Numerous scholars have conducted research on institution name normalization and achieved promising performance. According to the key techniques for institution name normalization, these methods can be divided into 4 groups: the string similarity matching–based method, rule-based method, machine learning–based method, and deep learning–based method.

The string similarity matching–based method, dominated by the Edit Distance and Jaccard Index, has been applied to determine whether 2 names represent the same institution by comparing their literal similarity, thereby realizing the identification of 1 institution despite its different names. Edit Distance has been used to calculate differences in the strings of institution names and to recognize spelling mistakes and other name variants since 1997 [2]. Subsequently, James et al [3] further improved the clustering algorithm by combining the Jaccard Index and Edit Distance and clustered similar strings by replacing each string in the cluster with a standard form to achieve name normalization. Jacob et al [5] designed the sCooL system, which can be used to normalize institution names in the resumes using Levenshtein, N-gram, Jaccard similarity, and other algorithms. However, some issues cannot be addressed by relying solely on the string similarity matching–based method. For example, similar names of different institutions may easily cause mismatching.

The principle of the rule-based method is to manually build the rules using the author’s address information, institution attributes, and other features to recognize possible matching institution names based on the rules and eliminate the wrong matching pairs according to the threshold. To eliminate the limitations of string similarity matching, Huang et al [6] built a rule-based institution name disambiguation algorithm using data mining technology, which has reached a high precision in mathematics, psychology, and other fields. Huang et al [7] also established institution name matching rules by introducing the unique identifier of the institution and determined whether 2 names represent the same institution by comparing the Global Research Identifier Database ID, International Standard Name ID, and other identifiers. Some institutions have also conducted corresponding research. For example, the bibliometric database built by the Swedish Research Council based on WoS according to the rules was subjected to affiliation disambiguation for the matching of institutions, cities, addresses, etc. These rules can be used to process more than 99% of the Swedish address strings [8]. Generally, this type of method has poor portability owing to the manually developed rules.

Various machine learning models have been trained for institution name normalization and name disambiguation, such as naive Bayes, support vector machine (SVM), and K-means clustering. Han et al [9] compared the disambiguation effects of naive Bayes and SVM on the ambiguity caused by name abbreviations and spelling mistakes in different scenarios. Balsmeier et al [10] disambiguated the inventor’s name in the patent database using the K-means clustering algorithm.

Compared with the machine learning–based methods involving much human labor, the deep learning–based models may automatically learn the institutional features through less annotated data. He et al [11] established a deep neural network–based entity disambiguation model, which was used to measure the entity similarity in combination with the context and optimize the representations of entities and documents, thus achieving promising disambiguation results. With regard to the neglect of the word order in text by the Bag of Word model, Phan et al [4] firstly used long short-term memory (LSTM) and attention mechanism for entity disambiguation. Jiang et al [12] developed a Dual-Channel Hybrid Network model by fusing a convolutional neural network (CNN) model and a capsule model with a self-attention mechanism, outperforming the mainstream deep learning model in Chinese short text disambiguation. Entity normalization and disambiguation based on deep learning models have higher accuracy and recall rates than traditional methods and therefore deserve more research attention.

On the basis of the initial success of machine learning, deep learning models have been widely used in speech recognition, image classification and search, natural language processing (NLP), and other fields. Deep learning models can learn features from the training data and exhibit a stronger data processing ability [13]. The institution name normalization mentioned in this study actually refers to the completion of NLP tasks using deep learning models. Common models mainly include CNN, recurrent neural network (RNN), LSTM, and BERT models. Previous developmental studies have demonstrated that both CNN and RNN models have showed their respective advantages in emotion analysis, entity recognition, part-of-speech tagging, and other NLP tasks [14]. Initially, the CNN model was primarily used for image processing and then applied to effectively mine the semantic information of the context [15], which is a typical NLP task. With a relatively simple internal architecture and low requirements for computing resources, the RNN model cannot effectively deal with the correlation between long sequences and thus easily causes a vanishing gradient or exploding gradient [16]. Therefore, the gating mechanism of LSTM was proposed to address the gradient issues.

