Effective data-driven feature learning for detecting name errors in automatic speech recognition

Abstract

This paper addresses the problem of detecting name errors in automatic speech recognition (ASR) output. The highly skewed label distributions (i.e. name errors are infrequent), sparse training data, and large number of potential lexical features pose significant challenges for training name error classification systems. Data-driven feature learning is needed for handling multiple languages but is sensitive to over fitting. We address the problem by designing aggregate features using a related (sentence-level name detection) task, and reduce dimensionality of the lexical features using word classes. Experiments on conversational domain data in both English and Iraqi Arabic show that best results are obtained using all feature mapping methods plus feature selection using L1 regularization.

Open-Domain Name Error Detection using a Multi-Task RNN

“…These features are combined in a maximum entropy (ME) classifier trained to predict name errors directly. This is the same as the baseline used in (He et al, 2014;Marin, 2015;, but with a different ASR system.…”

“…In this paper, we address these general problems -detecting rare events in an open-domain taskspecifically for name error detection. Prior work addressed the problem of skewed priors by artificially increasing the error rate by holding names out of the vocabulary (Chen et al, 2013) or by factoring the problem into sentence-level name detection and OOV word detection (He et al, 2014) (since OOV errors in general are more frequent than name errors). Sentence-level features are also shown to be more robust than local context in direct name error prediction (Marin, 2015).…”

“…In addition, since the words abutting the name are less reliable in a recognition hypothesis, the information lost by working at the sentence level is minimal. The idea of using sentence-level name prediction is proposed in (He et al, 2014), but in that work the sentence name posterior is a feature in the ME model (optionally with word cues learned by the sentence-level predictor). In our work, the problem is factored to use the acoustic confusibility and local word class features for OOV error prediction, which is combined with the sentence-level name posterior for name error prediction.…”

“…• Word Context. This is the baseline ME system described in Section 2 and also used as a baseline in (He et al, 2014;, which directly predicts the word-level name error using WCN structural features, current word Brown class, and up-to trigram left and right word context information. • Word Class.…”