JMIR Publications

ABSTRACT

Background:

The normalization of institution names is of great importance for literature retrieval, statistics of academic achievements, and evaluation of competitiveness of research institutions. Differences in authors' writing habits and spelling mistakes lead to variant names of institutions, which affects the analysis of publication data. With the development of deep learning models and the increasing maturity of natural language processing methods, training a deep learning-based institution name normalization model can heighten the accuracy of institution name normalization from the semantic level.

Objective:

This study aimed to train a deep learning-based model for institution name normalization based on the multi-source literature feature fusion of institutional address data, which can realize the normalization of institution name variants with the help of authority files, and achieve a high specification accuracy after several rounds of training and optimization.

Methods:

In this study, an institution name normalization model was trained based on Bidirectional Encoder Representation from Transformers (BERT) and other deep learning models, mainly including the classification model, hierarchical relation extraction model, matching and merging model of institutions. Then the model was trained to automatically learn institutional features by pre-training and fine-tuning, and institution names were extracted from affiliation data of 3 databases: Dimensions, Web of Science, and Scopus to complete the normalization process.

Results:

It was found that the trained model could achieve at least three functions as follows: Firstly, the model could identify the institution name that is consistent with the authority files and associate the name with the files through the unique institution ID; Secondly, it could identify the non-standard institution name variants, such as singular, plural changes, abbreviations and update the authority files; Thirdly, it could identify the unregistered institutions and add them to the authority files, so that when the institution appeared again, the model could identify and treat it as a registered institution. Moreover, testing results showed that the accuracy of the normalization model reached 93.79%, indicating the promising performance of the model in the normalization of institution names.

Conclusions:

The deep learning based institution name normalization model trained in this study exhibits high accuracy. Therefore, it could be widely applied in the evaluation of competitiveness of research institutions, analysis of research fields of institutions, and construction of inter-institutional cooperation networks, etc., showing high application value.