Holistic Spatial Reasoning for Chinese Spatial Language Understanding

Abstract

Spatial language understanding (SLU) is an important task in the field of information extraction, and it involves complex spatial analyses and reasoning processes. Unlike English SLU, in the case of the Chinese language, there may be some language-specific challenges, such as the phenomenon of polysemia and the substitution of synonyms. In this work, we explore Chinese SLU by taking advantage of large language models. Inspired by recent chain-of-thought (CoT) strategies, in this study, we propose the Spatial-CoT template to help improve LLMs’ reasoning abilities in order to deal with the challenges of Chinese SLU. Spatial-CoT offers LLMs three steps of instructions from different perspectives, namely, entity extraction, context analysis, and common knowledge analysis. We evaluate our framework on the Chinese SLU dataset SpaCE, which contains three subtasks: abnormal spatial semantics recognition, spatial role labeling, and spatial scene matching. The experimental results show that our Spatial-CoT outperforms vanilla prompt learning on ChatGPT and achieves competitive performance in comparison with traditional supervised models. Further analysis revealed that our method could address the phenomenon of polysemia and the substitution of synonyms in Chinese spatial language understanding.