Adversarial Semantic Decoupling for Recognizing Open-Vocabulary Slots

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 6070–6075, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 6070 Adversarial Semantic Decoupling for Recognizing Open-Vocabulary Slots Yuanmeng Yan1∗, Keqing He1∗, Hong Xu1, Sihong Liu1, Fanyu Meng2, Min Hu2, Weiran Xu1 1Beijing University of Posts and Telecommunications, Beijing, China 2China Mobile Research Institute, Beijing, China {yanyuanmeng,kqin,xuhong,liusihong,xuweiran}@bupt.edu.cn {mengfanyu,humin}@chinamobile.com Abstract Open-vocabulary slots, such as file name, al- bum name, or schedule title, significantly de- grade the performance of neural-based slot fill- ing models since these slots can take on val- ues from a virtually unlimited set and have no semantic restriction nor a length limit. In this paper, we propose a robust adversarial model-agnostic slot filling method that ex- plicitly decouples local semantics inherent in open-vocabulary slot words from the global context. We aim to depart entangled contex- tual semantics and focus more on the holistic context at the level of the whole sentence. Ex- periments on two public datasets show that our method consistently outperforms other meth- ods with a statistically significant margin on all the open-vocabulary slots without deterio- rating the performance of normal slots. 1 Introduction Slot filling is a critical component of spoken lan- guage understanding (SLU) in task-oriented dia- logue systems. It aims at extracting semantic con- stituents from the user queries. Given an immense amount of labeled training data, recent neural net- works (Mesnil et al., 2015; Liu and Lane, 2015, 2016; Goo et al., 2018; Haihong et al., 2019; Chen et al., 2019; He et al., 2020a,b) have been actively applied to slot filling task and achieved good re- sults. Although most previous neural-based models achieve state-of-the-art performance across a wide range of slot filling datasets, they often suffer from poor slot filling accuracy while dealing with ‘open- vocabulary’ slots. Open-vocabulary slots signify slot types that can take on values from a virtually unlimited set, such as file name, album name, text body, or schedule title. Typically, these slot values ∗The first two authors contribute equally. Weiran Xu is the corresponding author. add the song don’t drink the water to my playlist add the song don’t drink the water to my playlist Baseline Pred Ground Truth O O Figure 1: An error case of open-vocabulary slot “playlist” in Snips dataset (Coucke et al., 2018). Here “water” is mistakenly recognized as “entity name” type by the baseline model (Liu and Lane, 2016) due to the local context “don’t drink the water”. However, it rep- resents a playlist at the level of the whole sentence. album track poi genre served- dish entity- name movie- name cuisine restaurant- name object- name 0 10 20 30 40 50 60 70 Error rate (%) open-vocabulary slot normal slot Figure 2: Error rates of open-vocabulary slots com- pared to normal slots in Snips from Baseline (Liu and Lane, 2016). We display the top10 slot types of the highest error rates. have no constraints on the length and specific se- mantic patterns of content. Besides, these words are employed …

Content truncated to 500 words