Schema2QA: High-Quality and Low-Cost Q&A Agents for the Structured Web

• DOI: 10.48550/arxiv.2001.05609
• Publication Date: 2020-01-01
• Authors: Silei Xu; Giovanni Campagna; Jian Li; Monica S. Lam
• Keywords: Schema (genetic algorithms); Paraphrase

Building a question-answering agent currently requires large annotated datasets, which are prohibitively expensive. This paper proposes Schema2QA, an open-source toolkit that can generate Q&A system from database schema augmented with few annotations for each field. The key concept is to cover the space of possible compound queries on number in-domain questions synthesized help corpus generic query templates. data and small paraphrase set used train novel neural network based BERT pretrained model. We use Schema2QA systems five Schema.org domains, restaurants, people, movies, books music, obtain overall accuracy between 64% 75% crowdsourced these domains. Once paraphrases obtained schema, no additional manual effort needed create any website uses same schema. Furthermore, we demonstrate learning be transferred restaurant hotel domain, obtaining effort. achieves 60% popular answered using Schema.org. Its performance comparable Google Assistant, 7% lower than Siri, 15% higher Alexa. It outperforms all assistants by at least 18% more complex, long-tail questions.