Exploiting Sentence and Context Representations in Deep Neural Models for Spoken Language Understanding

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 258–267, Osaka, Japan, December 11-17 2016. Exploiting Sentence and Context Representations in Deep Neural Models for Spoken Language Understanding Lina M. Rojas-Barahona, Milica Gaˇsi´c, Nikola Mrkˇsi´c, Pei-Hao Su Stefan Ultes, Tsung-Hsien Wen and Steve Young Department of Engineering, University of Cambridge, Cambridge, UK lmr46, mg436, nm480, phs26, su259, thw28, sjy@cam.ac.uk Abstract This paper presents a deep learning architecture for the semantic decoder component of a Sta- tistical Spoken Dialogue System. In a slot-filling dialogue, the semantic decoder predicts the dialogue act and a set of slot-value pairs from a set of n-best hypotheses returned by the Auto- matic Speech Recognition. Most current models for spoken language understanding assume (i) word-aligned semantic annotations as in sequence taggers and (ii) delexicalisation, or a mapping of input words to domain-specific concepts using heuristics that try to capture morphological variation but that do not scale to other domains nor to language variation (e.g., morphology, syn- onyms, paraphrasing ). In this work the semantic decoder is trained using unaligned semantic annotations and it uses distributed semantic representation learning to overcome the limitations of explicit delexicalisation. The proposed architecture uses a convolutional neural network for the sentence representation and a long-short term memory network for the context representa- tion. Results are presented for the publicly available DSTC2 corpus and an In-car corpus which is similar to DSTC2 but has a significantly higher word error rate (WER). 1 Introduction In most existing work on Spoken Language Understanding (SLU), semantic decoding is usually seen as a sequence tagging problem with models trained and tested on datasets with word-level annotations (T¨ur et al., 2013; Mesnil et al., 2015; Yao et al., 2013; Sarikaya et al., 2011; Deoras and Sarikaya, 2013; Sarikaya et al., 2014). Spoken language understanding from unaligned data, in which utterances are annotated with an abstract semantics, faces the additional challenge of not knowing which specific words are relevant for extracting the semantics. This problem was tackled in (Zhou and He, 2011), by using conditional random fields (CRFs) driven by finely-tuned hand-crafted features. Other discriminative approaches that deal with unaligned data use some form of delexicalisation or mapping of the input to known ontological concepts (Henderson et al., 2012; Henderson et al., 2014a). The main disadvantage of delexicalisation is the difficulty in scaling it, not only to larger and more complex dialogue domains but also to handle the many forms of language variation. We propose in this paper a semantic decoder that learns from unaligned data (Figure 1) and that exploits rich semantic distributed word representations instead of delexicalisation. The semantic decoder predicts the dialogue act and the set of slot-value pairs from a set of n-best hypotheses returned by an automatic speech recognition (ASR). The prediction is made in two steps. First, a deep learning architecture is used for the joint prediction of dialogue acts and the presence or absence of slots. Second, the same architecture is reused for predicting the values …

Content truncated to 500 words