Improving Spoken Language Understanding with information retrieval and active learning methods

Abstract

In the context of deployed spoken dialogue telecom services, we introduce a preprocessor called Fiction into the Spoken Language Understanding (SLU) component. It acts as an intermediate between the speech recognition and interpretation processes in order to increase the rate of utterances that are correctly rejected (CRR for Correctly Rejected Rate) without decreasing the rate of appropriately interpreted utterances. This component is based on statistical approaches of natural language treatment and contextual information. We also use active learning methods to determine the best training corpus size. On a deployed test corpus, the CRR increases from 60% to 86% and active learning method's results show that better performance can be achieved using fewer training data.

Data Balancing for Efficient Training of Hybrid ANN/HMM Automatic Speech Recognition Systems

“…However, with the growth of databases in the last few years, a question about the eventual saturation of that lemma has been raised [15] adding a preoccupation about an adequate treatment of erroneously labeled samples. Several active learning solutions implying a selection of the data have been proposed not only in ASR but also in Spoken Language Processing [31], [32], emotion recognition [33], or language identification [34] among others. It is, however, worth noting that the main goal of active learning is the improvement in the precision of the target classifier disregarding, most of the time, the computational costs.…”