Convolutional neural networks (CNNs) have shown promising results on many natural language understanding (NLU) tasks owing to their ability to capture informative features in local patches. However, they extract informative patterns in a static manner; they use the same set of filters regardless of different inputs. In this article, we propose an adaptive convolution to provide greater flexibility to traditional CNNs. Unlike the traditional convolution, the adaptive convolution utilizes adaptively generated convolution filters which are conditioned on inputs. We achieve this by attaching filter-generating networks, which are carefully designed to generate input-specific filters, to each convolution block in existing CNNs. We show the efficacy of our approach for existing CNNs through extensive performance evaluations. Our results indicate that adaptive convolutions improve all the baselines, without any exception, by as much as 2.6 percentage points (%p) on sentiment analysis, 1.6 %p on text classification, and 3.6 %p on textual entailment.
Bibliographical noteFunding Information:
We thank the anonymous reviewers for their helpful comments. This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) ( 2021R1A2C30 10430 ) and the Basic Research Program through the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) ( 2020R1A4A1018309 ).
© 2021 Elsevier Ltd
- Adaptive convolution
- Natural language understanding
- Network adaptation
ASJC Scopus subject areas
- Computer Science Applications
- Artificial Intelligence