Automated Event Log Generation from Unstructured Text Using Finetuned LLMs
A research paper proposes a scalable framework using finetuned large language models (LLMs) to automatically generate structured event logs from unstructured text for process mining. The authors created a text-to-log dataset and finetuned LLMs, showing that finetuning outperforms few-shot and zero-shot prompting by a large margin.
Process mining relies on structured event logs, which are typically created manually by experts, leaving unstructured data like incident tickets and manuals underutilized. This research addresses the bottleneck by finetuning LLMs on a newly created text-to-log dataset to extract high-fidelity event logs from unstructured resources. The finetuned models significantly outperform few-shot and zero-shot prompting, indicating finetuning is necessary for reliable event data generation. The method offers a promising pipeline for leveraging previously unused organizational knowledge.
The key technical finding is that finetuning LLMs on a domain-specific text-to-log dataset yields substantially better event log extraction than prompting-based approaches. This suggests that task-specific finetuning is critical for structured data generation tasks where output format and semantic fidelity are paramount. The creation of a dedicated dataset enables the model to learn the mapping from natural language descriptions to event log schemas, including case IDs, activities, and timestamps.
This approach could lower the cost and expertise barrier for process mining adoption by automating the most labor-intensive step: event log creation. Organizations with large volumes of unstructured operational text (e.g., IT service management, customer support, healthcare) could unlock process insights without manual log construction. The need for finetuning implies that vendors may need to offer domain-specific model customization or pre-finetuned models for common log formats.
Automating event log generation reduces manual effort and expertise requirements, enabling faster and cheaper process mining projects. It can unlock value from previously untapped unstructured data, leading to better process optimization, compliance monitoring, and operational efficiency. The technology could be commercialized as a feature in process mining software or as a standalone data preparation service.
Next observable signals include: release of the text-to-log dataset and finetuned model weights; follow-up work extending the framework to multilingual or multi-domain logs; and pilot deployments in enterprise process mining tools. If the approach proves robust, it could become a standard preprocessing step in process mining pipelines, potentially integrated into commercial platforms.