This workshop will be held on Sept. 12, 2023 (BST).
Session | Time | Speaker | Title |
---|---|---|---|
Opening Introduction | 9:00 - 9:10 | ||
Paper Session 1 | 9:10 - 9:30 | Guiye Li | Bayesian Super-resolution using Deep Generative Models |
9:30 - 9:50 | Joe Tuccillo | SONET++: A Knowledge Graph of Geographic Categories based on OSM Tag Representation | |
Coffee break | 9:50 - 10:20 | ||
Paper Session 2 | 10:20 - 10:40 | Sergios Kefalidis | The Question Answering System GeoQA2 |
10:40 - 11:00 | Stefano De Sabbata | Learning urban form through unsupervised graph-convolutional neural networks | |
Keynote | 11:00 - 12:00 | Anthony G Cohn | Evaluating the Spatial Reasoning Capabilities of Large Language Models |
Closing Remark | 12:00 - 12:10 |
The rapid increase in multimodal data, advanced deep learning algorithms, and the availability of fast hardware have largely contributed to a renewed interest in Artificial Intelligence (AI). Despite many successful stories in computer vision and natural language processing, there are challenges that remain to be solved, such as large scale neural-symbolic reasoning and automatic knowledge graph construction. Unsurprisingly, one of the most prominent topics in AI, nowadays, is the combination of representation learning (connectionist AI) with symbolic representation and reasoning (symbolic AI). By combining these two aspects, we are capable of developing scalable and explainable machine learning models. Such a trend becomes even more pressing with the foundation models (e.g., GPT-4) have been used everywhere just within the past few months. While embracing their surprisingly fascinating performance, the society also calls for a more explainable and responsible use of these foundation models.
From a geospatial point-of-view, GeoAI, as an interdisciplinary field of GIScience and AI, advocating the development of spatially explicit machine learning approaches and the use of novel AI techniques in geography and earth science. Graphs are at the core of GeoAI, as they have been shown to allow effective representations of semantics as well as spatial and temporal relationships. Geospatial knowledge graphs, particularly as symbolic representations of geospatial knowledge, facilitate many intelligent applications such as geospatial data integration and knowledge discovery. Nevertheless, many deep learning models treat geographic entities as ordinary entities in which spatial characteristics, such as spatial footprint or distance decay, are often ignored. This results in suboptimal performance in many geospatial related tasks including geospatial knowledge graph completion, geographic question answering, geographic entity alignment, as well as geographic knowledge graph summarization.
Following its former success in GIScience 2021, this workshop at GIScience 2023 will continue highlighting the importance of geospatial information and principles in designing, developing, and utilizing geospatial knowledge graphs and other GeoAI techniques to discover knowledge in geosciences. We invite researchers from disparate disciplines (e.g., environmental studies, GIScience, AI, cognition, supply chain, humanities, etc.) to submit papers in the following three formats. All submitted papers will be peer-reviewed by our Program Committee.
Title: Evaluating the Spatial Reasoning Capabilities of Large Language Models
Abstract: In this talk I will present some initial results on evaluating the spatial reasoning capabilities of Large Language Models (LLMs). Whilst LLMs have shown remarkable apparent abilities in many areas of question answering, their abilities to perform reasoning is less clear. I will present some results, particularly focussing on qualitative spatial representations and reasoning for some LLMs. One way to probe the limits of LLM capabilities is to conduct an extended conversation (which we call “dialectical evaluation”) rather than simply using a pre-designed benchmark which usually has a fixed set of answer possibilities and I will show some results based on this evaluation method.
Submissions should be updated to the workshop EasyChair page. All submissions will be peer-reviewed. Submissions should use the Lecture Notes in Computer Science template. Additionally, accepted papers will be invited to submit an extended version to a special issue at the International Journal of Applied Earth Observation and Geoinformation on the topic of Spatially Explicit Machine Learning and Artificial Intelligence.