Name
Abstract | ||
|---|---|---|
Intent classification and slot filling are two classical problems for spoken language understanding and dialog systems. The existing works, either accomplishing intent classification or slot filling separately or using a joint model, are all human-designed models with trial and error. In order to explore the variety of network architecture and to find whether there exist possible network architectures with better results, we proposed the D-GHNAS (Deep deterministic policy gradient based Graph Hypernetwork Neural Architecture Search) to accomplish intent classification and slot filling via a NAS (Neural Architecture Search) method. NAS based techniques can automatically search for network architectures without experts’ trial and error. Different from early NAS methods with hundreds of GPU days to find an ideal neural architecture that takes too much computation resource, in this work, hypernetwork is used to decrease the computation cost. Experimental results demonstrate that our model improves intent classification and slot filling results on public benchmark datasets ATIS and SNIPS compared with other joint models for these tasks. |
Year | DOI | Venue |
|---|---|---|
2020 | 10.1007/978-3-030-60259-8_58 | Interational Conference on Web-Age Information Management |
DocType | Citations | PageRank |
Conference | 0 | 0.34 |
References | Authors | |
0 | 5 | |
Name | Order | Citations | PageRank |
|---|---|---|---|
| Yanxi Tang | 1 | 0 | 0.34 |
| Jianzong Wang | 2 | 61 | 34.65 |
| Xiaoyang Qu | 3 | 1 | 0.68 |
| Nan Zhang | 4 | 1 | 1.70 |
| jing xiao | 5 | 80 | 42.68 |