In this study, we trained the normalization model of institution names developed on BERT, which is a pretrained language model that has been broadly used for entity extraction, text classification, emotion analysis, and other NLP tasks with the best performance [17]. With the architecture of a bidirectional transformer [18], BERT can complete parallel computing and thus greatly improve the operation efficiency. In addition, the core mechanism of the transformer, the self-attention mechanism, enables the BERT model to pay more attention to the valuable information among the input data and assign different weights to the words by fully learning contextual features [19]. The BERT model has been broadly applied to named entity recognition (NER) tasks and provides promising results [20,21]. Commonly used pretrained language models include generative pretrained transformer, Global Vectors for Word Representation (GloVe), and Embeddings from Language Models. As the BERT model is state-of-the-art in many NLP tasks [22] and the study [23] shows that the BERT-based model is easily fine-tuned and can be applied to any form of entity name normalization in the biomedical field, it was adopted in this study. It could empower this study and enable the institution name normalization model to achieve higher accuracy.

NER is a common NLP task. Initially, NER was mainly used for string matching based on dictionaries and manual rules, but it showed poor feasibility and portability when dealing with complex data [24]. With the deepening of research, machine learning–based methods have been gradually adopted in NER, such as Hidden Markov Model [25], conditional random field (CRF) [26], and SVM [27] models, but they require much manual work. Deep learning methods are widely used in NER tasks. Hammerton et al [28] first used LSTM to complete NER tasks with data from the Reuters Corpus and the European Corpus Initiative Multilingual Corpus, and Peng et al [29] proposed the LSTM-CRF model on the basis of improved existing models and methods, which has significantly optimized the word segmentation effect compared with traditional methods. Lample et al [30] proposed a neural network model combining bidirectional LSTM (BiLSTM) and CRF, and the context sequence information can be obtained with this bidirectional structure that has been widely used in NER tasks. To further recognize the overlapping nested entities in sentences, span-based methods have been proposed in the NER field. For example, Mandar et al [31] proposed a pretrained model, SpanBERT, which was experimentally demonstrated to generally outperform the BERT model in terms of relation extraction and coreference resolution. The extraction of institutional hierarchical relationships is based on the accurate identification of institution names. In this study, we used the NER model to construct the institutional hierarchical relation extraction model.

The affiliation data from the literature database contain a large number of institution names and attribute information, and it is difficult for the machine to distinguish institutional information from noninstitutional information. Therefore, the accurate extraction of institutional names is the first step in institution name normalization. Owing to the complex and variable forms of addresses extracted from the affiliation data, this field often contains multitier institutions and detailed addresses of the institutions (including the information on street, zip code, region, and country). For example, in “Division of Biology and Biological Engineering, California Institute of Technology, 1200 E. California Blvd., MC 114-96, Pasadena, CA 91125, USA,” it is difficult to recognize the institution name accurately according to the general rules. Considering that the specific keywords and syntactic structures of institution names differ from the detailed address information, a supervised learning–based classifier model was developed in this study to recognize the address information as institutional phrases and noninstitutional phrases through automatic learning of institutional features, so as to extract institution names.

The accurate separation of institutional and detailed address information (noninstitutional part) from the original address is essentially a binary classification problem, specifically for short texts. In the process of institution name recognition, a machine’s understanding of word semantics is crucial. The institution classification model consists of a pretrained BERT model, transformer, and fully connected classifier. Considering that the institution name usually appears before the noninstitutional information (detailed address) in the full address field, the model uses a reverse classification method to segment the original address. That is, it identifies each word in reverse order until it reads a word that represents institutional information and then stops classifying. At this point, the first half of the address is considered to be institutional information and the second half is considered to be noninstitutional information. The BERT [32] model is a deep learning–based language representation model and can provide richer semantic information of words (especially keywords), such as “Institute,” “Department,” and “University,” by virtue of its multilayer transformer and the ability of converting the input texts into word vector representation. The transformer encodes the word order and background information of a word in a phrase. For example, the institutional hierarchical order and the differences between institutional and noninstitutional features can be encoded by the transformer. At present, the commonly used encoders include transformer [18], BiLSTM [30], and CNN [33]. The transformer incorporates a self-attention mechanism that focuses on the valuable information among the input data, which are beneficial for encoding important words; thus, it was adopted as an encoder herein. The output result from the transformer is mapped into a dichotomy by a fully connected classifier, with 0 representing noninstitution and 1 representing institution, thereby achieving the separation of institution names and detailed addresses.

