Unsupervised Transfer Learning for Spoken Language Understanding in Intelligent Agents
Authors: Aditya Siddhant, Anuj Goyal, Angeliki Metallinou4959-4966
AAAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our methods on various tasks and datasets, including data from Alexa, a popular commercial intelligent agent. Our results show that unsupervised transfer using unlabeled utterances can outperform both training from scratch and supervised pre-training. |
| Researcher Affiliation | Collaboration | Aditya Siddhant,1 Anuj Goyal,2 Angeliki Metallinou2 asiddhan@cs.cmu.edu, anujgoya@amazon.com, ametalli@amazon.com 1Carnegie Mellon University, 2Amazon Alexa AI |
| Pseudocode | No | The paper includes architectural diagrams (Figure 1 and Figure 2) but no formal pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide an explicit statement or link for open-source code related to the described methodology. |
| Open Datasets | Yes | For benchmarking our proposed methods we use two labeled public SLU datasets: ATIS and SNIPS. ATIS is a common SLU benchmark from the travel planning domain which contains 5K utterances (Hemphill, Godfrey, and Doddington 1990). SNIPS is a more recent SLU benchmark created by the company snips.ai... (Coucke et al. 2017). We also used two unlabeled datasets, the 1B Word Benchmark (1BWB)... (Chelba et al. 2014) and the 1M SLU Benchmark data (1MSLU)... including training splits from ATIS, SNIPS, DSTC2 (Henderson, Thomson, and Williams 2014), and others. |
| Dataset Splits | Yes | Table 2: Internal and public target datasets statistics: #Training Samples #Dev Samples #Test Samples. Domain-A 43168 3680 4752; Domain-B 100000 8227 8695; ATIS 4478 500 893; SNIPS 13084 700 700. Training is done upto 25 epochs with early-stopping based on sum of IC and ET scores on development set. |
| Hardware Specification | No | The paper mentions 'memory requirement' and 'faster inference' in general terms but does not specify any particular hardware components like CPU models, GPU models, or exact memory sizes used for experiments. |
| Software Dependencies | No | The paper mentions 'Adam optimizer (Kinmga and Ba 2015)' but does not provide specific version numbers for any software, libraries, or frameworks used (e.g., Python, TensorFlow, PyTorch versions). |
| Experiment Setup | Yes | Network Hyper-parameters The experimental setup was kept same for all 4 datasets with minor adjustments in hyperparameters. We use 200 hidden units for all three LSTM layers in our multi-task architecture. We use Adam optimizer (Kinmga and Ba 2015) with initial learning rate 0.0001 for internal datasets (Domain-A and Domain-B) and 0.0005 for ATIS and SNIPS. Both IC and ET losses are weighted equally in the total loss. Training is done upto 25 epochs with early-stopping based on sum of IC and ET scores on development set. Dropout probability is 0.5 for ATIS and SNIPS and 0.2 for internal datasets and we use L2 regularization on all weights with lambda=0.0001. |