Entity resolution for noisy ASR transcripts

Abstract

Large vocabulary domain-agnostic Automatic Speech Recognition (ASR) systems often mistranscribe domain-specific words and phrases. Since these generic ASR systems are the first component of most voice assistants in production, building Natural Language Understanding (NLU) systems that are robust to these errors can be a challenging task. In this paper, we focus on handling ASR errors in named entities, specifically person names, for a voicebased collaboration assistant. We demonstrate an effective method for resolving person names that are mistranscribed by black-box ASR systems, using character and phonemebased information retrieval techniques and contextual information, which improves accuracy by 40.8% on our production system. We provide a live interactive demo to further illustrate the nuances of this problem and the effectiveness of our solution. 1

KG-ECO: Knowledge Graph Enhanced Entity Correction for Query Rewriting

“…Our work is also closely related to two entity-level tasks: entity correction and entity linking. Entity correction [7,8,9,10] aims to tackle errors occurring in Automatic Speech Recognition (ASR) systems. These studies adopt an "entityto-entity" approach for entity correction, however, we take the query as context, and perform contextualized entity correction.…”

“…It is a two-layer system, retrieving and re-ranks utterances and NLU hypotheses. Since we do not assume the entities are tagged in utterances, "entity-to-entity" approaches [7,8,9,10] are not appropriate baselines. Besides gUFS-QR, we also consider our two-layer entity correction system without the KG components, which is a similar design as popular entity linking system [11], as a baseline.…”

End-to-End Spoken Language Understanding for Generalized Voice Assistants

“…Errors encountered in either model do not inform the other; in practice this means incorrect ASR transcriptions might be "correctly" interpreted by the NLU, fail to provide the user's desired response. While work is ongoing in detecting [8], quantifying [9,10], and rectifying [11,12] these ASR driven NLU misclassifications, end-to-end (E2E) approaches are a promising way to address this issue.…”

Neural text normalization leveraging similarities of strings and sounds

“…Noise generator.3.2 Deep MetaphoneRaghuvanshi et al (2019) revealed that relying only on surface text similarities cannot capture phonetic differences between words. Furthermore,Han et al (2013) showed that sound-related features are effective in text normalization.…”