Integrating element correlation with prompt-based spatial relation extraction

Spatial relations in text refer to how a geographical entity is located in space in relation to a reference entity. Extracting spatial relations from text is a fundamental task in natural language understanding. Previous studies have only focused on generic fine-tuning methods with additional classifiers, ignoring the importance of the semantic correlation between different spatial elements and the large offset between the relation extraction task and the pre-trained models.

To solve the problems, a research team led by Feng WANG published their new research on 15 Feb 2025 in Frontiers of Computer Science co-published by Higher Education Press and Springer Nature.

The team proposed a spatial relation extraction model based on Dual-view Prompt and Element Correlation (DPEC) that encodes and marks spatial discriminative information, and used prompts in two formats to extract spatial relations and confidence in spatial relation triggers. This approach significantly enhances the accuracy of spatial relation extraction during the inference phase. The authors first reformulated spatial relation extraction as a mask language model with a dual-view prompt, consisting of a Link Prompt and a Confidence Prompt. Link Prompt can assist the model in incorporating contextual information related to spatial relation extraction and adapting to the original pre-training task of language models. Confidence Prompt can measure the confidence of candidate triplets in Link Prompt and supplement the identification of easily confused examples. 

In the stage of candidate triplet extraction, the team first introduce a BERT-CRF to identify spatial elements and then obtain the set of candidate triplets by arranging the spatial elements.

In the stage of spatial relation classification, they first generate two prompt templates, i.e., Link and Confidence Prompt templates, respectively, according to the set of candidate triplets, and then concatenate the original sequence text and two prompt templates as two input sequences to BERT. Thus, we obtain the representations of the [MASK] tokens in Link Prompt for spatial relation extraction and those in Confidence Prompt for trigger recognition. Finally, taking into account the inherent clustering of spatial elements in terms of semantic correlation between different elements, they fuse the representation of the element correlation between the spatial elements in the Link Prompt classifier. Moreover, they train the two tasks simultaneously in the training step and take the [MASK] result of Confidence Prompt as the evaluation for the Link Prompt classifier in the inference step.

Future work can focus on constructing large-scale spatial relation datasets and improving the performance of OLINK.

DOI: 10.1007/s11704-023-3305-4