Classification or categorization of texts has been defined as the process using which text is classified into certain distinct groups or categories. Text classification is a complex process that involves procedures like model training, data processing, reduction of noise, and data transformation. Used in complex systems, this process amalgamates the recent procedures with their use in such systems. Recent years have seen the development of new procedures for text classification that can produce accurate results[1][2][3]. The use of Machine learning techniques to classify texts has gained prominence after digital versions of documents were made available. The large number of such documents being published regularly in a digital form has made the process of classifying texts a necessity.
Texts can be classified based on the topics and also based on the genre. While topic-based classification techniques classify documents based upon the nature of the topics [4], genre-based classification techniques classify documents based upon reviews, articles, etc.[5]. The data for classification is normally obtained from different sources and has a heterogeneous character, ie each data differs in items like the style of writing, use of vocabulary, formatting style, etc.
This study classifies documents on Chemistry from various sources into custom classes and multilevel subdivisions using a variety of machine learning algorithms to arrive at the best predictive model. The classification model has been built considering four important parameters which include title, abstract, text, and initial keywords. The model also provides the users an opportunity of adding more features. The first step in the process of text classification involves data cleaning which includes removing punctuation, breaking the text into separate words, stemming of broken words, etc.
pg. 2
The data has been extracted from multiple sources and a data frame has been constructed by selecting 100,000 most relevant research papers on Chemistry from each sub-category. Libraries like numpy, pandas, re, and nltk libraries have also been imported for the purpose of this study.
After importing the required data, we have classified these based on four different parameters: abstract, text, title, and index keywords, allowing the users to add more parameters. We have used the Bag of Words (BoW) model to convert the cleaned text documents into a matrix of token counts. We have also replaced the categorical values with numerical values by providing different integral values to all the sub-topics using Label encoding on the Topic column. The dataset was then subdivided into a training dataset (70%) and a testing dataset (30%).
Various Machine learning algorithms from the Scikit Learn Library including Multinomial NB, Linear Support Vector Machine, Decision Tree, Random Forest Classifier, and XGB Classifier have been used to classify the texts.
Our analysis shows that XGB Classifier Algorithm was able to classify the text with a very high degree of accuracy, followed by the Decision Tree Algorithm. Arranging in the descending order of accuracy for the remaining models, the Machine learning models are Random Forest, Linear Support Vector Machine, and Multinomial Naïve Bayes.
We have tuned the hyperparameters in all the algorithms using the randomized search cross-validation process which validates the model and splits the entire data into numerous training and testing datasets.

Classifying Chemistry literature using Machine Learning (ML) Techniques

Join us for ETD2021 (Electronic Theses and Dissertations) Conference in November 2021 in the United Arab Emirates.

Our theme, ‘Open scholarship in a post-pandemic world’ addresses the enhanced need for discoverability of research findings in a virtual world.

Since the global impact of the COVID-19 pandemic has constrained international travel and mobility, we have all come to rely on working in the virtual world. These developments have merely accelerated the existing trends towards more electronic and digital repositories.

Since 1991, the Networked Digital Libraries of Theses and Dissertations (NDLTD) have contributed to research collaboration by identifying and aggregating digital repositories of theses and dissertations as a previously fugitive class of research resources.