Jointly Modeling Inter-Slot Relations by Random Walk on Knowledge Graphs for Unsupervised Spoken Language Understanding

Human Language Technologies: The 2015 Annual Conference of the North American Chapter of the ACL, pages 619–629, Denver, Colorado, May 31 – June 5, 2015. c⃝2015 Association for Computational Linguistics Jointly Modeling Inter-Slot Relations by Random Walk on Knowledge Graphs for Unsupervised Spoken Language Understanding Yun-Nung Chen, William Yang Wang, and Alexander I. Rudnicky School of Computer Science, Carnegie Mellon University 5000 Forbes Avenue, Pittsburgh, PA 15213-3891, USA {yvchen, yww, air}@cs.cmu.edu Abstract A key challenge of designing coherent seman- tic ontology for spoken language understand- ing is to consider inter-slot relations. In prac- tice, however, it is difficult for domain experts and professional annotators to define a coher- ent slot set, while considering various lexi- cal, syntactic, and semantic dependencies. In this paper, we exploit the typed syntactic de- pendency theory for unsupervised induction and filling of semantics slots in spoken dia- logue systems. More specifically, we build two knowledge graphs: a slot-based seman- tic graph, and a word-based lexical graph. To jointly consider word-to-word, word-to- slot, and slot-to-slot relations, we use a ran- dom walk inference algorithm to combine the two knowledge graphs, guided by dependency grammars. The experiments show that con- sidering inter-slot relations is crucial for gen- erating a more coherent and compete slot set, resulting in a better spoken language under- standing model, while enhancing the inter- pretability of semantic slots. 1 Introduction An important requirement for building a success- ful spoken dialogue system (SDS) is to define a co- herent slot set and the corresponding slot-fillers for the spoken language understanding (SLU) compo- nent. Unfortunately, since the semantic slots are of- ten mutually-related, it is non-trivial for domain ex- perts and professional annotators to design a such slot set for semantic representation of SLU. Considering a restaurant domain (Henderson et al., 2012), “restaurant” is the target slot, and impor- tant adjective modifiers such as “Asian” (the restau- rant type) and “cheap” (the price of the restaurant) should be included in the slot set, so that the se- mantic representation of SLU can be more coherent and complete. In this case, it is challenging to de- sign such a coherent and complete slot set manually, while considering various lexical, syntactic, and se- mantic dependencies. Instead of considering slots independently, this paper takes a data-driven approach to model word- to-word relations via syntactic dependencies and further infer slot-to-slot relations. To do this, we incorporate the typed dependency grammar the- ory (De Marneffe and Manning, 2008) in a state- of-the-art frame-semantic driven unsupervised slot induction framework (Chen et al., 2013b). In par- ticular, we build two knowledge graphs: a slot- based semantic knowledge graph, and a word-based lexical knowledge graph. Using typed dependency triples, we then study the stochastic relations be- tween slots and words, using a mutually-reinforced random walk inference procedure to combine the two knowledge graphs. In evaluations, we use the jointly learned inter-slot relations to induce a coher- ent slot set in an unsupervised fashion. Our contri- butions are three-fold: • We are among the …

Content truncated to 500 words