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
- 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.
paths.py
Data file, log file and other file pathssnips_data.py
Process and load Snips data, and SnipsDataset (torch.utils.data.Dataset)plot_data.py
Plot data informationmodel.py
Slotfilling network architectures (torch.nn.Module), including both Coarse and Coach SF models as introduced in our paperslot_filling.py
Coarse SF class, including train and test functionsslot_filling_coach.py
Coach SF class, including train and test functionscrf.py
Implementation of CRFtext.py
&utils.py
&tools.py
Some helper functionstrain_dyn.py
Train Coarse CDSF via dynamic transfer, i.e, sequentially adding source domains sorted by the shared slot numberstrain_dyn_coach.py
Train Coach CDSF via dynamic transfer, i.e, sequentially adding source domains sorted by the shared slot numbersanalyze_dyn.py
Analyze and plot the logged results
- 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
ortrain_dyn_coach.py
, and then runanalyze_dyn.py
to show the results
- 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]