Exploring deep learning architectures coupled with CRF based prediction for slot-filling

Abstract

Slot-filling is one of the most crucial module of any dialogue system that focuses on extracting relevant and necessary information from the user utterances. In this paper, we propose variants of Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) models for the task of slot-filling which includes LSTM/GRU networks, Bi-directional LSTM/GRU (Bi-LSTM/GRU) networks, LSTM/GRU-CRF and Bi-LSTM/GRU-CRF networks. Variants of LSTM/GRU is used for discourse modeling i.e., to capture long term dependencies in the input sentences. A Conditional Random Field (CRF) layer is integrated with the above network to capture the sentence level tag information. We show the experimental results of our proposed model on the benchmark Air Travel Information System (ATIS) dataset which indicate that our model performed exceptionally well compared to the state of the art.