The training process of the model is as follows:

After training, when a phrase was entered, such as “California Institute of Technology,” the model returned “True” (institution) or “False” (noninstitution), and when an original address was entered, the model split it into detailed address and institution information. The operating effects of the institutional classification model are shown in Table 2.

Hierarchical relation is embedded in the institution information. For example, in “Medical Research Council Population Health Research Unit, Nuffield Department of Population Health, University of Oxford,” the “University of Oxford” is the first-tier institution, “Nuffield Department of Population Health” is the second, and “Medical Research Council Population Health Research Unit” is the third. In this study, an institutional hierarchical relation extraction model was constructed with a structure of the NER model plus rules plus hierarchical tree to maximize the effectiveness from multiple perspectives, such as rules and semantics, and normalize institutions at a finer granularity. First, the NER model sequentially labels the original address field with the Beginning Inside and Outside method. Once the labeling is completed, the deep learning model identifies the distinguishing features of institutions across different levels and provides an output for the labeled address field. At this stage, the institution-related information in the original address is labeled based on the primary, secondary, and tertiary levels. To ensure the precise extraction of hierarchical structures, this study also incorporates rules and hierarchical trees, which will be further explained in the latter part of this section.

It is worth noting that most existing techniques for institution name normalization and entity disambiguation operate at a broad level, with fewer studies focusing on the subdivision of institutional hierarchies. In this study, the model not only normalizes the names of institutions at each level but also finely organizes the hierarchical relationships among primary, secondary, and tertiary institutions, which is convenient for researchers to use the hierarchical structure of institutions for literature retrieval and statistics.

The X-CRF framework was adopted as the main model for institution extraction, as this task does not involve the issue of span overlapping. The BiLSTM-CRF and BERT-CRF models were used in the specific experiments for the institutional NER. The experimental results showed that the effect of the BiLSTM model was restricted by the training data, resulting in inaccurate labels predicted by the model on the one hand and out of vocabulary when a word was used as a token on the other hand. Therefore, in this study, a large-scale pretrained BERT model with word pieces as its token was adopted to obtain text representation, which alleviated out of vocabulary problems to some extent. The institutional hierarchical relation in the address was extracted using the BERT-CRF model, and the institutional span and label were obtained. However, there might be span deviation, label error, institutional deficiencies and redundancies, or other problems in the extraction results. Therefore, in this study, the model’s prediction results were corrected primarily through the use of rules and hierarchical trees. The formulated rules include the institution name completion rule and the institutional order correction rule. The former determines whether the extracted content is a complete field, typically by checking whether it ends with a comma. For instance, the institution name is “University of Antwerp (Campus Drie Eiken),” but the NER model only extracts “University of Antwerp.” Meanwhile, the latter corrects the institutional hierarchical order in original addresses that are not standardized, such as “Queen Mary University of London, Institute of Population Health Sciences,” by adjusting the order according to the characteristic words of each level of institutions. Once the rules were manually established, they were incorporated into the model. Finally, for the special hierarchical relationships that cannot be identified by the machine, the institutional hierarchical relationships are imported into the model by constructing institutional hierarchical trees to complement the extraction results of the model. The data size in this study was small; therefore, the hierarchical tree was constructed using the open Neo4j database to correct the institutional hierarchical order. The constructed hierarchical relationship is shown in Figure 3.

