Drilling engineering reports record geological information about oil and gas reservoirs as well as various drilling engineering parameters. The automatic extraction of unstructured information from these reports can significantly improve the efficiency of data integration into data lakes, thereby enabling more efficient data management. However, these reports typically have domain-specific characteristics, and the diversity of their structure and language presents considerable challenges for accurate named entity recognition (NER). Currently, deep neural network models commonly used for NER are typically trained or fine-tuned on small-scale annotated datasets, leading to two main issues. First, the lack of large-scale annotated corpora limits the diversity of training samples, which in turn causes poor performance when the model encounters new or unseen data, decreasing the model’s generalization ability across different types of data. Second, existing models lack the ability to effectively model long-distance contextual information in texts. Since relevant entities may be scattered across long text segments in drilling engineering reports, these methods often struggle to capture and recognize relationships between named entities in complex documents. To address the aforementioned issues, this paper proposes an intelligent method for named entity recognition in drilling engineering that integrates rotational position embedding and masked conditional random fields. The proposed method is based on a Transformer encoder, a bidirectional long short-term memory network (BiLSTM), and a conditional random field (CRF) architecture. The Transformer encoder leverages pre-trained language models to provide rich contextual semantic representations, BiLSTM captures sequential dependencies, and CRF is used for sequence labeling. Moreover, the traditional CRF is improved by designing a masked modeling mechanism, which restricts the generation of inverted sequences, thereby enhancing the consistency of sequence labeling order. The integration of rotational position embedding further enhances the model's awareness of relative positional information in the text, allowing the model to better capture dependencies between distant words. This improves the model's ability to recognize named entities spread across larger contextual ranges. In addition to model improvements, this paper also addresses the issue of insufficient training data by constructing a domain-specific named entity corpus. This corpus includes annotations for 12 categories of entities, covering a total of 20,727 entity labels across 4,000 text segments. This enriched dataset provides more diverse training samples, which helps improve the model's generalization ability. Experimental results show that the proposed model achieves an F1 score of 86.49 on the test set, representing an improvement of 2.65 percentage points over the previous best-performing model. Furthermore, the model demonstrates significant improvements in recognizing entities with long-tail distributions, which are often underrepresented in typical training datasets. This method not only expands the application of named entity recognition in the field of drilling engineering but also provides engineers with an efficient tool for extracting critical information. By accelerating the analysis of drilling data, it improves the efficiency of drilling operations management and enhances data lake integration, ultimately bringing positive impacts to the decision-making process in drilling projects.