Relation Extraction with Explanation

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 609–618 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 609 Learning Low-Resource End-To-End Goal-Oriented Dialog for Fast and Reliable System Deployment Yinpei Dai†, Hangyu Li†, Chengguang Tang†, Yongbin Li†∗, Jian Sun†, Xiaodan Zhu‡ †Alibaba Group, Beijing ‡Ingenuity Labs Research Institute & ECE, Queen’s University {yinpei.dyp,hangyu.lhy,chengguang.tcg}@alibaba-inc.com {shuide.lyb,jian.sun}@alibaba-inc.com, zhu2048@gmail.com Abstract Existing end-to-end dialog systems perform less effectively when data is scarce. To ob- tain an acceptable success in real-life online services with only a handful of training exam- ples, both fast adaptability and reliable perfor- mance are highly desirable for dialog systems. In this paper, we propose the Meta-Dialog System (MDS), which combines the advan- tages of both meta-learning approaches and human-machine collaboration. We evaluate our methods on a new extended-bAbI dataset and a transformed MultiWOZ dataset for low- resource goal-oriented dialog learning. Exper- imental results show that MDS significantly outperforms non-meta-learning baselines and can achieve more than 90% per-turn accura- cies with only 10 dialogs on the extended- bAbI dataset. 1 Introduction End-to-end neural models have shown a great po- tential in building flexible goal-oriented dialog sys- tems. They can be directly trained on past dialogs without any domain-specific handcrafting, which makes it easy to automatically scale up to new do- mains (Bordes et al., 2017). However, these mod- els are normally data-hungry and have only been successfully applied to domains with rich datasets (Perez et al., 2017; Luo et al., 2019; Kim et al., 2019). In real-world scenarios, common issues with end-to-end dialog models include: (1) the short- age of proper training dialogs because of the high cost of data collection and cleaning, i.e., the data scarcity problem (Zhao and Eskenazi, 2018), and (2) a large gap between limited data and unknown online test examples, i.e., the covariate shift effect (Liu et al.). Such problems can lead to a significant performance degradation in dialog systems, which ∗Corresponding author may harm the users’ experience and result in loss of customers in commercial applications. Therefore, both fast adaptability and reliable performance are strongly desirable for practical system deployment. Fast adaptability reflects the efficiency of adapt- ing dialog systems to domains with low-resource data. Reliable performance reflects the robustness of handling unpredictable user behaviors in online services. To boost the online performance of dialog sys- tems, there have been some recent work (Rajendran et al., 2019; Wang et al., 2019; Lu et al., 2019) on designing end-to-end models in a human-machine joint-teaming manner. For instance, the dialog sys- tem in (Rajendran et al., 2019) can identify an on- going dialog during testing when the system might fail and transfer it to a human agent. But all these methods are trained with sufficient data, which hin- ders the possibility of rapidly prototyping the mod- els in new domains with restricted resources. In this paper, we formulate the low-resource goal-oriented dialog learning as a few-shot learning problem, where a limited numbers of dialogs are used for training and the remaining …

Content truncated to 500 words