Balanced Joint Adversarial Training for Robust Intent Detection and Slot Filling

Proceedings of the 28th International Conference on Computational Linguistics, pages 4926–4936 Barcelona, Spain (Online), December 8-13, 2020 4926 Balanced Joint Adversarial Training for Robust Intent Detection and Slot Filling Xu Cao1,3, Deyi Xiong2, Chongyang Shi1∗, Chao Wang3, Yao Meng3, Changjian Hu3 1 School of Computer Science, Beijing Institute of Technology, Beijing , China 2 College of Intelligence and Computing, Tianjin University, Tianjin, China 3 Lenovo Research AI Lab, Beijing, China bitcaoxu@gmail.com dyxiong@tju.edu.cn cy shi@bit.edu.cn {wangchao31, mengyao1, hucj1}@lenovo.com Abstract Joint intent detection and slot filling has recently achieved tremendous success in advancing the performance of utterance understanding. However, many joint models still suffer from the robustness problem, especially on noisy inputs or rare/unseen events. To address this issue, we propose a Joint Adversarial Training (JAT) model to improve the robustness of joint intent detection and slot filling, which consists of two parts: (1) automatically generating joint adversarial examples to attack the joint model, and (2) training the model to defend against the joint adversarial examples so as to robustify the model on small perturbations. As the generated joint adversarial examples have different impacts on the intent detection and slot filling loss, we further propose a Balanced Joint Adversarial Training (BJAT) model that applies a balance factor as a regularization term to the final loss function, which yields a stable training procedure. Extensive experiments and analyses on the lightweight models show that our proposed methods achieve significantly higher scores and substantially improve the robustness of both intent detection and slot filling. In addition, the combination of our BJAT with BERT-large achieves state-of-the-art results on two datasets. 1 Introduction Intent detection and slot filling are two critical components in dialogue systems. Although these two tasks are normally considered as parallel tasks, they may inherently correlate with each other as one is helpful in defining the other. Some existing works (Liu and Lane, 2016; Vu, 2016; Zhang and Wang, 2016) indicate that sharing parameters in an encoder module to simultaneously detect the intent and fill the slots of an utterance can utilize such correlation between the two tasks. The joint modeling of them has therefore achieved tremendous success (Goo et al., 2018; Li et al., 2018; E et al., 2019) and further improved the performance of spoken language understanding (SLU) systems. However, recent outstanding joint models still show insufficient robustness. For example, some small changes of inputs can mislead the models to give wrong prediction. As shown in Figure 1 (a), misspelled words “SanFranciso” (missing a blank) and “pitsburgh” (missing “t”) are easily to fool the neural models to produce wrong slot predictions. One might think that such wrong predictions are related to the input word embeddings and using character embeddings can avoid these prediction errors. However, we find in our preliminary experiments that character embeddings cannot fully eliminate such errors as it is not easy to reconstruct the meaning of corresponding slot words based on character embeddings. In Figure 1 (b), word-level perturbations such as inserted word “video” or substituted word “movie” can …

Content truncated to 500 words