MONETA: Multimodal Industry Classification through Geographic Information with Multi Agent Systems
Summary: arXiv:2604.07956v2 Announce Type: replace
Abstract
Industry classification schemes are integral parts of public and corporate databases as they classify businesses based on economic activity. Due to the size of the company registers, manual annotation is costly, and fine-tuning models with every update in industry classification schemes requires significant data collection. To address these challenges, we replicate the manual expert verification process by utilizing existing or easily retrievable multimodal resources for industry classification.
Introducing MONETA
We present MONETA, the first multimodal industry classification benchmark that combines text-based and geospatial sources. The text sources include widely-used platforms such as:
- Websites
- Wikipedia
- Wikidata
In addition, geospatial data is sourced from:
- OpenStreetMap
- Satellite imagery
Dataset Overview
Our dataset comprises 1,000 businesses located in Europe, categorized according to 20 economic activity labels as per EU guidelines (NACE). This dataset is designed not only to facilitate efficient industry classification but also to support the ongoing evolution of classification schemes in a rapidly changing economic landscape.
Performance and Results
Our training-free baseline model achieves an accuracy of 62.10% when utilizing open-source Multimodal Large Language Models (MLLM) and 74.10% with closed-source MLLM. These results highlight the potential of multimodal approaches in enhancing classification accuracy without the need for extensive training datasets.
Enhancements and Future Directions
We have observed a significant performance increase of up to 22.80% by integrating multi-turn design, context enrichment, and providing classification explanations. These enhancements not only boost accuracy but also improve the interpretability of the classification results. Our commitment to transparency and accessibility will lead to the release of our dataset along with enhanced guidelines for further industry classification research.
Conclusion
MONETA represents a significant advancement in the field of industry classification by leveraging multimodal resources. By streamlining the classification process and reducing reliance on manual annotation, we aim to foster a more efficient and adaptive approach to understanding economic activities across Europe. As we prepare to release our dataset and guidelines, we invite collaboration and feedback from the research community to further refine and enhance this innovative framework.