Word embeddings: A semi-supervised learning method for slot-filling in spoken dialog systems

One of the key components in spoken dialog systems is semantic slot-filling, a sequence tagging task. There are several state-of-the-art supervised approaches to model the slot-filling problem such as conditional random fields (CRF), support vector machine (SVM) and stochastic finite state transducers (SFST). A general way to improve their performance is to use unsupervised word embeddings as extra input features. In this paper we evaluate two kinds of word embeddings on all the three approaches for slot-filling. We use three near state-of-the-art supervised baselines, and find that each of them can be improved by plugging word embeddings into the existing systems. Experiments on the ATIS benchmark show that our work outperforms the baseline by 2.2% in relative F1-score increase at least. Under a noisy automatic speech recognition (ASR) condition, our best system outperforms the state-of-the-art CRF baseline by 9.6% in relative F1-score increase.