To avoid the repetition of any noun, appearing in a sentence in the course of a dialogue, pronouns are often used in subsequent sentences to replace that particular noun. This economy of articulations is a phenomenon commonly observed in human languages. The Anaphora resolution is essentially required in any natural language dealing-based system for a variety of functions, such as information extraction, question and answer, machine translation, document classification, document summarization, dialogue agent systems and the like [1,2,3]. To correctly analyze the meaning that a text recalls, the co-relationship and reference relationship of the objects included in the text should be made clear.

The Tapper algorithm, which can extract the step-wise information of documents and the accumulated data of medical organizations by applying the Bernoulli theorem and the Bayesian theorem, was developed for classification of the documents [4]. Thus, it became possible to automatically generate medical terms and abbreviation dictionaries from medical data by using a statistical learning method and logistic regression analysis [5]. A prior study also used filtering and searching methods of information to look up medical records similar to the symptoms of a certain patient [6].

Computerization of laboratory medicines has been used to report, check, and manage all the test results rapidly, which increased the overall efficiency of the examination and treatment processes [7]. In this regard, it is worth mentioning that, though the use of computer-based expert systems has increased in decoding test results and interpretive reports, such processes are still in their rudimentary stages [8,9]. The primary reason that computerized analysis has gained immense popularity is because it enables clinicians to provide more elaborate information on treatments, diagnoses, and healing progress than simple briefings on the resulting test values [10]. In other words, the purpose of making interpretive reports has been not only to reduce any possible error in the interpretation of the test results, but also to be able to select additional tests and treatments, if required [11].

With the introduction of Electronic Medical Records, where all the medical records are saved as computer-based documents instead of the conventional paper-based ones, medical organizations started processing all data that has been accumulated over a long period to extract meaningful information and knowledge from them. The processing of these accumulated data collected through clinical tests and medical studies has enabled the reutilization of old data to share information and has also improved the quality of medical services through the gaining of new knowledge and the development of a rapid decision-making support system [12].

To extract the right meaning of a pronoun for medical information out of a natural language dialogue, it is important to consider the features of the pronoun, to determine anaphora. To understand the meaning of a natural language dialogue, a rule-based method is needed to form a system that can replace the personal and demonstrative pronouns of a Korean language dialogue with personal and demonstrative words. Many researchers have used the machine learning-based method for natural language dealings to check its classification function in a number of studies [13,14,15,16,17,18]. Machine learning-based methods have also been reported to translate anaphora in the past [19]. However, the machine learning-based method has a limitation that it depends on a small number of data. It also has a difficulty in annotating a variety of information to translate anaphora. On the other hand, an information searching and processing system using meaningful information was also used in the semantic-based method [20,21,22,23,24]. The semantic-based method involves encoding of cultural memories or facts through a graph data structure. One past studies used the semantic-based method to check the possibility of making a meaningful combination between predicates and referent candidates [25,26].

The current study proposes a method of translating anaphora in a dialogue unit of language processing for a dialogue processing system. The proposed method was designed to solve several types of anaphora phenomena. Interestingly, anaphora recognition can appear in a variety of forms, depending upon the various expressions made in the course of a dialogue; however, they are not considered as different phenomena. That is why various types of anaphora, instead of a single type, were considered together at the same time in past studies related to the Korean anaphora recognition systems. In terms of language processing, a better system can be constructed in which a variety of language forms and anaphora phenomena can be analyzed efficiently. However, straightforward application of the abovementioned methods to resolve anaphora has several limitations. Therefore, in the current study, a new system was developed, a so-called hybrid-type method, which combined the rule-based, semantic-based, and machine learning-based methods to achieve better functionality. This new system with a hybrid-based method has the potential to be used actively as a translation engine or dialogue agent in medical information systems.

Functional comparisons were made depending on the algorithms applied to machine learning and the size of training data of an entity name recognizer.

To estimate its functional difference from that of the hybrid-based system, the machine learning-based method was employed solely using the learning algorithms, which are related only to machine learning-based anaphora recognition. As shown in Table 6, no significant difference was observed in the precision rate, depending on the learning algorithms. However, it is assumed that the feature vector failed to reflect the whole structure of the targeted anaphora recognition domain.

Due to dependence on the semantic scope classifying capacity of the nouns/noun phrases of the entity name recognizer, an experiment was conducted in which the size of the training data used in the NER model was treated as an independent variable. The results shown in Table 7 indicate that the precision rate of the anaphora recognition model increased log-linearly in proportion to the size of the training data of the entity name recognition model. At the training time of the entity name recognition model, its internal features were disregarded to utilize only the feature vector of its neighboring morphemes, so that the function of the entity name recognition model can be proved unfavorable.

The system evaluation results show that the rule-based method achieves 54.9% precision, whereas our suggested system, which is a hybridization of rule and machine learning, achieves 63.7% precision as shown in Table 8. In terms of sensitivity and specificity, the results are consistent.

The purpose of the current research was to check a hypothesis that the rule-based and machine learning-based algorithms could impart positive effects to the functionality of the system. As predicted, it was confirmed that anaphora recognition precision rate could be improved as compared to systems using only a rule or classification model singly or that calculated the mapping table score in the hybrid-based method.

Table 1 also showed that the rule-based system achieves superior function of types 2 and 3, and the machine learning-based system achieves superior function of type 1. Since the hybrid system is designed to take a wider context into consideration than rule-based system, it is natural that it shows better performance in anaphora resolution.

In the current study, a pronoun-type anaphora recognition system was developed. To optimize the mapping scoring model for referent candidates and referred expressions, a hybrid-based method was generated. This hybrid-based method, developed by the combination of rule-based and machine learning-based methods, represents a new system with enhanced functional capabilities as compared to other pre-existing individual methods.

The anaphora resolution system has great potential for use in the medical domain in that it is an essential component in semantic analysis systems; it can be used either for deep semantic analysis of medical free-text to extract information and knowledge in an automatic manner or to design a dialogue system which can provide a healthcare consulting service.