End-to-end named entity recognition for Vietnamese speech
Thu-Hien T. Nguyen, Thai Binh Nguyen, Quoc Truong Do, Tuan-Linh Nguyen · 2022
One of the first steps in comprehending natural or spoken language is named-entity recognition. It plays a critical role in natural language processing applications such as text clustering, subject detection, topic detection, and text summarization. The most common techniques in information extraction using named-entity recognition can be applied to building automatic question-answering systems, semantic web technology, automatic translation machines, and so on. For input text with correct formatting, studies on the named entity recognition (NER) task have achieved excellent and nearly human-equal results. However, for the output documents of the speech recognition system (ASR), the normative signs of the text, such as punctuation and uppercase, no longer exist, causing difficulties for the researchers. In this paper, we present the process of building a Vietnamese speech dataset for the NER task and propose a new end-to-end approach for the NER of Vietnamese speech. In addition, we combine the multi-task learning model with the punctuation and uppercase (CaPu) recovery model and demonstrate the combination’s effectiveness when documenting the approximately 5% improvement in the F1 score.