Semantic complexity of geographic questions: A comparison in terms of conceptual transformations of answers

There is an increasing trend of applying AI-based automated methods to geoscience problems. An important example is a geographic question answering (geoQA) focused on answer generation via GIS workflows rather than retrieval of a factual answer. However, a representative question corpus is necessary for developing, testing, and validating such generative geoQA systems. We compare five manually constructed geographical question corpora, GeoAnQu, Giki, GeoCLEF, GeoQuestions201, and Geoquery, by applying a conceptual transformation parser. The parser infers geo-analytical concepts and their transformations from a geographical question, akin to an abstract GIS workflow. Transformations thus represent the complexity of geo-analytical operations necessary to answer a question. By estimating the variety of concepts and the number of transformations for each corpus, the five corpora can be compared on the level of geo-analytical complexity, which cannot be done with purely NLP-based methods. Results indicate that the questions in \textit{GeoAnQu}, which were compiled from GIS literature, require a higher number as well as more diverse geo-analytical operations than questions from the four other corpora. Furthermore, constructing a corpus with a sufficient representation (including GIS) may require an approach targeting a uniquely qualified group of users as a source. In contrast, sampling questions from large-scale online repositories like Google, Microsoft, and Yahoo may not provide the quality necessary for testing generative geoQA systems