Loading...

Improving the Training Process of Understanding Unit in Spoken Dialog Systems Using Active Learning Methods

Hadian, Hossein | 2013

771 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 45348 (19)
  4. University: Sharif University of Technology
  5. Department: Computer Engineering
  6. Advisor(s): Sameti, Hossein
  7. Abstract:
  8. This thesis aims at reducing the need for labeled data in the SLU domain by the means of active Learning methods. This need is due to the lack of labeled datasets for Spoken Language Understanding (SLU) in the Persian language, and fairly high labeling costs. Active learning methods enables the learner to choose the most informative instances to be labeled and used for training, and prevents labeling uninformative or redundant instances. For modeling the SLU system, several statistical models namely MLN (Markov Logic Networks), CRF (Conditional Random Fields), HMM (Hidden Markov Model) and HVS (Hidden Vector State) were reviewed, and finally CRF was chosen for its superior performance. The unlabeled ATIS (Air Travel Information System) dataset was labeled and was prepared to be used with CRF. State-of-the-art methods for active learning such as query-by-committee, uncertainty-sampling and information-density based methods were implemented and evaluated. All active learning methods were evaluated using the area under the accuracy learning curves on ATIS. To complete the research, a noise model for ATIS was designed. Using this noise model, different levels of noise were generated and active learning methods were evaluated in noisy conditions. To improve the existing methods in noisy conditions, three new methods were introduced: two based on the local density principle and the other based on the Fisher information framework. The results indicated that local density is a better measure than the total density for detecting noisy instances. In addition, evaluation results indicated that the three newly proposed methods perform better than the existing methods, both in noisy and noise-free conditions. As a result, the Local Density (LD) and Weighted Gradient Entropy methods (WGE) respectively show an improvement of 0.3% and 0.6% over the best known method being query-by-committee. Finally the Strict Local Density method (SLD) shows an improvement of 1%.
    Keywords: SLU, active learning, CRF, ATIS dataset, noise model, noise, information density, Fisher information, local density, strict local density, weighted gradient entropy
  9. Keywords:
  10. Active Learning ; Fault Model ; Noise ; Spoken Langauge Understanding ; Fisher Information Matrix ; Local Density Approximation ; Conditional Random Fields (CRF) ; Air Travel Information System (ATIS)Dataset ; Weighted Gradient Entropy

 Digital Object List

 Bookmark

No TOC