Transfer Learning for Sequences via Learning to Collocate

Wanyun Cui, Guangyu Zheng, Zhiqiang Shen, Sihang Jiang, Wei Wang

Abstract:

Transfer learning aims to solve the data sparsity for a specific domain by applying information of another domain. Given a sequence (e.g. a natural language sentence), the transfer learning, usually enabled by recurrent neural network (RNN), represent the sequential information transfer. RNN uses a chain of repeating cells to model the sequence data. However, previous studies of neural network based transfer learning simply transfer the information across the whole layers, which are unfeasible for seq2seq and sequence labeling. Meanwhile, such layer-wise transfer learning mechanisms also lose the fine-grained cell-level information from the source domain.

In this paper, we proposed the aligned recurrent transfer, ART, to achieve cell-level information transfer. ART is in a recurrent manner that different cells share the same parameters. Besides transferring the corresponding information at the same position, ART transfers information from all collocated words in the source domain. This strategy enables ART to capture the word collocation across domains in a more flexible way. We conducted extensive experiments on both sequence labeling tasks (POS tagging, NER) and sentence classification (sentiment analysis). ART outperforms the state-of-the-arts over all experiments.

Keywords: transfer learning, recurrent neural network, attention, natural language processing

TL;DR: Transfer learning for sequence via learning to align cell-level information across domains.