Improved class-specific vector for biomedical question type classification

ABSTRACT

Polysemy is a constant issue in biomedical terms which also affects its QA system. In our work, we consider polysemous words as weak aspect in biomedical question classification and propose two vector model-based solutions that determine the class-specific features of biomedical terms. In first approach, label independent class vector and general word vector are combined using linear compositionality property of vector to generate multiple class-specific embeddings of words. Second is the feature fusion approach, which combines the class-specific sense vector of a word with vectors generated in the first approach. Besides this, a survey dataset COVID-S is introduced in this paper, which is a collection of public queries, myths, and doubts about novel COVID-19 diseases. The series of experiments are performed on two biomedical questions datasets, BioASQ8b and COVID-S, and the results of comparisons with other state-of-arts prove its integrity using SVM and naïve Bayes.