Neural Machine Translation for Semantic-Driven Q&A Systems in the Factory Planning

Abstract

Shorter lifecycles, increasing product variance and the integration of new products and technologies into existing factories lead to a high complexity in today’s factory planning. In order to master this complexity, many companies attempt to improve their processes by using digitalization tools. This generates enormous amounts of data, which are currently only partially managed centrally in the company. In order to simplify the associated difficulties regarding the access to information, semantic technologies for the generation and application of knowledge graphs (KG) are currently being investigated intensively by research and industry. To retrieve information - stored in the KG - query languages such as SPARQL are used which require knowledge regarding the structure of the KG as well as a profound knowledge of query languages. To also enable non-expert users to retrieve the stored information in the KG, an intuitive user interface for question answering (Q&A) is needed.In this context, we propose a translation model using artificial neural networks (ANN), which translates questions in german into the query language SPARQL. Due to a lack of suitable datasets, we first developed a method to automatically generate synthetic data sets for training the translation model based on existing ontologies. Based on similar approaches from research, we develop an ANN architecture for the translation model. The method for data generation and the ANN were tested and validated at a german car manufacturer using a knowledge graph. It was shown that the developed architecture is particularly suitable for our field of application.