Slot Filling with Weighted Multi-Encoders for Out-of-Domain Values

Abstract

This paper proposes a new method for slot filling of out-ofdomain (OOD) slot values, which are not included in the training data, in spoken dialogue systems. Word embeddings have been proposed to estimate the OOD slot values included in the word embedding model from keyword information. At the same time, context information is an important clue for estimation because the values in a given slot tend to appear in similar contexts. The proper use of either or both keyword and context information depends on the sentence. Conventional methods input a whole sentence into an encoder and extract important clues by the attention mechanism. However, it is difficult to properly distinguish context and keyword information from the encoder outputs because these two features are already mixed. Our proposed method uses two encoders, which distinctly encode contexts and keywords, respectively. The model calculates weights for the two encoders based on a user utterance and estimates a slot with weighted outputs from the two encoders. Experimental results show that the proposed method achieves a 50% relative improvement in F1 score compared with a baseline model, which detects slot values from user utterances and estimates slots at once with a single encoder.

Interactive Teaching for Conversational AI

“…This paper also introduced a new dataset for this problem that is publicly available, which is used in our work as one of the datasets for evaluating model performances. In [11,12] the authors proposed augmenting the training data for the slot-filling model, by injecting noise tokens into the regular slot values of training data, to force the model to learn about the context of the respective slots. This augmentation is done to improve the model robustness in predicting the correct span, on unseen slot values.…”

A Study on the Impacts of Slot Types and Training Data on Joint Natural Language Understanding in a Spanish Medication Management Assistant Scenario

“…In previous years, several studies were carried out to improve and evaluate the performance of NLU tasks. Former works evaluated the impact of using different techniques to improve slot filling [ 4 , 5 , 6 ] and intent detection [ 7 , 8 ]. Interestingly, research in joint NLU (jointly learning both intent detection and slot filling) achieved better results in both tasks [ 9 , 10 ].…”