Graph LSTM with Context-Gated Mechanism for Spoken Language Understanding
Authors: Linhao Zhang, Dehong Ma, Xiaodong Zhang, Xiaohui Yan, Houfeng Wang9539-9546
AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our extensive evaluation shows that the proposed model outperforms the state-of-the-art results by a large margin. Experiments Datasets To fully evaluate the proposed model, we conducted experiments on the Snips and ATIS datasets. The statistics of these two datasets are shown in Table 1. |
| Researcher Affiliation | Collaboration | Linhao Zhang,1 Dehong Ma,1 Xiaodong Zhang,1 Xiaohui Yan,2 Houfeng Wang1 1MOE Key Lab of Computational Linguistics, Peking University, Beijing, 100871, China 2CBG Intelligence Engineering Dept, Huawei Technologies, China {zhanglinhao, madehong, zxdcs, wanghf}@pku.edu.cn yanxiaohui2@huawei.com |
| Pseudocode | No | The paper does not contain any structured pseudocode or algorithm blocks. It describes the model components and equations but not in pseudocode format. |
| Open Source Code | No | The paper does not include any statement about releasing source code or provide a link to a code repository. |
| Open Datasets | Yes | To fully evaluate the proposed model, we conducted experiments on the Snips and ATIS datasets. The Snips dataset was created by snips.ai (Coucke et al. 2018). ... ATIS The Airline Travel Information Systems (ATIS) (Hemphill, Godfrey, and Doddington 1990) dataset has long been used as benchmark in SLU. ... In this work, we use the same one as used in Goo et al.; Niu et al.(2018; 2019). |
| Dataset Splits | Yes | Snips ... There are 13,084 utterances in the training set and 700 utterances in the test set, with a development set of 700 utterances. ... ATIS ... There are 4,478 utterances in the training set, 500 in the valid set and 893 in the test set, with a total of 120 distinct slot labels and 21 different intent types. |
| Hardware Specification | Yes | The model was implemented in Py Torch and trained on a single NVIDIA Ge Force GTX 1080 GPU. |
| Software Dependencies | No | The paper states "The model was implemented in Py Torch" and mentions using "Adam (Kingma and Ba 2015)" for optimization. However, it does not specify version numbers for PyTorch or any other software dependencies. |
| Experiment Setup | Yes | The dimensions of S-LSTM hidden state is set to 150. The dimensions of ELMo embeddings are 1024. ... The batch size is set to 32. Dropout (Hinton et al. 2012) layers are applied on both input and output vectors during training for regularization, with a dropout rate of 0.5. We use Adam (Kingma and Ba 2015) for the training process to minimize the cross-entropy loss, with learning rate = 10 3, β1 = 0.9, β2 = 0.98 and ϵ = 10 9. ... we set α in Equation 14 to be 1. ... We set χ = 3 and T = 6 for the Snips dataset. For the ATIS, we set χ = 3 and T = 4. |