Data Augmentation for Spoken Language Understanding via Joint Variational Generation
Authors: Kang Min Yoo, Youhyun Shin, Sang-goo Lee7402-7409
AAAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experiments show that existing SLU models trained on the additional synthetic examples achieve performance gains. Our approach not only helps alleviate the data scarcity issue in the SLU task for many datasets but also indiscriminately improves language understanding performances for various SLU models, supported by extensive experiments and rigorous statistical testing. |
| Researcher Affiliation | Academia | Kang Min Yoo, Youhyun Shin, Sang-goo Lee Seoul National University, Seoul 08826, Korea {kangminyoo, shinu89, sglee}@europa.snu.ac.kr |
| Pseudocode | Yes | Algorithm 1: Monte Carlo posterior sampling. input : a sufficiently large number m given : Dw, θ, φ output: synthetic utterance list U initialize U as an empty list; while U has less than m samples do sample a real utterance w from Dw; estimate the mean z of the posterior qφ (z|w); sample ˆw from the likelihood pθ (w| z); append ˆw to U; end return U |
| Open Source Code | Yes | The code is available on github (kaniblu/ludus-jluva). |
| Open Datasets | Yes | In this paper, we carry out experiments on the following language understanding datasets. ATIS: Airline Travel Information System (ATIS) (Hemphill, Godfrey, and Doddington 1990) is a representative dataset in the SLU task... Snips: The snips dataset is an open source virtual-assistant corpus. |
| Dataset Splits | Yes | Table 1: Dataset statistics. Training sets of ATIS (Small) and ATIS (Medium) have been chunked from the training set of ATIS (Full). The table provides specific numbers for 'Train Val Test' splits for ATIS-small, ATIS-medium, ATIS, Snips, MIT Movie Eng, MIT Movie Trivia, and MIT Restaurant datasets. |
| Hardware Specification | No | The paper does not provide specific details about the hardware (e.g., CPU, GPU models, memory) used for running the experiments. |
| Software Dependencies | No | The paper mentions using the Adam optimizer and the conlleval perl script, but it does not specify versions for any key software dependencies or libraries. |
| Experiment Setup | Yes | Implementation Details section provides specific experimental setup details, including: 'The word (Ww), slot label (Ws), and intent (Wy) embeddings have dimensions of 300, 200, and 100 respectively... The encoder network, a single-layer Bi LSTM-Max model... (1024 hidden dimensions). The decoders are uni-directional single-layer LSTMs with the same hidden dimensions (1024)... The beam search size was set to 15... Exploratory hyperparameter λs was 0.18... KLD annealing rate (kd) was 0.03 and word dropout rate pw was 0.5. We used Adam optimizer with 0.001 initial learning rate.' |