Authors:
Sergio Jimenez, Youlin Avila, Youlin Avila, George Dueñas, Alexander Gelbukh
Abstract:
The decision of reading or not a research paper is commonly made while reading its title and abstract. Although content and merit should lead to that decision, other factors such as writing style may intervene. Eventually, more readings could produce more citations. We investigated the stylistic factors in the title and abstract of research papers that affect their “citability”, and built a prediction model for citations at 5, 10, and 15 years. Since the number of citations is the preferred ranking function of several academic search engine (...)
The decision of reading or not a research paper is commonly made while reading its title and abstract. Although content and merit should lead to that decision, other factors such as writing style may intervene. Eventually, more readings could produce more citations. We investigated the stylistic factors in the title and abstract of research papers that affect their “citability”, and built a prediction model for citations at 5, 10, and 15 years. Since the number of citations is the preferred ranking function of several academic search engines, our “citability” function could alleviate the under-representation of recent not-yet-cited papers in query results. For this study, we collected a large dataset of around 750,000 titles and abstracts from articles in Scopus, intended to be representative of the entire science. For each instance, we extracted a relatively large set of 3578 stylistic features that were extracted at different linguistic levels, i.e. characters, syllables, tokens (i.e. words), sentences, stop/content words, and part-of-speech (POS) tags. Particularly, we present a novel set of corpus-based stylistic features that we called Corpus Spectral Signatures (CSS). We found out that a linear prediction model for citations (binned into quartiles) build with only the top-250 correlated features achieved a mean absolute error of 0.805 quartiles, and that on average, predictions were highly correlated with their real values (Spearman’s $$rho=0.515$$ ). CSS features were among the top correlated features, but POS features were the most predictive group of features in an ablation study.
(Read More)