Abstract
Recent developments in deep learning have led to a significant innovation invarious classic and practical subjects, including speech recognition, computervision, question answering, information retrieval and so on. In the context ofnatural language processing (NLP), language representations have shown giantsuccesses in many downstream tasks, so the school of studies have become amajor stream of research recently. Because the immenseness of multimedia dataalong with speech have spread around the world in our daily life, spokendocument retrieval (SDR) has become an important research subject in the pastdecades. Targeting on enhancing the SDR performance, the paper concentrates onproposing a neural retrieval framework, which assembles the merits of usinglanguage modeling (LM) mechanism in SDR and leveraging the abstractiveinformation learned by the language representation models. Consequently, to ourknowledge, this is a pioneer study on supervised training of a neural LM-basedSDR framework, especially combined with the pretrained language representationmethods.