Abstract
The abstract is an extensive summary of a scientific paper that supports making a quick decision about reading it. The employment of a structured abstract is useful to represent the major components of the paper. This, in turn, enhances extracting information about the study. Regardless of the importance of the structured abstract, many computer science research papers do not apply it. This may lead to weak abstracts. This paper aims at implementing the natural language processing (NLP) techniques and machine learning on conventional abstracts to automatically generate structured abstracts that are formatted using the IMRaD (Introduction, Methods, Results, and Discussion) format which is considered as a predominant in medical, scientific writing. The effectiveness of such sentence classification, which is the capability of a method to produce an expected outcome of classifying unstructured abstracts in computer science research papers into IMRAD sections, depends on both feature selection and classification algorithm. This can be achieved via IMRaD Classifier by measuring the similarity of sentences between the structured and the unstructured abstracts of different research papers. After that, it can be classified the sentences into one of the IMRaD format tags based on the measured similarity value. Finally, the IMRaD Classifier is evaluated by applying Naive Bayes (NB) and Support Vector Machine (SVM) classifiers on the same dataset. To conduct this work, we use dataset contains 250 conventional Computer Science abstracts for periods 2015 to 2018. This dataset is collected from two main websites: DBLP and IOS Press content library. In this paper, 200 xml based files are used for training, and 50 xml based files are used for testing. Thus, the dataset is 4x250 files where each file contains a set of sentences that belong to different abstracts but belong to the same IMRaD sections. The experimental results show that Naive Bayes (NB) can predict better outcomes for each class (Introduction, method, results, Discussion and Conclusion) than Support Vector Machine (SVM). Furthermore, the performance of the classifier depends on an appropriate number of the representative feature selected from the text.