Slot Filling. In: Proceedings of the 2022 IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP'2022), online conference, Singapore, 2022.
  • [BibTex]