The zip code extraction model was applied to extract the zip code from the address information and integrate it with other affiliation data. In the process of canonical information matching, the zip code information plays an auxiliary role in determination. As mentioned in the section of Institution Classification Model before, the institution classification model divides the original address into 2 parts: institutional information and noninstitutional information. The noninstitutional information may contain city, zip code, administrative division, and country. First, the model reads the noninstitutional information in reverse order with the help of the International Organization for Standardization (ISO) standard; eliminates the country, city, administrative division, etc; and then manually researches the coding rules of zip codes of different countries (eg, the zip code of China consists of 6 consecutive digits), builds a zip code rule base, and extracts the zip code from the remaining information, which has been tested to be more accurate than directly extracting the zip code. The specific process for extracting zip code is as follows:

The extraction results are shown in Multimedia Appendix 1. The zip code rules for China, the United States, the United Kingdom, and Germany were included in the database; thus, the zip code extraction model achieved an accuracy of almost 100%.

The institution name extracted from the institution address will be compared and matched with the name in authority files, which are computer-identifiable documents consisting of specification records. If the name matches the institution name in the authority files, then the institution will become a registered institution, and its specification information, such as the new alias, zip code, and official website, can be completed in authority files; otherwise, the institution is unregistered and will be included in the authority files as a new institution after manual review. In the process of matching and merging, the recall rate using traditional matching methods is relatively low because of the differences in text transform, punctuation, and word expression order among the institution names. In this study, the institution matching and merging model was developed using country grouping and vector clustering to improve the matching recall rate and reduce the running time of the model. The specific process is as follows: all registered and unregistered institutions are divided by countries, and rules are established for the preliminary processing of unregistered institutions, which include converting the names to lowercase for text vectorization, and eliminating nonalphabetic characters other than “&” in names and converting “&” to “and.” The GloVe model was used to transform the processed data into vectors, and the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm was then used to cluster institution names under the same country. Subsequently, the new unregistered organizations are added or aliases of registered organizations are included in the authority files based on the clustering results. This operation can achieve multithreaded matching of different countries, greatly improving the matching efficiency.

Text representation is a critical step in the text clustering process, and the method and effect of text representation have a great influence on the model clustering effect. The most commonly used methods in text representation include bag of word and term frequency–inverse document frequency, which perform well in classification and clustering tasks; however, there are still some problems, such as an extremely high vector dimension, sparse data, failure to focus on the word order in sentences, and failure to learn text semantic information [34,35]. Multiple word embedding representation methods have been developed to overcome this limitation, such as Word2Vec [36], GloVe [37,38], and Embeddings from Language Models [39], which can effectively address the semantic problems of words in the text. In this study, GloVe was applied for text representation owing to its advantages of high accuracy and a short training period.

The classical clustering algorithms include K-means, Gaussian mixture model, and DBSCAN. For the K-means and Gaussian mixture models, the number of clusters must be set manually, and the same institution cannot be clustered automatically. However, the DBSCAN algorithm speeds up a rapid clustering and requires no manual setting for the number of clusters [40]. Moreover, the fault tolerance rate of the model can be reduced by setting a shorter intercluster distance when the DBSCAN algorithm is used. Therefore, the DBSCAN algorithm was applied in this study. The normalization of institution names was achieved using vector clustering. When an institution name vector cluster contains both registered and unregistered institutions, the unregistered institutions are grouped under the registered ones and listed as aliases in authority files. In cases where the cluster only contains unregistered institution names, the institution name with the largest number is taken as the base data for grouping and written in authority files after manual review.

In summary, the institution matching and merging model consists of 3 components: GloVe, DBSCAN, and a set of rules. First, we manually formulated rules to transform and process the data. Next, the organization names were represented as vectors using GloVe. Finally, the DBSCAN algorithm clusters the vectors, and the institution names are matched and merged with the authority files according to the clustering results. These 3 components work together to guarantee the precision of institution name matching while maintaining high operational efficiency.