<aside>
📄
Paper Unified Question Generation with Continual Lifelong Learning (Unified-QG)
Authors Wei Yuan, Hongzhi Yin, Tieke He, Tong Chen, Qiufeng Wang, Lizhen Cui
Venue WWW 2022 · arXiv 2022.01
Links arXiv
Keywords Question Generation Continual Learning Catastrophic Forgetting
</aside>
TL;DR
Unified-QG learns question generation continually across datasets and formats (extractive, abstractive, multiple-choice, boolean) with a single model. A format-converting encoding unifies inputs, and STRIDER (similarity-regularized difficult-example replay) prevents catastrophic forgetting.
1. Problem Definition
Existing QG methods are trained for specific datasets, which causes two limitations:
- They are dedicated to specific QG formats, so a new format requires redesigning the model.
- Best performance is only achieved on the dataset they were just trained on, so many QG models must be trained and maintained — resource-intensive and ungeneralizable.
2. Method
- Format-convert encoding: transform different QG formats into a unified text-to-text representation so one model can handle all of them.
- STRIDER (SimilariTy RegularIzed Difficult Example Replay):
- Difficult example replay: keep a small memory of hard examples from previous tasks and replay them when learning a new task.
- Similarity regularization: constrain parameter updates so that the model doesn't drift too far from what it learned on earlier tasks.
3. Findings
- On 8 QG datasets across 4 formats, Unified-QG continually adapts as datasets and formats change.
- A single trained Unified-QG model generates synthetic QA data that improves 8 different QA systems.
4. Takeaways
<aside>
💭
- My question at the time: could STRIDER's regularization be applied to parameter updates of a retriever as it adapts to new domains one after another?
- For retrieval, the analogous problem is adapting a dense retriever to a stream of new corpora without losing performance on earlier ones; a replay buffer of hard query–passage pairs is a natural fit.
- A single QG model that handles many formats is also useful for synthetic-query pipelines like QGen/GPL, which currently use one generator trained on one dataset.
</aside>
5. Related Papers
- Zero-shot Neural Passage Retrieval via Domain-targeted Synthetic Question Generation (QGen)
- Unsupervised Domain Adaptation for Question Generation with Domain Data Selection and Self-training