عنوان مقاله [English]
Objective: The social web have provided a platform for publicizing open peer review reports. In the new open reviewing sphere, journal readers, authors, editors, and reviewers can involve in multilateral discussions on the reviewed papers and share their comments and viewpoints on the merits and probable pitfalls of the papers. The open peer review comments may, hence, reflect the features of their mother articles. In order to identify this potential, the present study investigates how probable similar comments are in accurately predicting similar papers.
Methods: Using a quantitative content analysis method, the present study applies natural language processing techniques to analyze the contents of a sample of papers in medicine and life sciences and their comments. To do so, it builds a test collection extracted from F1000Research. F1000 Research is an open access publishing platform that adheres to an open peer reviewing process by transparently providing the public with peer review reports, authors’ responses, and users’ comments about the papers openly published on the platform. The test collection consists of 2212 papers and their comments. 100 papers are randomly selected as seed documents that serve as queries. The similarities between the comments and the contents of the papers are calculated using Cosine similarity of TF-IDF values. The TF-IDF values are calculated for both unigrams and bigrams extracted from the contents and comments. The correlation between the content and comment similarities is analyzed using Spearman correlation, given the non-normality of the data distributions. The accuracy of prediction of the papers’ content similarity by the similarity of their comments is tested using Receiver Operating Characteristic (ROC) curves.
Findings: The results of the Spearman correlation reveals a significant correlation between the content and comment similarities. This signifies that more similar papers are more likely to receive similar comments and vice versa. The ROC curves show that similar comments can significantly identify the similar papers, either at unigram or bigram level. The prediction is highly accurate.
Conclusion: Similar comments are found effective in representing similar papers. In other words, similar comments are expected to present similar papers. This finding has implications for interactive information retrieval systems where users are interested in reading experts’ comments on a given paper before viewing or downloading the paper itself. The findings also may pave the path towards new studies about the application of the comments in such spheres as information retrieval, evaluation or classification, where content similarity is of importance.