Corpus Reacción: consumers engagement in Facebook posts

We took on the task of collecting and standardizing a large dataset, ≈ 14, 000 of Facebook posts (https://github.com/lyr-uam/CorpusReaccion) from 10 brands that have an important presence in Mexico. Collected corpus (Table 1), in Spanish language, can be used for training and evaluating automatic systems that aim at predicting several customer’s engagement metrics, specifically Face- book’s reactions (i.e., Like, Love, Haha, Wow, Sad and Angry), sharing amount, and the number of comments generated by a post. Therefore, we define the task of predicting consumer’s engagement as the process of classifying whether a post will have higher (or lower) impact volume than the average seen in training data. Accordingly, our collected dataset defines six binary classification prob- lems, namely: i) comments (|C|), ii) sharing (|S|), ii) total reactions (|R|), iv) positive reactions (|R+|), v) negative reactions (|R−|) and, vi) neutral reactions (|R |). Each classification problem has categories high-impact and low-impact. The methodology for assigning each post’s category was: for each classification problem, we compute the average value of metric k among all the posts from the ten brands, this is referred as x̄k . Once we know the value x̄k , for each post contained in brand i, we review the value of metric k in post pi , thus, if pi,k > x̄k the category of the post is assigned to high-impact, or low-impact otherwise.

Table 1. Number of high- and low- impact instances for each problem.