Exploring and Predicting Transferability across NLP Tasks
arXiv:2005.00770
Abstract
Recent advances in NLP demonstrate the effectiveness of training large-scale language models and transferring them to downstream tasks. Can fine-tuning these models on tasks other than language modeling further improve performance? In this paper, we conduct an extensive study of the transferability between 33 NLP tasks across three broad classes of problems (text classification, question answering, and sequence labeling). Our results show that transfer learning is more beneficial than previously thought, especially when target task data is scarce, and can improve performance even when the source task is small or differs substantially from the target task (e.g., part-of-speech tagging transfers well to the DROP QA dataset). We also develop task embeddings that can be used to predict the most transferable source tasks for a given target task, and we validate their effectiveness in experiments controlled for source and target data size. Overall, our experiments reveal that factors such as source data size, task and domain similarity, and task complexity all play a role in determining transferability.
arXiv commentAccepted as a conference paper at EMNLP 2020, 45 pages, 3 figures, 34 tables
References in corpus (9)
- HuggingFace's Transformers: State-of-the-art Natural Language Processing
- An Overview of Multi-Task Learning in Deep Neural Networks
- The Natural Language Decathlon: Multitask Learning as Question Answering
- Sentence Encoders on STILTs: Supplementary Training on Intermediate Labeled-data Tasks
- Compositional Sequence Labeling Models for Error Detection in Learner Writing
- How Transferable are Neural Networks in NLP Applications?
- Multi-Task Learning for Sequence Tagging: An Empirical Study
- Semantic Tagging with Deep Residual Networks
- Modelling Latent Skills for Multitask Language Generation