Identify Novel Elements of Knowledge with Word Embedding

As novelty is a core value in scientific advancement and technological progress, several bibliometric measures for novelty have been proposed. Nonetheless, a few fundamental limitations remain, which this study aims to address by offering a validated and field-universal approach to compute element novelty. To this end, we first developed a word embedding model based on large-scale scientific text data. The word embedding technique draws on machine learning to extract semantic information from text data as high-dimensional vectors assigned to every word. Second, using the developed word embedding model, we computed element novelty measures. The validation with our survey scores suggests that the proposed novelty measure is strongly correlated with <em>Novel_B</em> ("discovering and identifying") across most scientific fields. Therefore, our element novelty measure based on our own word embedding model seems to provide a strong approach to quantify element novelty in terms of discovering or identifying something new. This study offers a few contributions. First, we provide a robust and versatile measure of element novelty. Despite its fundamental importance, element novelty has been underserved in comparison to recombinant novelty. Our approach based on word embeddings is robust and efficient in comparison to existing measures of element novelty based on keywords and manually developed controlled dictionaries.