Joint Multiple Intent Detection and Slot Labeling for Goal-Oriented Dialog

Proceedings of NAACL-HLT 2019, pages 564–569 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 564 Joint Multiple Intent Detection and Slot Labeling for Goal-Oriented Dialog Rashmi Gangadharaiah AWS AI, Amazon rgangad@amazon.com Balakrishnan Narayanaswamy AWS AI, Amazon muralibn@amazon.com Abstract Neural network models have recently gained traction for sentence-level intent classification and token-based slot-label identification. In many real-world scenarios, users have multi- ple intents in the same utterance, and a token- level slot label can belong to more than one intent. We investigate an attention-based neu- ral network model that performs multi-label classification for identifying multiple intents and produces labels for both intents and slot- labels at the token-level. We show state- of-the-art performance for both intent detec- tion and slot-label identification by comparing against strong, recently proposed models. Our model provides a small but statistically signif- icant improvement of 0.2% on the predomi- nantly single-intent ATIS public data set, and 55% intent accuracy improvement on an inter- nal multi-intent dataset. 1 Introduction In dialog systems, the natural language under- standing component (NLU) is responsible for identifying the user’s request and creating a se- mantic frame that succinctly summarizes the user’s needs. These semantic frames are typi- cally constructed using intents and slot-labels (T¨ur et al., 2010). As the names imply, an intent cap- tures the intention of the user and slot-labels cap- ture any additional information or constraints the user provides. These constraints must be satisfied in order to fulfill the user’s request. The example below shows a user’s request, “how is the weather in Dallas ?”. We need to identify the intent (“GetWeatherInfo”) as well as the values for the slot-labels (SL), here, “City” (value=“Dallas”). It is crucial that intents and slot-labels are identi- fied with high accuracy as an error made by the NLU component propagates through downstream components such as the dialog state tracker, the dialog policy and the natural language generator components, leading to a substantial degradation of the performance of the entire dialog system. NLU Semantic Frame how is the weather in Dallas ? ↕ ↕ ↕ ↕ ↕ ↕ ↕ SL: O O O O O City O Intent: GetWeatherInfo Intent detection has been modeled as a sentence classification task where an intent (yI) is assigned to the user’s utterance. Slot labeling is typically modeled as a sequential labeling problem, where a user’s sentence, x1, x2, ...xN, is labeled with yS 1 , yS 2 , ..yS N, and yS i is the slot label assigned to the token at position i (xi). In the example above, the sequence of slot labels would be, “O O O O O City O”, where, “O” stands for “Other”. Sequential models such as Maximum Entropy Markov models (Toutanova and Manning, 2000; McCallum et al., 2000; Berger et al., 1996) and Conditional Random Fields, CRFs (Lafferty et al., 2001; Jeong and Geunbae Lee, 2008) are popu- lar approaches for slot-labeling while intent pre- diction is often performed using standard classi- fication approaches such as …

Content truncated to 500 words