Table 2 Description of the heuristics used to evaluate the recommendations

Using the 500 papers’ first authors (we call this heuristic “the first-author heuristic”)

A recommendation is successful for a researcher r if, based on the profile of r built from the dataset of abstracts, some of the top statements of ignorance recommended to r from the augmented dataset in fact come from a paper written by r.

Predict the interest of the “next papers” of r without having to see them during training.

The quality of the recommendation may partially depend on particular words and expressions r uses in their papers.

Using the co-authors of the 500 papers’ first authors (we call this heuristic “the co-authors heuristic”)

A recommendation is successful for a researcher r if, based on the profile of r built from the dataset of abstracts, some of the top statements of ignorance recommended to r from the augmented dataset come from a paper for which the first author is a co-author of r in the dataset.

The co-authors’ interests of r can be related to the interests of r AND the quality of the recommendation does not depend on particular words and expressions r uses in their papers.

The co-authors’ interests can be really different than r’s interests.

Using the concepts of the 500 papers’ first authors (we call this heuristic “the concepts heuristic”)

A recommendation is successful for a researcher r if, based on the profile of r built from the dataset of abstracts, some of the top statements of ignorance recommended to r from the augmented dataset contains concepts that are also present in the abstracts of r in the dataset of abstracts.

Does not depend on the writing style of r or of their co-authors.

The quality of the recommendation may rely on a concept in the statement of ignorance that is in fact present in the abstracts of most, if not all, of the 500 authors (i.e., the statement of ignorance would be considered relevant for most, if not all, of the 500 authors).