Unsupervised KB-to-Text Generation with Auxiliary Triple Extraction using Dual Learning

Proceedings of the 1st Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 10th International Joint Conference on Natural Language Processing, pages 258–268 December 4 - 7, 2020. c⃝2020 Association for Computational Linguistics 258 Unsupervised KB-to-Text Generation with Auxiliary Triple Extraction using Dual Learning∗ Zihao Fu1, Bei Shi2, Lidong Bing3, Wai Lam1 1Department of Systems Engineering and Engineering Management, The Chinese University of Hong Kong 2Tencent AI Lab 3DAMO Academy, Alibaba Group zhfu@se.cuhk.edu.hk; beishi@tencent.com; l.bing@alibaba-inc.com; wlam@se.cuhk.edu.hk Abstract The KB-to-text task aims at generating texts based on the given KB triples. Traditional methods usually map KB triples to sentences via a supervised seq-to-seq model. However, existing annotated datasets are very limited and human labeling is very expensive. In this paper, we propose a method which trains the generation model in a completely unsuper- vised way with unaligned raw text data and KB triples. Our method exploits a novel dual train- ing framework which leverages the inverse re- lationship between the KB-to-text generation task and an auxiliary triple extraction task. In our architecture, we reconstruct KB triples or texts via a closed-loop framework via linking a generator and an extractor. Therefore the loss function that accounts for the reconstruc- tion error of KB triples and texts can be used to train the generator and extractor. To re- solve the cold start problem in training, we propose a method using a pseudo data gen- erator which generates pseudo texts and KB triples for learning an initial model. To resolve the multiple-triple problem, we design an allo- cated reinforcement learning component to op- timize the reconstruction loss. The experimen- tal results demonstrate that our model can out- perform other unsupervised generation meth- ods and close to the bound of supervised meth- ods. 1 Introduction Knowledge Base (KB)-to-text task focuses on gen- erating plain text descriptions from given knowl- edge bases (KB) triples which makes them accessi- ble to users. For instance, given a KB triple <101 Helena, discoverer, James Craig Watson>, it is ex- pected to generate a description sentence such as ∗ The work described in this paper is substantially sup- ported by a grant from the Research Grant Council of the Hong Kong Special Administrative Region, China (Project Codes: 14204418). “101 Helena is discovered by James Craig Watson.”. Recently, many research works have been proposed for this task. For example, Gardent et al. (2017a,b) create the WebNLG dataset to generate descrip- tion for triples sampled from DBPedia (Auer et al., 2007). Lebret et al.’s (2016) method generates people’s biographies from extracted Wikipedia in- fobox. Novikova et al. (2017) propose to generate restaurant reviews by some given attributes and Fu et al. (2020a) create the WikiEvent dataset to generate text based on an event chain. However, the works mentioned above usually map structured triples to text via a supervised seq-to-seq (Sutskever et al., 2014) model, in which large amounts of an- notated data is necessary and the annotation is very expensive and time-consuming. We aim to tackle the problem …

Content truncated to 500 words