An Unsupervised Joint System for Text Generation from Knowledge Graphs and Semantic Parsing

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 7117–7130, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 7117 An Unsupervised Joint System for Text Generation from Knowledge Graphs and Semantic Parsing Martin Schmitt1 Sahand Sharifzadeh2 Volker Tresp2,3 Hinrich Sch¨utze1 1Center for Information and Language Processing (CIS), LMU Munich 2Department of Informatics, LMU Munich 3Siemens AG Munich martin@cis.lmu.de Abstract Knowledge graphs (KGs) can vary greatly from one domain to another. Therefore su- pervised approaches to both graph-to-text gen- eration and text-to-graph knowledge extrac- tion (semantic parsing) will always suffer from a shortage of domain-specific parallel graph- text data; at the same time, adapting a model trained on a different domain is often impos- sible due to little or no overlap in entities and relations. This situation calls for an approach that (1) does not need large amounts of anno- tated data and thus (2) does not need to rely on domain adaptation techniques to work well in different domains. To this end, we present the first approach to unsupervised text gener- ation from KGs and show simultaneously how it can be used for unsupervised semantic pars- ing. We evaluate our approach on WebNLG v2.1 and a new benchmark leveraging scene graphs from Visual Genome. Our system out- performs strong baselines for both text↔graph conversion tasks without any manual adapta- tion from one dataset to the other. In additional experiments, we investigate the impact of us- ing different unsupervised objectives.1 1 Introduction Knowledge graphs (KGs) are a general-purpose approach for storing information in a structured, machine-accessible way (Van Harmelen et al., 2008). They are used in various fields and domains to model knowledge about topics as different as lex- ical semantics (Fellbaum, 2005; van Assem et al., 2006), common sense (Speer et al., 2017; Sap et al., 2019), biomedical research (Wishart et al., 2018) and visual relations in images (Lu et al., 2016). This ubiquity of KGs necessitates interpretabil- ity because diverse users – both experts and non- experts – work with them. Even though, in prin- 1https://github.com/mnschmit/ unsupervised-graph-text-conversion ciple, a KG is human-interpretable, non-experts may have difficulty making sense of it. Thus, there is a need for methods, such as automatic natural language generation (“graph→text”), that support them. Semantic parsing, i.e., the conversion of a text to a formal meaning representation, such as a KG, (“text→graph”) is equally important because it makes information that only exists in text form accessible to machines, thus assisting knowledge base engineers in KG creation and completion. As KGs are so flexible in expressing various kinds of knowledge, separately created KGs vary a lot. This unavoidably leads to a shortage of training data for both graph↔text tasks. We therefore pro- pose an unsupervised model that (1) easily adapts to new KG domains and (2) only requires unla- beled (i.e., non-parallel) texts and graphs from the target domain, together with a few fact extraction heuristics, but no manual annotation. To show the effectiveness of our approach, we conduct experiments on the latest …

Content truncated to 500 words